Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
SEPARATION OF STATES OF MECHANICAL PRESSES BY ANALYSING TRAINED PATTERNS IN A NEURAL NETWORK
Document Type and Number:
WIPO Patent Application WO/2021/023449
Kind Code:
A1
Abstract:
The present invention is related to a computer-implemented method of, a data processing system for and a computer program product for indicating a failure of a manufacturing process as well as to a corresponding manufacturing machine and further to a computer- implemented method of training a machine learning system (MLS) for indicating states of a manufacturing process. An input signal of a sensor is transformed into a parameter. The parameter is provided to the MLS, which derives latent features. The latent features are mapped into one of several distinct clusters each representing a mode of the manufacturing process. Finally, a failure of the manufacturing process based on the different states of the manufacturing process may be indicated.

Inventors:
REIMANN THORSTEN (DE)
Application Number:
EP2020/068923
Publication Date:
February 11, 2021
Filing Date:
July 06, 2020
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
SIEMENS AG (DE)
International Classes:
B30B15/00; B30B15/26; G05B23/02; G06N3/04; G06N3/08
Foreign References:
US20190118443A12019-04-25
US4633720A1987-01-06
EP3381646A12018-10-03
Download PDF:
Claims:
Patent claims

1. Computer-implemented method of indicating a failure of a manufacturing process, comprising the steps: a) receiving (1) at least one input signal (II..15) based on at least one physical quantity monitored during the manufacturing process by means of at least one sensor (S1..S4); b) transforming (2) the received at least one input sig nal (II..15) into at least one parameter (P1..P5), having a different domain and/or a different refer ence value than the received at least one input sig nal (II..15); c) deriving (3) latent features (LF) based on the at least one parameter (P1..P5) by means of a machine learning system, MLS, which is trained on deriving latent features (LF) based on the at least one param eter (P1..P5) indicative of specific states of the manufacturing process; d) mapping (4) the derived latent features (LF) into one of several distinct clusters (C1..C3) in a two dimen sional, 2D, cluster space, wherein the clusters (cl..C3) represent different sates of the manufactur ing process; and e) optionally (5) indicating a failure of the manufac turing process based on the different states of the manufacturing process.

2. Computer-implemented method according to claim 1, wherein the at least one monitored physical quantity is a press ing force (F) of a press (40) and the at least one gener ated input signal (II..15) is a force-over-time signal and wherein a failure in a pressed workpiece is indicated as the failure of the manufacturing process.

3. Computer-implemented method according to claim 2, wherein the monitored pressing force (F) is a force of a plunger (41) of the press (40) and/or a force on a mould (42) of the press (40).

4. Computer-implemented method according to any preceding claim, wherein the received at least one input signal (II..15) is transformed by selecting a predefined time slice of the signal (II..15) and/or a data cleansing and/or a normalization and/or a centring.

5. Computer-implemented method according to any preceding claim, wherein the MLS is a neural network, NN, prefera bly a deep NN and most preferably a convolutional NN.

6. Computer-implemented method according to any preceding claim, wherein the MLS is deployed on a cloud-based sys tem or on a local computer system of a premise where the manufacturing process is conducted.

7. Computer-implemented method according to any preceding claim, wherein a t-distributed Stochastic Neighbor Embed ding, t-SNE, method is used to map the derived latent features (LF) into one of the several distinct clusters (Cl..C3).

8. Data processing system (10) for indicating a failure of a manufacturing process, comprising means for carrying out the steps of the method according to any preceding claim.

9. Computer program product for indicating a failure of a manufacturing process, comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to any of claims 1 to 7.

10. Manufacturing system (30) comprising: the data processing system (10) according to claim 8; and at least one sensor (SI, S2, S3, S4) generating at least one input signal (II..15) by monitoring at least one physical quantity during conducting a manu facturing process, wherein the generated at least one input signal (II..15) is provided to the data processing system (10).

11. Manufacturing system (30) according to claim 10, further comprising a press (40), wherein the at least one sensor (S1..S4) is a force sensor generating a force signal by monitoring a pressing force (F) of the press (40).

12. Manufacturing system (30) according to claim 10 or 11, wherein the monitored pressing force (F) is a force of a plunger (41) of the press (40) and/or a force on a mould (42) of the press (40).

13. Computer-implemented method of training a machine learn ing system, MLS, for indicating states of a manufacturing process, comprising the steps: i) generating (101) a set of at least one training pa rameter based on training input signals (T1..T5) based on at least one physical quantity monitored during several manufacturing processes by means of at least one sensor (S1..S4) and of corresponding train ing results of the respective manufacturing process es; ii) training the MLS using the set of the at least one training parameter and of corresponding training re sults.

14. Computer-implemented method according to claim 13, where in the training input signals (T1..T5) are real input signals from a real machine (40) of a manufacturing sys tem (30) or manufacturing process or simulated input sig nals from an artificial model of the machine (40) of the manufacturing system (30) or manufacturing process or a combination of both.

15. Computer-implemented method according to claim 13 or 14, wherein the MLS is a neural network, NN, and preferably a pre-trained NN.

16. Computer-implemented method according to any of claims 13 to 15, wherein the training input signals (T1..T5) are transformed into the at least one training parameter of the set by selecting a predefined time slice of the training input signals (T1..T5) and/or a data cleansing and/or a normalization and/or a centring and/or using parts of predefined length of the training input signals (T1..T5), where preferably, the parts of the predefined length of the training input signals (T1..T5) are select ed according to a Gaussian distribution.

Description:
Description

Separation of states of mechanical presses by analysing trained patterns in a neural network

The present invention is related to a computer-implemented method of, a data processing system for and a computer pro gram product for indicating a failure of a manufacturing pro cess as well as to a corresponding manufacturing machine and further to a computer-implemented method of training a ma chine learning system (MLS) for indicating states of a manu facturing process.

In pressing sheets or plates of different materials (e.g. steel, copper, polymer etc.) cracks in a sheet/plate (or ra ther manufactured workpiece), ripples of the sheet/plate or rather manufactured workpiece) and other failures of the man ufacturing process regularly occur. These failures of the manufacturing process are not detected during the pressing process, but only after several pressing cycles during an au tomatically or manually conducted visual inspection. Further, during the pressing process micro-cracks can occur, which are not detected during the visual inspection, but only much lat er during final assembly or even operation.

The later such failures of the manufacturing process are de tected the more expensive their correction becomes.

It is therefore an objective of the present invention to pro vide a reliable and fast automatic indication of a failure of a manufacturing process.

The present invention provides a computer-implemented method of indicating a failure of a manufacturing process according to independent claim 1 and a corresponding data processing system and a corresponding computer program product as well as a corresponding manufacturing system and, further, a com puter-implemented method of training a machine learning sys- tem (MLS) for indicating states of a manufacturing process. Refinements of the present invention are subject of the re spective dependent claims.

According to a first aspect of the present invention a com puter-implemented method of indicating a failure of a manu facturing process comprises the steps a) receiving at least one input signal, b) transforming the at least one input sig nal, c) deriving latent features, d) mapping the derived la tent features and e) optionally indicating a failure of the manufacturing process. In step a) at least one input signal is received. The received at least one input signal is based on at least one physical quantity, which is monitored during the manufacturing process by means of at least one sensor. In step b) the received at least one input signal is transformed into at least one parameter. The at least one parameter has a different domain and additionally or alternatively a differ ent reference value than the received at least one input sig nal. In step c) latent features are derived based on the at least one parameter by means of a machine learning system (MLS). The MLS is trained on deriving latent features based on the at least one parameter indicative of specific states of the manufacturing process. In step d) the derived latent features are mapped into one of several distinct clusters in a two dimensional (2D) cluster space, wherein the clusters represent different sates of the manufacturing process. In step e) a failure of the manufacturing process is indicated based on the different states of the manufacturing process.

According to a second aspect of the present invention a data processing system for indicating a failure of a manufacturing process comprises means for carrying out the steps of the method according to the first aspect of the present inven tion.

According to a third aspect of the present invention a com puter program product for indicating a failure of a manufac turing process comprises instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to the first aspect of the present invention.

According to a fourth aspect of the present invention a manu facturing system comprises the data processing system accord ing to the second aspect of the present invention and at least one sensor. The at least one sensor generates at least one input signal by monitoring at least one physical quantity during conducting a manufacturing process. The generated at least one input signal is provided to the data processing system.

At least one physical quantity like vibrations, temperature, force, torque, rotational/angular/linear speed/acceleration, and/or the like occurring during the manufacturing process is monitored by the at least one sensor. Therefore, at least one appropriate sensor is provided, such as a temperature sensor, a vibration sensor, an acceleration sensor, a position sen sor, a force sensor etc. The sensor is positioned at a corre sponding position where the respective physical quantity can be monitored by the sensor.

More than one physical quantity may be monitored with appro priate sensors, for example vibrations (e.g. vibrations of a machine part or a workpiece) may be monitored with a vibra tion sensor. A temperature during the manufacturing process (e.g. a temperature of a machine part, of a chamber or of a workpiece) may be monitored with a temperature sensor. A ro tational speed may be monitored with a rotation sensor. An acceleration or force (or speed derived from the accelera tion/force) may be monitored with an acceleration or force sensor.

The respective sensor generates a corresponding input signal based on the respective monitored physical quantity. In step a) the generated input signal is provided to and received by the data processing system (e.g. at an interface of a receiv er).

In step b) the received at least one input signal is trans formed into at least one parameter. For example, the input signal may be normalized and additionally or alternatively KPIs like RMS, crest factor, and/or standard deviation may be derived as features of the input signal. Further the input signal may (additionally) be transformed into another domain, for example a vibration signal may be transformed from the time domain (F(t)) into the frequency domain (F(f)) e.g. by means of a Fast Fourier Transformation (FFT). An input signal based on a measured acceleration may be transformed from the time domain (F(t)) into the 3D-domain (F(x,y,z)). Further, the input signal may be transformed such that the feature has a different reference value than the input signal, like an input signal based on a measured temperature may be trans formed from degree Celsius (°C) into Kelvin (K) or an input signal base on a measured speed which is a relative speed has to be transformed (superimposed) into an absolute speed (e.g. rotating element moved by an arm of a machine of the manufac turing system). Further, statistical values may be determined in the step of transforming. For example, in transforming the input signal may first be transformed into another domain and then statistical features may be derived therefrom. All these transformations may be referred to as parameter engineering.

Machine learning algorithms or MLS build a mathematical model based on sample data, known as "training data", in order to make predictions or decisions without being explicitly pro grammed to perform the task. The at least one "engineered" parameter is provided to the MLS in step c). The MLS has been trained such that it can derive latent features based on the provided at least one parameter indicative of specific states of the manufacturing process. Specific states such as fail ures of the manufacturing process (e.g. wrong manufacturing parameters, bad orientation of work pieces, malfunctioning machine parts etc.) manifest themselves in specific patterns of the at least one parameter or more than one parameter. Therefore, the MLS has been trained based on such patterns in a set of at least one training parameter indicative of at least one certain failure (state) of the manufacturing pro cess.

The latent features derived by the MLS from the at least one provided parameter resemble one of several specific trained patterns. The MLS extracts latent information provided with the at least one parameter and connects this latent infor mation internally (e.g. based on trained weights) into the latent features, which provide different information about states such as failures of the manufacturing process than the at least one provided parameter.

In step d) the derived latent features are mapped into one cluster. The cluster is one of several distinct clusters in the 2D cluster space. Thereto, the latent features derived by the MLS are mapped with a specific algorithm into two values (e.g. an X-coordinate and a Y-coordinate), which resemble their difference to other latent features derived by the MLS based on another provided at least one parameter. The latent features derived by the MLS form the at least one parameter have a specific pattern. The pattern can be transformed by the algorithm into the two values specific for the respective pattern of the latent features. Each of the distinct clusters resembles a different state of the manufacturing process (e.g. idle, start-up phase, normal operation and one or sev eral different failures of the manufacturing process). Also a class and additionally or alternatively a severity of a fail ure of the manufacturing process may be derivable from the cluster.

In the optional step e) a failure of the manufacturing pro cess (and its class and/or severity) is indicated based on the state of the manufacturing process resembled by the clus ter. The failure of the manufacturing process, which is based on the pattern of the latent features derived by the MLS from the at least one parameter, may be reported to a user (e.g. technician, mechanic personnel, shift supervisor, facility engineer etc.) who can initiate the necessary steps for bringing the manufacturing process back to normal operation (eliminating the cause of the failure of the manufacturing process).

Additionally, the method may comprise a further step of auto matically initiating appropriate counter measures, like repo sitioning of movable machine parts or workpieces, refilling of coolant or lubricant etc. to bring the manufacturing pro cess back to normal operation.

The manufacturing system may comprise machines and machine parts by means of which the manufacturing process can be con ducted. The at least one sensor generates the at least one input signal based on the at least one physical quantity mon itored during the manufacturing process.

With the present invention it is possible to detect states of the manufacturing process and optionally to automatically in dicate failures in the manufacturing process, such that the manufacturing process can be brought back to normal operation faster and more reliable.

According to a refinement of the present invention the at least one monitored physical quantity is a pressing force of a press and the at least one generated input signal is a force-over-time signal. Further, a failure in a pressed work- piece is indicated as the failure of the manufacturing pro cess.

According to a further refinement of the present invention the manufacturing system further comprises a press. The at least one sensor is a force sensor generating a force signal by monitoring a pressing force of the press. At least one force sensor and preferably four force sensors equally distributed over the press is/are monitoring the pressing force of the press. The at least one force sensor may be a piezo-electric sensor or the like. The at least one force sensor measures the pressing force over time and pro duces the at least one corresponding input signal.

The at least one input signal based on the pressing force of the press is transformed into the at least one parameter. The at least one parameter is provided to the MLS which has been trained on respective training parameters based on training input signals based on pressing forces. The MLS derives the latent features which are mapped into one of the several re spective clusters that represent states of the manufacturing process like a failure where a work piece formed by the press from a sheet or plate has cracks or ripples or micro-cracks etc.

With the present invention states and in particular failures of the manufacturing process like cracks, ripples, micro cracks and the like in the manufactured workpieces can be re liably and fast detected.

According to a refinement of the present invention the moni tored pressing force is a force of a plunger (or punch) of the press () and additionally or alternatively a force on a mould of the press ().

Preferably, at least four force sensors are positioned either at the mould or at the plunger. Each of the four force sen sors is arranged in one of a front left portion, a front right portion, a rear left portion or a rear right portion of the mould or the plunger, respectively.

The pressing force of the plunger or on the mould can be par ticularly effectively measured and processed. According to a refinement of the present invention the re ceived at least one input signal is transformed by selecting a predefined time slice of the signal and additionally or al ternatively a data cleansing and additionally or alternative ly a normalization and/or a centring.

Only a predefined time slice of the at least one input signal can be selected and used as the at least one parameter. For example, a pressing force over time signal may be transformed by selecting only a time slice of said signal, where there is a measured an actual pressing force and preferably a time slice of said signal, from where the plunger makes contact to the sheet or plate until the pressing is finished and the plunger leaves the work piece pressed from the sheet or the plate.

The received at least one input signal may also be processed by a data cleansing. In the data cleansing gaps, jumps and the like in the at least one received input signal are filled, corrected and the like, such that a continuous and smooth signal is provided as parameter.

With normalization, the at least one input signal is normal ised to a predefined value range (e.g. from 0 to 1) and addi tionally or alternatively the metering time period is normal ised. For example each input signal may be normalized to a predefined number of time steps of a predefined length, such that a too long signal is cut and a too short signal is ex tended.

The received at least one input signal may also be centred around a specific value or raise in the input signal. For ex ample, the input signals based on the pressing force of the press may be centred by convolving each input signal with it self. With the transformation the at least one input signal is pre pared such that the MLS can optimally derive the latent fea tures.

According to a refinement of the present invention the MLS is a neural network (NN) preferably a deep NN and most prefera bly a convolutional NN. This provides for storing specific coherent patterns of the timeseries in the latent space so that they can be distinguished.

Artificial neural networks (ANN) are systems, in particular computing systems, inspired by biological neural networks that constitute animal brains. ANNs "learn" to perform tasks by considering (labelled) examples or training data, general ly without being designed with any task-specific rules. Dur ing an initial learning or training phase ANNs automatically generate identifying characteristics from the (labelled) training data. ANNs comprise a collection of connected nodes called artificial neurons, which loosely model the neurons in a biological brain. Each connection (synapses in the biologi cal brain) can transmit a signal from one node to another. A node that receives a signal can process it and then signal to subsequent neurons connected to it. In common ANN implementa tions, the signal at a connection between nodes is a real number (e.g. 0...1), and the output of each artificial neuron is computed by some non-linear function of the sum of its in puts (from other nodes). The connections between nodes are called "edges". The edges in ANNs may each have a weight that is adjusted during training of the ANNs. The weight increases or decreases the strength of the signal at the corresponding edge. Nodes may each have a threshold such that the signal is only sent if an aggregate signal exceeds that threshold. Typ ically, nodes are aggregated into layers. Different layers may perform different kinds of transformations on their in puts. Signals travel from a first layer or input layer to a last layer or output layer, possibly after traversing the layers multiple times.

In other words, an ANN is a network of simple elements, the so called nodes or artificial neurons, which receive input. After receiving input the nodes change their internal state (activation) according to that input, and produce output de pending on the input and activation. The network forms by connecting the output of certain nodes to the input of other nodes forming a directed, weighted graph. The weights as well as the functions that compute the activation of each node can be modified during initial learning / training, which is gov erned by a learning rule or paradigm.

A node receiving an input from at least one predecessor neu ron consists of the following components: an activation, the node's state, depending on a discrete time parameter, option ally a threshold, which stays fixed unless changed by a learning / training function, an activation function (e.g. hyperbolic tangent function, sigmoid function, softmax func tion, rectifier function etc.) that computes the new activa tion at a given time and the net input and an output function computing the output from the activation (often the output function is the identity function). An important characteris tic of the activation function is that it provides a smooth transition as input values change, i.e. a small change in in put produces a small change in output.

An input node has no predecessor but serves as input inter face for the whole ANN. Similarly an output node has no suc cessor and thus serves as output interface of the whole ANN. An ANN consists of edges / connections, each edge transfer ring the output of a node (predecessor) to the input of an- other, succeeding node (successor). Additionally to the as signed weight an edge may have a bias term added to a total weighted sum of inputs to serve as a threshold to shift the activation function. The propagation function computes the input to the succeeding node (successor) from the outputs of preceding nodes (predecessors) and may include the bias val ue.

The deep NN comprises more than one layer, preferably more than four layers, more preferably more than seven layers and most preferably ten or more layers. Each layer may comprise several neurons or nodes. Preferably each layer may contain ten or more, more preferably 50 or more and most preferably 100 or more neurons.

The convolutional NN is a deep NN with convolutional layers. In the convolutional layers, there are filters that are con volved with the input. Each filter is equivalent to a weights vector that has to be trained.

The accuracy of the mapped clusters and, thus, the certainty of the state of the manufacturing process based on the de rived latent features of the deep or convolutional NN are in creased.

According to a refinement of the present invention the MLS is deployed on a cloud-based system or on a local computer sys tem of a premise where the manufacturing process is conduct ed.

The cloud-based system may be located at a side of a manufac turer of the manufacturing system, which conducts the manu facturing process, or at a side of the user of the manufac turing system. In case the MLS is deployed on the cloud-based system the MLS can be used for several different manufacturing systems and manufacturing processes of the (exact) same type.

In case the MLS is deployed on the local computer system of a premise where the manufacturing process is conducted, the MLS is exclusively used for the respective manufacturing process.

Consequently, the MLS can either be used (globally) for dif ferent manufacturing systems and for different manufacturing processes and, thus, the capacity of the MLS can be optimally used or the MLS can be locally used and, thus, the MLS can be more specialised for the respective manufacturing system or rather manufacturing process.

According to a refinement of the present invention a t- distributed Stochastic Neighbor Embedding (t-SNE) method is used to map the derived latent features into one of the sev eral distinct clusters. t-SNE is a machine learning algorithm for visualization. It is a nonlinear dimensionality reduction technique well-suited for embedding high-dimensional data for visualization in a low-dimensional space of two or three dimensions. Specifical ly, it models each high-dimensional object (here the latent features) by a 2D or three-dimensional (3D) point in such a way that similar objects are modelled by nearby points (clus ters) and dissimilar objects are modelled by distant points with high probability. Here, the latent features derived by the MLS (e.g. several hundred values) are reduced or mapped into two values resembling points in the 2D cluster space, where the different states of the manufacturing process can be discriminated based on the distinct clusters in the 2d cluster space.

According to a fifth aspect of the present invention a com puter-implemented method of training a machine learning sys tem (MLS) for indicating states of a manufacturing process comprises the steps i) generating a set of at least one training parameter and of corresponding training results and ii) training the machine learning system. In step i) a set of at least one training parameter is generated based on train ing input signals based on at least one physical quantity monitored during several manufacturing processes by means of at least one sensor. Further corresponding training results of the respective manufacturing processes are included in the set. In step ii) the MLS is trained using the set of the at least one training parameter and of corresponding training results.

A learning or rather training rule or paradigm is an algo rithm which modifies the parameters of a respective ANN, in order for a given input to the ANN to produce a favoured out put. This training typically amounts to modifying the weights and thresholds of the variables within the ANN. Given a spe cific task to solve and a class of functions, learning means using a set of observations to find the one function of the class of functions, which solves the task in some optimal sense. This entails defining a cost function such that for the optimal solution the cost is minimal and no other solu tion has a cost less than the cost of the optimal solution. The cost function is an important concept in learning, as it is a measure of how far away a particular solution is from an optimal solution to the problem to be solved. Learning algo rithms search through the solution space to find a function that has the smallest possible cost. For applications where the solution is data dependent, the cost must necessarily be a function of the observations, otherwise the model would not relate to the data. It is frequently defined as a statistic to which only approximations can be made. It is possible to define an arbitrary cost function, however, a particular cost function may be used either because it has desirable proper- ties (e.g. convexity) or because it arises naturally from a particular formulation of the problem.

An ANN can be discriminatively trained with a standard back- propagation algorithm. Backpropagation is a method to calcu late the gradient of a loss function (produces the cost asso ciated with a given state) with respect to the weights in the ANN. The weight updates of backpropagation can be done via stochastic gradient descent. The choice of the cost function depends on factors such as the learning type (e.g. super vised, unsupervised, reinforcement etc.) and the activation function. Commonly, the activation function and cost function are the softmax function and cross entropy function, respec tively.

In other words, training an ANN essentially means selecting one model from the set of allowed models (or, in a Bayesian framework, determining a distribution over the set of allowed models) that minimizes the cost. Commonly some form of gradi ent descent is deployed, using backpropagation to compute the actual gradients. This is done by simply taking the deriva tive of the cost function with respect to the network parame ters and then changing those parameters in a gradient-related direction. Backpropagation training algorithms fall into three categories: steepest descent (with variable learning rate and momentum, resilient backpropagation), quasi-Newton (Broyden-Fletcher-Goldfarb-Shanno, one step secant), Leven- berg-Marquardt and conjugate gradient (Fletcher-Reeves up date, Polak-Ribiere update, Powell-Beale restart, scaled con jugate gradient).

Common training paradigms include supervised learning, unsu pervised learning and reinforcement learning. Supervised learning uses a set of example pairs and the aim is to find a function in the allowed class of functions that matches the examples. In other words, the mapping implied by the data is inferred; the cost function is related to the mismatch be tween the mapping of the ANN and the data and it implicitly contains prior knowledge about the problem domain. The cost may be the mean-squared error, which tries to minimize the average squared error between the ANN's output and a target value over all the example pairs. Minimizing this cost using gradient descent for the class of ANNs called multilayer per- ceptrons (MLP), produces the backpropagation algorithm for training ANNs. In unsupervised learning, some data is given and the cost function to be minimized that can be any func tion of the data and the ANN's output. The cost function is dependent on the task and any a priori assumptions (e.g. im plicit properties or parameters of the model, observed varia bles etc.). In reinforcement learning, data is usually not given, but generated by an agent's interactions with the en vironment. At each point in time the agent performs an action and the environment generates an observation and an instanta neous cost according to some (usually unknown) dynamics. The aim is to discover a policy for selecting actions that mini mizes some measure of a long-term cost, e.g. the expected cu mulative cost. The environment's dynamics and the long-term cost for each policy are usually unknown, but may also be es timated. The environment is commonly modelled as a Markov de cision process (MDP) with states and actions with the follow ing probability distributions: the instantaneous cost distri bution, the observation distribution and the transition, while a policy is defined as the conditional distribution over actions given the observations. Taken together, the two then define a Markov chain (MC). The aim is to discover the policy (i.e., the MC) that minimizes the cost.

To each at least one training parameter in the generated set belongs a corresponding result. There may be at least 100, preferably at least 500 and most preferably at least 1000 pairs of training parameter and corresponding training result included in the set. The at least one training parameter cor responds to the at least one parameter described above for the first to fourth aspect of the present invention. The training input signals, which correspond to the at least one input signal described above for the first to fourth aspect of the present invention, is transformed into the at least one training parameter. The transformation is the same as de scribed above for the first to fourth aspect of the present invention (cf. step b)). For example, at least one physical quantity (e.g. a pressing force or four pressing forces at different locations of a press) is monitored with at least one respective sensor during several cycles (e.g. 1000 cy cles) of a manufacturing process. At each cycle the at least one sensor generates a training input signal corresponding to the monitored physical quantity (e.g. a force-over-time sig nal). Further, in each cycle the state of the manufacturing process is determined (e.g. normal operation, failure of the manufacturing process (cracks in a workpiece etc.) and the like). The determined state of the manufacturing process is the trainings result for the respective trainings input sig nal or rather trainings parameter.

Based on the generated set of the at least one training pa rameter and of corresponding training results the MLS is trained. Thereto, the MLS is provided with one training pa rameter of the set and the MLS is optimised until the train ing result can be identified in a cluster space into which the latent features generated by the MLS have been mapped (e.g. by a t-SNE method).

The MLS used for deriving the latent features needs to learn cues in patterns of the provided at least one parameter. Therefore, the generated set contains training parameters based on training input signals generated during a certain failure of the manufacturing process and also based on train ing input signals generated during normal operation. Further, the set may contain training parameters based on training in put signals generated during idle times and/or a start-up phase. The set of the at least one training parameter is "la belled" with the respective states of the manufacturing pro cess (trainings results) present while the corresponding training input signal was generated (by the respective sen sor (s)).

According to a refinement of the present invention the train ing input signals are real input signals from a real machine of the manufacturing system or manufacturing process or simu lated input signals from an artificial model of the machine of the manufacturing system or manufacturing process or a combination of both.

Thus, only a small set of "real" training parameters and cor responding training results needs to be generated from real (test) cycles of the manufacturing process. This significant ly reduces the cost for generating the set. Further, "incom plete" sets (e.g. training signals during a specific state of the manufacturing process like a certain failure of the manu facturing process are missing) can be complemented with vir tually generated training signals during specific states of the manufacturing process adjusted in the model.

According to a refinement of the present invention the MLS is a neural network (NN) and preferably a pre-trained NN.

Especially by training a pre-trained neural network that has been generally conditioned for deriving information from in dustrial data the time and amount of sets of training fea ture (s) can be significantly reduced.

According to a refinement of the present invention the train ing input signals are transformed into the at least one training parameter of the set by selecting a predefined time slice of the training input signals and/or a data cleansing and/or a normalization and/or a centring and/or using parts of predefined length of the training input signals, where preferably, the parts of the predefined length of the train- ing input signals are selected according to a Gaussian dis tribution.

The selecting of the predefined time slice of the signal, the data cleansing, the normalization and the centring are the same as for the at least one input signal as described above for the first to fourth aspect of the present invention.

Instead of the whole training input signal only parts of the predefined length may be used as parameters for training the MLS. Preferably, the predefined length of the parts is equal to or smaller than 10% and most preferably equal to or small er than 2% of the length of the training signal. For example, a training input signal (e.g. a force-over-time-signal) that has been normalised to 1000 time steps, where each time step has the same predefined time length, is divided into the parts of the predefined length. For example the normalised training input signal is divided into parts of a length of 16 time steps. For training the MLS only those "important" parts of the training input signal are used as training parameter, which contain the most information about the state of the manufacturing process or rather the trainings result.

Preferably, the parts of one single training input signal may be used several times for training the MLS.

Further preferably, the parts of the training input signals may be selected according to a Gaussian distribution and used as the training parameter. In selecting the parts of the training input signals based on the Gaussian distribution the maximum of the Gaussian distribution may be placed at the part of the input training signals, where the most infor mation about the state of the manufacturing process is con tained. Thereby, the training of the MLS can be accelerated and the accuracy of the trained MLS may also be improved.

The present invention and its technical field are subsequent ly explained in further detail by exemplary embodiments shown in the drawings. The exemplary embodiments only conduce bet ter understanding of the present invention and in no case are to be construed as limiting for the scope of the present in vention. Particularly, it is possible to extract aspects of the subject-matter described in the figures and to combine it with other components and findings of the present description or figures, if not explicitly described differently. Equal reference signs refer to the same objects, such that explana tions from other figures may be supplementarily used.

FIG 1 shows a schematic flow-chart of the computer- implemented method of indicating a failure of a manufacturing process.

FIG 2 shows a schematic view of the data processing sys tem for indicating a failure of a manufacturing process.

FIG 3 shows a schematic view of a computer-readable medi um having stored thereon the computer program prod uct for indicating a failure of a manufacturing process.

FIG 4 shows a schematic view of the manufacturing system comprising the data processing system for indicat ing a failure of a manufacturing process and a press.

FIG 5 shows a schematic view of the computer-implemented method of training a machine learning system for indicating states of a manufacturing process.

In FIG 1 the computer-implemented method of indicating a failure of a manufacturing process is schematically depicted. Here the manufacturing process is forming work pieces by pressing of sheets or plates (e.g. made of steel).

The computer-implemented method comprises the steps a) re ceiving 1 at least one input signal, b) transforming 2 the at least one input signal, c) deriving 3 latent features, d) mapping 4 the derived latent features and e) optionally indi cating 5 a failure of the manufacturing process. In step a) the at least one input signal is received. Here, five input signals II..15 are received. Four of the five in put signals II..14 are force-over-time signals based on a pressing force over time of a plunger of a press on a mould of the press measured by four respective force sensors. The last of the five input signals 15 is a position-over-time signal based on a position over time of the plunger measured by a respective position sensor.

In step b) the five input signals II..15 are transformed. The step b) comprises for each input signal II..15 the sub-steps: selecting 2.1 a predefined time slice of the signal; data cleansing 2.2; normalization 2.3; and centring 2.4.

The sub-steps 2.1..2.4 may be applied in any order to the signals II..15. Each input signal II..15 is appropriately cut by selecting 2.1 the predefined time slice of the signal. On ly the time slice of the input signals II..15 where a press ing force or movement is present is selected. Then the time slices are cleansed, whereby gaps or jumps in the curve pro gression of the signals or rather selected time slices are filled or corrected such that a continuous and smooth curve progression is present in the signals/selected time slices. The time slices are further normalised. The in the normaliz ing 2.3 the values may be normalised to a predefined range (e.g. 0..1) or the time scale may be normalised to a prede fined number of time steps or both. Here, the time slices are all normalised to 1000 time steps of predefined length (e.g. a time step has a time length of 1ms [Millisecond]). Also the curve progressions of the signals/time slices are centred, such that in each signal/time slice the pressing force or po sition starts to change at the same time step. The centring 2.4 is done by convolving each signal or rather time slice with itself. The selected, cleansed, normalised and centred signals/time slices are forwarded as the parameters P1..P5. In step c) the latent features LF are derived from the param eters P1..P5 of step b). The five parameters P1..P5 (four force-based and one position-based parameter) are provided to a convolutional neural network (NN). The convolutional NN has one input layer, five hidden layers and one output layer. The parameters P1..P5 are provided to the input layer. The convo lutional NN has been trained to derive latent features LF based on the parameters P1..P5, which latent features LF pro vide information about a state of the manufacturing process (e.g. normal operation or a certain failure of the manufac turing process). The information about the manufacturing pro cess contained in the parameters P1..P5 is extracted and new ly combined into the latent features LF by the convolutional NN. Thereto, in each layer of the convolutional NN trained weights are applied. The derived latent features LF have a specific pattern indicative of the state of the manufacturing process. The latent features LF may either be provided by the output layer or by one of the hidden layers or a combination of both.

In step d) the derived latent features LF are mapped with a t-SNE algorithm to two values resembling a point in a two di mensional (2D) cluster space. With the t-SNE algorithm simi lar patterns of the latent features LF of different cycles of the manufacturing process are grouped together and different patterns of the latent features LF are put into other groups in the 2D cluster space. Thereby, clusters C1..C3 of latent features LF are generated, where each cluster C1..C3 resem bles one specific state of the manufacturing process. With the mapped cluster C1..C3 the state of the manufacturing pro cess becomes apparent.

In optional step e) a failure of the manufacturing process is indicated, for example, to a user like a technician, mechanic personnel, shift supervisor, facility engineer etc. The fail ure is determined based on the respective cluster C1..C3, which indicates the respective state of the manufacturing process and to which the pattern of latent features LF de rived by the convolutional NN belongs.

The computer-implemented method of indicating a failure of a manufacturing process may be provided in form of the computer program product for indicating a failure of a manufacturing process.

In FIG 2 the data processing system 10 for indicating a fail ure of a manufacturing process is schematically depicted. The data processing system 10 is arranged and configured to exe cute the computer-implemented method of indicating a failure of a manufacturing process of FIG 1.

The data processing system 10 may be a personal computer (PC), a laptop, a tablet, a server, a distributed system (e.g. cloud system) and the like. The data processing system 10 comprises a central processing unit (CPU) 11, a memory having a random access memory (RAM) 12 and a non-volatile memory (MEM, e.g. hard disk) 13, a human interface device (HID, e.g. keyboard, mouse, touchscreen etc.) 14 and an out put device (MON, e.g. monitor, printer, speaker, etc.) 15.

The CPU 11, RAM 12, HID 14 and MON 15 are communicatively connected via a data bus. The RAM 12 and MEM 13 are communi catively connected via another data bus. The computer program product for indicating a failure of a manufacturing process can be loaded into the RAM 12 from the MEM 13 or another com puter-readable medium. According to the computer program product the CPU executes the steps a) to d) and optionally e) of the computer-implemented method of indicating a failure of a manufacturing process of FIG 1. The execution can be initi ated and controlled by a user via the HID 14. The status and/or result of the executed computer program may be indi cated to the user by the MON 15. The result of the executed computer program may be permanently stored on the non volatile MEM 13 or another computer-readable medium. In particular, the CPU 11 and RAM 12 for executing the com puter program may comprise several CPUs 11 and several RAMs 12 for example in a computation cluster or a cloud system.

The HID 14 and MON 15 for controlling execution of the com puter program may be comprised by a different data processing system like a terminal communicatively connected to the data processing system 10 (e.g. cloud system).

In FIG 3 the computer-readable medium 20 having stored there on the computer program product for indicating a failure of a manufacturing process is schematically depicted.

Here, exemplarily a computer-readable storage disc 20 like a Compact Disc (CD), Digital Video Disc (DVD), High Definition DVD (HD DVD) or Blu-ray Disc (BD) has stored thereon the com puter program product for indicating a failure of a manufac turing process. However, the computer-readable medium may al so be a data storage like a magnetic storage/memory (e.g. magnetic-core memory, magnetic tape, magnetic card, magnet strip, magnet bubble storage, drum storage, hard disc drive, floppy disc or removable storage), an optical storage/memory (e.g. holographic memory, optical tape, Tesa tape, Laserdisc, Phasewriter (Phasewriter Dual, PD) or Ultra Density Optical (UDO)), a magneto-optical storage/memory (e.g. MiniDisc or Magneto-Optical Disk (MO-Disk)), a volatile semiconduc tor/solid state memory (e.g. Random Access Memory (RAM), Dy namic RAM (DRAM) or Static RAM (SRAM)), a non-volatile semi conductor/solid state memory (e.g. Read Only Memory (ROM), Programmable ROM (PROM), Erasable PROM (EPROM), Electrically EPROM (EEPROM), Flash-EEPROM (e.g. USB-Stick), Ferroelectric RAM (FRAM), Magnetoresistive RAM (MRAM) or Phase-change RAM).

In FIG 4 the manufacturing system 30 comprising the data pro cessing system 10 for indicating a failure of a manufacturing process and a press 40 is schematically depicted. The manu facturing system 30 further comprises four force sensors S1..S4 and one position sensor (not depicted). The press 40 comprises a plunger (or punch) 41 and a mould (or die) 42. The four force sensors S1..S4 are arranged in a front left portion, a front right portion, a rear left por tion and a rear right portion of the mould 42. The plunger 41 may be driven by a motor via an eccentric or by a pneumatic or hydraulic drive. The plunger 41 is moved towards the mould 42 with a pressing force F such that a sheet or plate 50 (e.g. made of steel) is pressed into the mould 42 by the plunger 41. Thereby, the sheet/plate 50 is deformed into a work piece by the pressing force F applied via the plunger 41. After the work piece has been formed from the sheet/plate 50 the plunger 41 is retracted. During this operation of the plunger 41, the four force sensors S1..S4 detect the pressing force F applied by the plunger 41 onto the mould 42 and gen erate corresponding force-over-time signals that are provided to the data processing system 10 for indicating a failure of a manufacturing process of FIG 2 as the input signals II..14. Further the position sensor determines the position of the plunger 41 relative to the mould 42 during the operation and generates a position-over-time signal that is provided to the data processing system 10 for indicating a failure of a manu facturing process of FIG 2 as the input signal 15.

In FIG 5 the computer-implemented method of training a ma chine learning system for indicating states of a manufactur ing process is schematically depicted. The computer-implemen ted method of training a machine learning system for indicat ing states of a manufacturing process comprises the steps i) generating 101 a set of at least one training parameter and of corresponding training results and ii) training 102 the machine learning system.

Here, five training parameters are used. The training parame ters are generated much like the parameters in the computer- implemented method of FIG 1. In a manufacturing system (e.g. the manufacturing system of FIG 4) several cycles of the man ufacturing process are run in order to generate training in put signals T1..T5. Here, 1000 training cycles are run, but the training input signals T1..T5 may also be acquired during productive operation. In each (training or productive) cycle of the manufacturing process the pressing force of a plunger on a mould while pressing a sheet or plate into a workpiece is detected by four force sensors. Further, the position of the plunger relative to the mould is detected by a position sensor during each cycle. The four force sensors generate force-over-time signals T1..T4 and the position sensor gener ates a position-over-time signal T5 in each cycle. The four force-over-time signals and the position-over-time signal of each of the 1000 cycles are provided as the training input signals T1..T5 of the set. Further, the state of the manufac turing process at each cycle is determined. These states of the manufacturing process comprise idle phase, start-up phase, normal operation and one or more failures of the manu facturing process. The failures of the manufacturing process may comprise cracks in the sheet/plate or rather workpiece, ripples of the sheet/plate or rather workpiece, micro-cracks of the sheet/plate or rather workpiece and the like after the pressing. These states of the manufacturing process are in cluded in the set as the corresponding training results. Thereby, respective five training input signals T1..T5 and the corresponding training result (state of the manufacturing process) form one pair of training data (the five training input signals as input data for the training and the corre sponding training result as (desired) output data for the training). The five training input signals T1..T5 may be la belled or tagged with the corresponding training result.

The generating 101 comprises the sub-steps: selecting 101.1 a predefined time slice of the training signal; data cleansing 101.2; normalization 101.3; centring 101.4; and using 101.5 parts of predefined length of the training input signals, where the parts of the predefined length of the training input signals are selected according to a Gaussian distribution.

The sub-steps 101.1..101.4 may be applied in any order to the training input signals T1..T5. The five training input sig nals T1..T5 are transformed as described for the input sig nals in the computer-implemented method of FIG 1. The sub steps 101.1..101.4 correspond to the sub-steps 2.1..2.4. Fur ther, the five input training signals T1..T5 or rather the selected time slices may not be used completely, but only parts of the predefined length may be used as training param eters. Thereto, the training input signals T1..T5 or rather selected time slices thereof may be divided into the parts of the predefined length. Here the time slices of 1000 time steps are divided into parts of 16 time steps.

Further, the part of the time slice of each training input signal T1..T5 that is used as the respective training parame ter is selected according to a Gaussian distribution. The maximum of the Gaussian distribution is placed in a region of the training input signals Tl..T5/time slices, where a maxi mal pressing force is detected. This region of the training input signals Tl..T5/time slices contains the most infor mation about the state of the manufacturing process. The cor responding parts are therefore more often selected as the in put parameters via the Gaussian distribution than the parts belonging to regions of the training input signals Tl..T5/time slices, where the plunger applies a lower press ing force onto the mould (e.g. from the time point where the plunger contacts the sheet/plate until before the plunger ex erts the maximal pressing force as well as from the time point where the plunger decreases the pressing force and is retracted from the mould).

In step ii) a convolutional NN is trained for indicating states of the manufacturing process. Iteratively the pairs of input data (five training parameters) and (desired) output data (corresponding training results) are used for training the convolutional NN. In each iteration, the five training parameters, which are based on the respective five training input signals T1..T5, are provided to the convolutional NN. The internal weights of the hidden layers of the convolution al NN are adjusted until the optimisation function converges to the corresponding training result. After the set of 1000 pairs of input data and output data have been used for train ing, the convolutional NN is sufficiently trained for indi cating states of the manufacturing process.

The convolutional NN trained by the computer-implemented method of training a machine learning system for indicating states of a manufacturing process of FIG 4 may be used in the computer-implemented method of indicating a failure of a man ufacturing process of FIG 1 and/or integrated in the data processing system for indicating a failure of a manufacturing process of FIG 2, which may be included in the manufacturing system of FIG 4.

Although specific embodiments have been illustrated and de scribed herein, it will be appreciated by those of ordinary skill in the art that a variety of alternate and/or equiva lent implementations exist. It should be appreciated that the exemplary embodiment or exemplary embodiments are only exam ples, and are not intended to limit the scope, applicability, or configuration in any way. Rather, the foregoing summary and detailed description will provide those skilled in the art with a convenient road map for implementing at least one exemplary embodiment, it being understood that various chang es may be made in the function and arrangement of elements described in an exemplary embodiment without departing from the scope as set forth in the appended claims and their legal equivalents. Generally, this application is intended to cover any adaptations or variations of the specific embodiments discussed herein.

In the foregoing detailed description, various features are grouped together in one or more examples for the purpose of streamlining the disclosure. It is understood that the above description is intended to be illustrative, and not restric tive. It is intended to cover all alternatives, modifications and equivalents as may be included within the scope of the invention. Many other examples will be apparent to one skilled in the art upon reviewing the above specification.

Specific nomenclature used in the foregoing specification is used to provide a thorough understanding of the invention. However, it will be apparent to one skilled in the art in light of the specification provided herein that the specific details are not required in order to practice the invention. Thus, the foregoing descriptions of specific embodiments of the present invention are presented for purposes of illustra tion and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed; ob viously many modifications and variations are possible in view of the above teachings.

The embodiments were chosen and described in order to best explain the principles of the invention and its practical ap plications, to thereby enable others skilled in the art to best utilize the invention and various embodiments with vari ous modifications as are suited to the particular use contem plated. Throughout the specification, the terms "including" and "in which" are used as the plain-English equivalents of the respective terms "comprising" and "wherein, " respective ly. Moreover, the terms "first," "second," and "third," etc., are used merely as labels, and are not intended to impose nu merical requirements on or to establish a certain ranking of importance of their objects. In the context of the present description and claims the conjunction "or" is to be under stood as including ("and/or") and not exclusive ("either ... or").