Description

The invention relates to a method and system for predicting the outside temperature of a rotary kiln shell for at least one slice n of the rotary kiln.

WO 2020/188549 A2 is concerned with a method and system for monitoring and optimizing the operation of an alumina rotary kiln, and discloses a temperature prediction unit 116 which is configured to predict a gas and a solid temperature profile at preselected locations in the alumina rotary kiln 102 over a predefined time period using the preprocessed real time data and a thermal model stored in a database 126, whereas the thermal model may comprise a first principles driven model, a data-driven model or a knowledge driven model.

EP 1 228 016 B1 is concerned with a kiln plant control system and discloses using a control matrix which includes values determinative of the relationships between the operating parameters and plant measurements, such as, e.g., a hood temperature (Thood) as disclosed in Fig. 6.

J. Xu, D. Fu, L. Shao, X. Zhang and G. Liu, "A Soft Sensor Modeling of Cement Rotary Kiln Temperature Field Based on Model-Driven and Data-Driven Methods," (in IEEE Sensors Journal, vol. 21 , no. 24, pp. 27632-27639, 15 Dec.15, 2021 , doi: 10.1109/JSEN.2021 .3116937) discloses a soft sensor taking the axial air speed, swirling air speed, coal mass flow, material mass flow, secondary air temperature, and x, y, z coordinates as the inputs, while the temperature at a certain position inside the kiln is the output..

During operation of a rotary kiln, it is essential to monitor the status of such a kiln in terms of the outside temperature of a rotary kiln shell. In case this temperature becomes too high, significant damage can occur to such a rotary kiln, e.g., due to (partial) melting of the shell in regions where the outside temperature exceeds a certain threshold. Such excessive temperatures can occur e.g., due to wear of the inside refractories, which might result in an increased thermal load to the shell.

The wear of the inside refractories in a rotary kiln cannot easily be monitored during operation, only after shutting down of the kiln an inside inspection can take place. Therefore, knowledge of different wear mechanisms and other processes in rotary kilns are sparse. Additionally other influences on the outside shell temperature, apart from the refractory wear, might be present, such as, e.g., regions in a rotary kiln which possess a higher internal temperature, e.g., caused by exothermic chemical reactions or caused directly by the vicinity of the burner, e.g., in the flame zone, where the temperature might also depend on the fuel, the fuel mix or the fuel volume used. Currently the outside shell temperature is usually monitored by a number of temperature sensors, and whenever the temperature reaches a certain upper temperature threshold, the kiln is shut down and the refractory lining inside the kiln is renewed. Further it is difficult to predict the outside shell temperature for the future production, e.g., in the following days or weeks. Thus, for security reasons, the upper temperature threshold is set to a value significantly below the break down temperature of the shell. In turn, the rotary kiln campaign often ends prematurely, not utilizing the refractory lining to its full potential, thereby increasing operational costs and decreasing productivity. In other words, a prediction of the outside kiln temperature would allow to maintain a rotary kiln in operation for a longer period of time before a renewal of the refractory lining is necessary. Additionally, also a lower temperature threshold is of interest for the operator, for example in cases where the temperature suddenly drops significantly (e.g., a peak down in temperature), as this can lead to an increased thermoshock, resulting in an increased wear of the refractories.

The inventors have realized, that for a reliable prediction of the outside temperature of a rotary kiln shell for at least one slice n of the rotary kiln, a model should be used which is based on a machine learning algorithm being trained with a training feature data set and a training outside temperature data set. Such a machine learning algorithm in a trained state can reliably predict an outside temperature of a rotary kiln shell for at least one slice n of the rotary kiln in the near future. This reliable prediction may be caused by the fact that a machine learning algorithm trained in such a way does not need any prior knowledge to be fed into such an algorithm or a model. Therefore, prior art models based on any physical models, thermodynamical variables, first principles driven models, data- driven models or knowledge driven models have regularly failed in the actual application, which is believed by the inventors to be caused by a lack of full understanding of the complexity of all processes and wear mechanisms taking place in a rotary kiln. The usage of a machine learning algorithm in the context of the present invention does not need any input in regard to such internal wear mechanisms or processes, as the resulting machine learning model is achieved by training the model with a training feature data set and a training outside temperature data set. The inventors have further realized that such models can be trained to very high accuracies when training is performed with a large outside temperature data set in terms of such data having a high spatial resolution on the outside of a rotary kiln shell. The inventors have realized that such large temperature data sets can preferably be provided by using one or more infrared cameras directed at the rotary kiln shell, and able to provide a large number of temperature readings during operation of a rotary kiln.

Therefore, it is an object of the invention to provide a method and a system for predicting the outside temperature of a rotary kiln shell for at least one slice n of the rotary kiln.

The object is achieved by a method according to claim 1 and by a prediction unit according to claim 12 and a prediction system according to claim 14 and a rotary kiln system according to claim 15.

The core idea of the invention is based on the finding, that machine learning can be used to determine the future outside temperature of a rotary kiln shell for at least one slice n of the rotary kiln. This can be achieved especially due to the selection of a machine learning algorithm, feeding this algorithm with a selection of features of the at least one slice n of the rotary kiln and by further training this machine learning algorithm to achieve a model. The model can then be used to predict an outside temperature for the at least one slice n of the rotary kiln based on a prediction feature data set.

The term “predicting”, as used herein, is defined as a declaration or indication in advance, in other words, predicting an outside temperature data set (T _n (t), t > 0) involves foretelling the (future) outside temperature data set in advance.

The term “machine learning”, as used herein comprises the use and development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyze and draw inferences from patterns in data.

The term “machine learning algorithm”, as used herein, is defined as an algorithm (an algorithm being a set of rules that must be followed when solving a particular problem) used in machine learning applications. Examples of such machine learning algorithms include DT (Decision Tree Regression), RF (Random Forest Regression), SVR (Support Vector Regression), NN (Neural Network). Machine learning algorithms build a model based on sample data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Generally, in a slice n the associated algorithm is denoted by A _n .

The term “supervised learning”, also referred to as supervised training, as used herein, might comprise use of labeled data sets to train algorithms that classify data or predict outcomes accurately. Supervised learning, also known as supervised machine learning, is a subcategory of machine learning. Supervised learning is the machine learning task of learning a function that maps an input to an output based on example input-output pairs. In the context of the present invention the input-output pair might comprise values for features for the at least one slice n of the rotary kiln and temperature data for the at least one slice n of the rotary kiln.

The term “model parameters”, as used herein, might comprise parameters in the model that must be determined by performing a training step, such as supervised learning (by using an input-output pair). These are fitted parameters of the model. Generally, in a slice n there can be a set of K model parameters for a certain Algorithm A _n , denoted by

The term “hyperparameters” as used herein, might comprise adjustable parameters that can be tuned to obtain a certain model. They are parameters whose values are used to control the learning process. In contrast, the values of the model parameters are derived via training. The obtained model will depend on the actual used hyperparameters. For example, the hyperparameters depend on the actually used machine learning algorithm. For a DT the hyperparameters might comprise the depth of the tree or split criteria. Generally, in a slice n there can be a set of J hyperparameters for a certain Algorithm A _n , denoted by

The term “machine learning model”, as used herein, might comprise a data file that comprises trained data in order to achieve a good fit to the output part of the inputoutput pair. This file is generated by using a machine learning algorithm with respective hyperparameters after supervised training.

The term “rotary kiln” (“Drehrohrofen”), as used herein, comprises a cylindrical vessel with a longitudinal axis L, the longitudinal axis L of the cylindrical vessel being slightly inclined to the horizontal, the cylindrical vessel rotating around its longitudinal axis L. The cylindrical vessel has an (outer) shell, usually made of a metal (called “rotary kiln shell”), which is lined with refractory bricks on the inside (that is between the inside of the rotary kiln shell and the longitudinal axis L of the rotary kiln). The rotary kiln shell comprises a surface on its side facing away from the longitudinal axis L of the rotary kiln (called “Surface of rotary kiln shell”), where the rotary kiln shell has its outside temperature (“the outside temperature of a rotary kiln shell”). The cylindrical vessel is heated (fired), usually by a burner at the lower end of the cylindrical vessel. During rotation of the cylindrical vessel, material moves gradually downwards from the upper end to the lower end of the cylindrical vessel. The material undergoes continuous stirring and mixing while being heated by the hot flame I hot gas from the burner. Usual application of a rotary kiln includes increasing the temperature of a material, e.g., for calcination in a continuous process. The term “the outside temperature at the surface of a rotary kiln”, as used herein, comprises the temperature measurements at the outside of the rotary kiln, which is at the surface of the rotary kiln shell. This temperature can be measured at this outside of the object by means of temperature measurement devices, e.g., thermal sensor(s), or contactless, e.g., by infrared camera(s) or alike.

The term “slice n of the rotary kiln” as used herein, comprises a section of the rotary kiln, i.e., a section of the cylindrical vessel. This section is preferably a cylinder, e.g., a right cylinder, whereby the cylinder defining the slice is cut out of the cylindrical vessel. In other words, the slice is preferably a cylinder with its longitudinal axis coinciding with the longitudinal axis L of the cylindrical vessel. The slices are numbered, such that slice n is the n-th (e.g., the 1 ^st , 2 ^nd , 3 ^rd ...) slice of the rotary kiln. Slices can be quite short (that is they have a short longitudinal dimension), e.g., in the width (which is to be understood as the dimension of the brick along the longitudinal axis L in a mounted form) of a single brick (such as 200 mm) or even a share of a single brick width (e.g. 1/10 ^th of a brick width, such as 20 mm), and preferably a slice comprises at least one single temperature measurement. In several regions the slices might be longer, e.g., up to 5 (such as 1000 mm) or even 10 brick widths (such as 2000 mm).

The term “feature for the at least one slice n of the rotary kiln” as used herein, comprises any input value or data extracted from the production and I or laboratory information management system of the rotary kiln. For example, the features can comprise process features such as the flame temperature, inlet chamber temperature, torque of the rotary kiln, energy consumption, as well as chemical features, such as alkali content, chloride content, sulfur content of the hot meal and I or clinker.

The term “temperature data” as used herein, comprises a machine-readable data structure, e.g., in the form of a single temperature value, or a set of temperature values in the form of a field / vector, or a matrix containing temperature values.

A prediction unit is understood to mean one or more devices for carrying out the respective method steps described below, and which, for this purpose, comprise either discrete electronic components in order to process signals, or which are implemented partially or completely as a computer program in a computer.

In a first aspect of the invention, the object is achieved by providing a method for predicting an outside temperature (T _n (t), t>0) of a rotary kiln shell for at least one slice n of the rotary kiln, with the steps:

(a) a training feature providing step for providing a training feature data set preferably comprising values for M* features of the rotary kiln;

(b) a training temperature providing step for providing a training outside temperature (T _n (t), t < 0) of a rotary kiln shell for the at least one slice n of the rotary kiln as the training outside temperature data set (T _n (t), t < 0);

(c) a model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln, comprising:

(i) an algorithm selection step for selecting at least one algorithm An for the at least one slice n of the rotary kiln from a machine learning algorithm, the at least one algorithm A _n comprising model parameters and hyperparameters;

(ii) an algorithm training step for training the at least one algorithm A _n for the at least one slice n of the rotary kiln by using the training feature data set and the training outside temperature data set (T _n (t), t < 0) leading to the at least one algorithm A _n for the at least one slice n of the rotary kiln in a trained state;

(iii) the algorithm training step is based on supervised learning;

(iv) whereby a prediction model A _n * is achieved, which comprises the at least one algorithm A _n for the at least one slice n of the rotary kiln in a trained state, and a set of hyperparameters

(d) a prediction feature providing step for providing a prediction feature data set preferably comprising values for the M features for the at least one slice n of the rotary kiln;

(e) a temperature prediction step for predicting an outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln based on the prediction feature data set by using the prediction model A _n *

Preferably the method predicts an outside temperature (T _n (t), t > 0) of a rotary kiln shell (i.e. , the outside temperature at the surface of the rotary kiln) for at least one slice n of the rotary kiln. Preferably the at least one slice n of the rotary kiln may cover all of a specific zone of the rotary kiln, such as an outlet zone, a lower transition zone, a burning zone, an upper transition zone, a safety zone, a calcining zone or an inlet zone. Alternatively, the method may predict an outside temperature for the whole rotary kiln shell (T _n (t), t > 0, n = 1 ,...,N).

Preferably the method runs at a certain point in time (here at t = 0) and can predict an outside temperature at that time or in the future (t > 0).

Preferably, a training feature data set preferably comprising values for M* features of the rotary kiln, is provided in a training feature providing step. The training feature data set is collected in the past, that is at a certain time t < 0, i.e., before the execution of the method. The number M* of such features can be in the range of 5 to 500, preferably 10 to 400, most preferably 50 to 300. The features are preferably selected from laboratory values or properties of clinker, kiln feed, hot meal, fuel type or from operational parameters of the kiln, such as pressure or flame temperature. The training feature data set comprises time series of values of each feature at specific times t < 0: t ₁₍ t _K . The times t ₁₍ t _K are preferably selected from a time interval between just before the execution of the method up to certain time in the past, such as at least 50 days in the past, preferably at least 100 days in the past, more preferably at least 300 days in the past, most preferably 500 days in the past.

Preferably, a training outside temperature data set (T _n (t), t < 0) comprising outside temperature data of a rotary kiln shell for the at least one slice n of the rotary kiln is provided in a training temperature providing step. The outside temperature data set T _n (t) is collected in the past, that is at a certain time t < 0, i.e. , before the execution of the method. The outside temperature data set T _n (t) may comprise a single temperature value per slice n of the rotary kiln at a certain time t or alternatively may comprise multiple temperature values per slice n of the rotary kiln at a certain time t. The outside temperature data set T _n (t) comprises time series of values of each temperature at times t < 0: t ₁₍ ... , t _K . The times t ₁₍ ... , t _K are preferably selected from a time interval between just before the execution of the method up to certain time in the past, such as at least 50 days in the past, preferably at least 100 days in the past, more preferably at least 300 days in the past, most preferably 500 days in the past. The time t is preferably selected from the same time interval as for the training feature data set Most preferably the outside temperature data set T _n (t) and the training feature data set comprise temperature data and values for M* features taken at the same time t. In other words, for all times t _k < 0 k = 1, ..., k there exist pairs of and T at the same time tk.

Preferably a pre-processing step for pre-processing of at least one data set is performed, whereas the at least one data set is selected from the list: a training feature data set training outside temperature data set (Tn(t), t < 0).

Preferably the pre-processing step for pre-processing of at least one data set, such as a training feature data set and or a training outside temperature data set (T _n (t), t < 0), comprises: checking of the data format and data structure of the data set, and I or checking of the consistency of the data set and I or outlier extraction of the data set, and I or removal of duplicates in the data set and / or missing value imputation. This pre-processing allows to achieve a faster convergence of the algorithm training step and thus saves computational time.

Preferably the pre-processing step for pre-processing of at least one data set comprises: scaling of the values in the data set. With this scaling the convergence of the algorithm training step is significantly faster and thus saves additional computational time. Further to that, scaling leads to a significant reduction of memory required for storing and handling the training feature data set

Preferably a model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln is performed.

Preferably the model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln further comprises a scoring function setup step for setting up a scoring function E _n (t) for the at least one slice n of the rotary kiln, the scoring function being selected from: RMSE (Root Mean Square Error), trend, weighted scoring function. The scoring function E _n (t) may be a function of at least one algorithm A _n , selected in an algorithm selection step, and the training outside temperature data set (T _n (t), t < 0). In other words, the scoring function E _n (t) may provide a score by comparing the output of at least one algorithm A _n (during the training phase, at t < 0) to the training outside temperature data set (T _n (t), t < 0).

As the training outside temperature data set (T _n (t), t < 0) is the target of the algorithm training step, the scoring function E _n (t) provides a measure on the accuracy of a certain algorithm A _n in a certain state. Preferably the scoring function E _n (t) is selected as the RMSE calculated at time t, in that for all t < Preferably the scoring function E _n (t) is selected as the trend calculated at time t, in that for all t < 0: t ₁₍ t _K , E _n (t) = Preferably the scoring function E _n (t) is selected as the weighted scoring function calculated at time t, in that for all t < Preferably a model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln, comprises an algorithm selection step for selecting at least one algorithm A _n for the at least one slice n of the rotary kiln from a machine learning algorithm, the at least one algorithm A _n comprising model parameters and hyperparameters The algorithm A _n is preferably set up such that an input, such as the training feature data set will provide an output A _n

Preferably in the algorithm selection step, the selected machine learning algorithm is selected from any of: tree-based machine learning algorithms, such as Decision Tree (DT), Random Forest (RF), gradient boosting regression; vector based machine learning algorithms, such as Support Vector Regression (SVR); neural network based, such as a simple Neural Network (NN); ensembled machine learning algorithms such as voting regression.

Preferably the model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln further comprises an algorithm feature selection step before a subsequent algorithm training step, whereas the algorithm feature selection step comprises the selection of M features from the training feature data set preferably comprising values for M* features for the at least one slice n of the rotary kiln, and providing a selected training feature data set preferably comprising values for M features for the at least one slice n of the rotary kiln. Preferably the selected training feature data set ( comprises a reduced number of M features compared to the training feature data set preferably comprising values for M* features. Preferably the reduced number of M features is less than 50% of the number M* more preferable less than 40%, most preferable less than 30%. Preferably the algorithm feature selection step comprises at least one algorithm selected from: correlation-based feature selection (CBFS), mutual information-based feature selection (MIFS), sequential forward selection (SFS), sequential backward selection (SBS). This step will in all cases reduce the runtime for the algorithm training step. Additionally, this step may reduce the number of features used in a subsequent algorithm training step in a way, that the accuracy of the results of the subsequent training step is not compromised, but maybe is even improved. Preferably the model setup step for setting up at least one model An* for the at least one slice n of the rotary kiln further comprises an algorithm training step for training the at least one algorithm A _n for the at least one slice n of the rotary kiln by using the training feature data set and the training outside temperature data set (T _n (t), t < 0) leading to the at least one algorithm An for the at least one slice n of the rotary kiln in a trained state. Preferably the algorithm training step is based on supervised learning. Preferably a prediction model A _n * is achieved, which comprises the at least one algorithm A _n for the at least one slice n of the rotary kiln in a trained state, and a set of hyperparameters . It has been found by the inventors that such a use of a machine learning algorithm, especially when trained by supervised learning, leads to very good predictions of the outside temperature of a rotary kiln shell for at least one slice n of a rotary kiln. The algorithm training step is preferably performed by a variation of the set of K model parameters 1, … , ^) of the algorithm A _n , such that the the training feature data set 1,...,M, t < 0, M ≤ M*) as the input to the algorithm A _n leads to an output of algorithm An that approximates the training outside temperature data set 0). In other words, the scoring function E _n (t) is optimized (here: minimized) by a variation of the set of K model parameters( ^ of the algorithm An. When the scoring function has reached a certain (minimum) target, such that the scoring function is below a scoring function threshold En(t) < Emin, the variation of the set of K model parameters of the algorithm A _n is stopped and the algorithm An is in a trained state, that is when the algorithm An is applied with the thereby achieved final set of K model parameters of the algorithm An. Thus a prediction model may be achieved, which comprises the at least one algorithm An for the at least one slice n of the rotary kiln in a trained state, which comprises a final set of K model parameters as well a set of hyperparameters ( of the algorithm Preferably a prediction feature providing step for providing a prediction feature data se preferably comprising values for the M features for the at least one slice n of the rotary kiln, is performed. Preferably the M features of the prediction feature data set consists of the same M features of the training feature data set

Preferably the prediction feature providing step for providing a prediction feature data set comprises providing the prediction feature data set based on the production plan of the rotary kiln; or providing the prediction feature data set based on a continuation of the values from the training feature data set 0, M < M*); or providing the prediction feature data set based on a subset of the values from the training feature data set 1 ,...,M, t < 0, M < M*). Preferably the prediction feature providing step for providing a prediction feature data set will comprise providing the prediction feature data set n the same form and order as the training feature data se

Preferably the model setup step for setting up at least one model A _n * for the at least one slice n of the rotary kiln further comprises a temperature prediction step for predicting an outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln based on the prediction feature data set by using the prediction mode Preferably the predicted outside temperature data set (T _n (t), t > 0) comprises temperature data in the future, that is after running the method (which is at t = 0), i.e. , at a time t > 0. The outside temperature data set (T _n (t), t > 0) comprises a time series of values of temperatures at specific times t > 0: t ₁₍ ... , t _K . The times t ₁₍ ... , t _K are preferably selected from a time interval between just before the execution of the method up to certain time in the future, preferably at least 5 days in the future, more preferably at least 10 days in the future, most preferably 30 days in the future.

Preferably, the model setup step for setting up at least one model for the at least one slice n of the rotary kiln comprises a scoring function setup step for setting up a scoring function E _n for the at least one slice n of the rotary kiln, whereas the algorithm setup step for setting up at least one algorithm A _n for the at least one slice n of the rotary kiln is repeated for different sets (i) of hyperparameters, leading to different models A _n ⁽ⁱ⁾ with different results of the scoring function E _n ⁽ⁱ⁾ , and subsequently the model A _n ^{( )} , with the best result of the scoring function E _n ^{( )} being selected as the prediction model A _n *

Preferably the training outside temperature data set (T _n (t), t < 0) provided in the training temperature providing step comprises temperature data for the at least one slice n of the rotary kiln obtained by at least one temperature surveillance unit, such as an infrared camera. Preferably the temperature surveillance unit (20) comprises at least one of: an infrared camera (25), a discrete temperature measuring probe such as a thermocouple, or an optical temperature measuring device, such as a fiber-optic temperature sensor.

Preferably the method according to this first aspect of the invention is performed for a number of N slices of the rotary kiln in parallel. Thus preferably the method according to this first aspect of the invention is a method for predicting an outside temperature (T _n (t), t > 0) of a rotary kiln shell for at least N slices n (n = 1 ,...,N) of the rotary kiln, preferably for at least 100, more preferably at least 500, most preferably at least 1000 slices, whereas for each slice n the respective method claims are performed. Thus, preferably the at least one slice n comprises preferably at least 100, more preferably at least 500, most preferably at least 1000 slices.

Preferably, the training outside temperature data set (T _n (t), t < 0) provided in the training temperature providing step comprises outside temperature data that is averaged within the at least one slice n of the rotary kiln, preferably by averaging the outer temperature data along the circumference of the at least one slice n of the rotary kiln. This allows to train the model algorithm without requiring to obtain a parameter for the rotation of the rotary kiln within a training feature data set. In a second aspect of the invention, the object is achieved by providing a prediction unit for predicting the outside temperature (T _n (t), t > 0) of a rotary kiln shell for at least one slice n of the rotary kiln, comprising:

(a) a training feature interface for receiving a training feature data set of the rotary kiln;

(b) a training temperature interface for receiving a training outside temperature data set (T _n (t), t < 0) comprising outside temperature data of a rotary kiln shell for the at least one slice n of the rotary kiln;

(c) a temperature prediction interface for providing an outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln; the prediction unit being programmed to:

(d) provide the training feature data set preferably comprising values for M* features of the rotary kiln, from the training feature interface;

(e) provide the training outside temperature data set (T _n (t), t < 0) from the training temperature interface;

(f) set up at least one model A _n * for the at least one slice n of the rotary kiln, by:

(i) selecting at least one algorithm A _n for the at least one slice n of the rotary kiln from a machine learning algorithm, the at least one algorithm A _n comprising model parameters and hyperparameters

(ii) training the at least one algorithm A _n for the at least one slice n of the rotary kiln by using the training feature data set and the training outside temperature data set (T _n (t), t < 0) resulting in the at least one algorithm A _n for the at least one slice n of the rotary kiln in a trained state;

(iii) whereas training the at least one algorithm A _n is based on supervised learning;

(iv) achieving a prediction model A _n *, which comprises the at least one algorithm A _n for the at least one slice n of the rotary kiln in a trained state, and a set hyperparameters;

(g) provide a prediction feature data se preferably comprising values for the M features for the at least one slice n of the rotary kiln;

(h) predict an outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln based on the prediction feature data set by using the prediction model A _n *

The training feature interface for receiving a training feature data set of the rotary kiln might be any interface able to receive a training feature data set from a feature providing unit, such as, e.g., from a source like a memory or a server or directly from a production system or the like.

The training temperature interface for receiving a training outside temperature data set (T _n (t), t < 0) for the at least one slice n of the rotary kiln might be any interface able to receive a temperature data set, e.g., from a source like a memory or a server or directly from a temperature surveillance unit, e.g., a unit comprising an infrared camera or the like. Preferably the prediction unit might further comprise a temperature surveillance unit for providing a training outside temperature data set (T _n (t), t < 0) comprising temperature data of the outside surface of a rotary kiln shell for the at least one slice n of the rotary kiln to a training temperature interface for receiving the training outside temperature data set (T _n (t), t < 0) comprising outside temperature data of a rotary kiln shell for the at least one slice n of the rotary kiln. The temperature prediction interface for providing an outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln might be any interface able to provide a temperature data set, e.g., to a source like a memory or a server or directly to a temperature prediction displaying device, such as a display or a monitor or alike.

Any interface is understood to mean one or more devices for receiving or providing a data set as described above, and which, for this purpose, comprise either discrete electronic components in order to handle the data, or which are implemented partially or completely as a computer program in a computer.

The prediction unit according to the second aspect of the invention is preferably programmed to perform any or all the steps according to the first aspect of the invention.

In a third aspect of the invention, the object is achieved by providing a prediction system for predicting the outside temperature (Tn(t), t > 0) of a rotary kiln shell for at least one slice n of the rotary kiln comprising: a prediction unit according to any of the second aspect of the invention; a feature providing unit for providing a training feature data set for the at least one slice n of the rotary kiln to the training feature interface of the prediction unit; a temperature surveillance unit for providing a training outside temperature data set (T _n (t), t < 0) comprising temperature data of the outside surface of a rotary kiln shell for the at least one slice n of the rotary kiln to the training temperature interface of the prediction unit; a temperature prediction displaying device for displaying information based on the outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln provided by the temperature prediction interface of the prediction unit. The feature providing unit for providing a training feature data set for the at least one slice n of the rotary kiln to the training feature interface of the prediction unit might be a source like a memory I data acquisition unit connected to a production and I or laboratory information management system or the like, which is able to provide features of the rotary kiln as the training feature data set It is preferable that the feature providing unit receives a feature data set ( preferably comprising values for M* features of the rotary kiln, from the production information management system and I or laboratory information management system of the rotary kiln.

The temperature surveillance unit for providing a training outside temperature data set (T _n (t), t < 0) comprising outside temperature data of a rotary kiln shell for the at least one slice n of the rotary kiln to the training temperature interface of the prediction unit might comprise discrete temperature measuring probes, such as thermocouples or alike, or optical temperature measuring devices, such as a fiberoptic temperature sensor. Preferably the temperature surveillance unit comprises an infrared camera, preferably at least one infrared camera. Preferably the infrared camera records the outside kiln shell temperature, and the temperature surveillance unit provides a training outside temperature data set (T _n (t), t < 0). Preferably the temperature surveillance unit comprises a fiber-optic temperature sensor. Preferably the fiber-optic temperature sensor is in thermal contact with the rotary kiln shell. Preferably the temperature surveillance unit provides a temperature data set (T _n (t), t < 0) comprising outside temperature data of a rotary kiln shell for a number of N slices n (n = 1 ,...,N) of the rotary kiln, preferably for at least a number of N = 100, more preferably at least a number of N = 500, most preferably at least a number of N = 1000 slices.

The temperature prediction displaying device for displaying information based on the outside temperature data set (T _n (t), t > 0) for the at least one slice n of the rotary kiln provided by the temperature prediction interface of the prediction unit might be a display or a monitor or alike. Exemplary embodiments of the invention are explained in more detail by means of illustrations:

Figure 1 shows a diagrammatic sketch of the first aspect of the invention,

Figure 2 shows a diagrammatic sketch of the second and third aspect of the invention,

Figure 3 shows an exemplary calculation where the training outside temperature data set (T _n (t), t < 0) is averaged within the at least one slice n of the rotary kiln,

Figure 4 shows the result of a prediction of the outside temperature (T _n (t), t > 0) of a rotary kiln shell.

Figure 1 shows a diagrammatic sketch of the first aspect of the invention, which is a method (100) for predicting an outside temperature (T _n (t), t > 0) of a rotary kiln shell (3) for slice n of a rotary kiln (1 ), with several steps. First, a training feature providing step (110) is shown, for providing a training feature data set 1 ,...,M* t < 0) comprising values for M* features of the rotary kiln (1). In a specific example, 150 such features (M* = 150) are provided as a time series 1 ,...,M* at a certain time t < 0, i.e., before the execution of the method (100). In a specific example, each time series for each feature m comprises values for the last 360 days in intervals of 1 hour (thus k = 8640 values per feature m are provided in this specific example). All features are collected at the same time; thus, the time series are synchronized in that each time series comprises values for the respective feature m at same times of t < 0: t ₁₍ K = 8640 in this specific example.

Then a training temperature providing step (120) provides a training outside temperature data set (T _n (t), t < 0) comprising temperature data of the outside surface (2) of a rotary kiln shell (3) for slice n of the rotary kiln (1 ). In the specific example the training outside temperature data set (T _n (t), t < 0) provided in the training temperature providing step (120) comprises outside temperature data for the slice n of the rotary kiln (1 ) obtained by the least one temperature surveillance unit (20), which in this example is an infrared camera (25). The infrared camera may obtain outside temperature data of a rotary kiln shell (3) for several slices n of the rotary kiln. This is shown in Figure 3, where temperature data are collected for a slice n (n = 1 ,...,N) and in a certain segment q (q = 1 ,...,Q). A segment would be a certain section of the rotary kiln shell (3) running along a z- direction (which is running along the direction of the longitudinal axis L) and comprising a certain range of the azimuth angle of the rotary kiln (1 ). The training outside temperature data set (T _n (t), t < 0) provided in the training temperature providing step (120) comprises outside temperature data averaged within the at least one slice n of the rotary kiln (1 ), thus in this specific example the provided training outside temperature data set (T _n (t), t < 0) of slice n is the average temperature T _n (t) in that slice, thus

In this specific example a pre-processing step (125) for pre-processing of at least one data set is performed, whereas in this example two data sets are used, namely the training feature data se and the training outside temperature data set (T _n (t), t < 0). For both of these data sets the data format and data structure are checked, as well as the consistency of the data set. Further an outlier extraction, a removal of duplicates and a missing value imputation are performed on both data sets. In this specific examples even a scaling of the values in the data set is performed by normalization and standardization.

Another step is the model setup step (130) for setting up at least one model A _n * for the slice n of the rotary kiln (1 ). This step comprises further sub steps. In this specific example a scoring function setup step (140) for setting up a scoring function E _n (t) for the slice n of the rotary kiln (1 ) is performed. In this specific example the scoring function is the RMSE scoring function, thus for all t < The RMSE scoring function is a function of the (to be selected) algorithm (4 _n (t)) and the provided training outside temperature data set (T _n (t), t < 0). The scoring function is a measure on how good the algorithm is trained, as it is a measure of the deviations between the output of the algorithm (4 _n (t)) and the training outside temperature data set (T _n (t), t < 0).

In an algorithm selection step (150) the at least one algorithm A _n for slice n of the rotary kiln (1 ) is selected from a machine learning algorithm, in this specific example the algorithm is the neural network (NN) algorithm, this algorithm A _n comprises model parameters and hyperparameters In this specific example a first set of hyperparameters (HP”) is chosen (in this case this is the optimizer chosen as “rmsprop”, the epochs chosen to be 100, and the batch size chosen to be 20). These specific choices of the hyperparameters were found from a technique for the optimization of hyperparameters , in this case the method of grid search (neural network grid search) was used.

An algorithm feature selection step (160) is performed before the algorithm training step (170), where the algorithm feature selection step (160) comprises the selection of M features from the training feature data se comprising values for M* features for the slice n of the rotary kiln (1 ) and providing a selected training feature data se comprising values for M features for the at least one slice n of the rotary kiln (1 ). In this example by using a correlation-based feature selection (CBFS), the above- mentioned values for M* = 150 features of the rotary kiln (1) could be reduced to M = 20 features.

In an algorithm training step (170) the algorithm A _n for the slice n of the rotary kiln (1 ) is trained by using the selected training feature data set comprising values for M = 20 features of the rotary kiln (1) and the training outside temperature data set (T _n (t), t < 0) resulting in an algorithm A _n * for the slice n of the rotary kiln (1) in a trained state. Training is achieved by supervised learning. Mathematically this is to be understood as a variation of the model parameters such that the scoring function E _n (t) (summed over all times) becomes a minimum, converges or at least that it lies below a certain threshold value (in this specific example the optimization achieved a minimum RMSE of 4.05°C). In such a trained algorithm, the output of the algorithm will generally be a good fit to the training outside temperature data set (T _n (t), t < 0). The algorithm in a trained stated with its optimized model parameters and its set of hyperparameters HP- ¹ is called the model I prediction model A _n * , which is a function of the input features and the time The output of a previously trained algorithm (trained on 21 .11 .2020, t = -10d; d = days), without prior knowledge of the outside temperature data set of t = -10d to t = 0; d = days) is shown in Figure 4 by the dotted line marked “p” for “predicted” in the region from t = -10d (that is from the 21 .11 .2020) until t = 0 (that is until the 1 .12.2020). This output from a previously trained model was used as a previous prediction, which can be compared with the obtained outside temperature data set (T _n (t), solid line marked “m” for “measured” in the region of t = -1 Od to t = 0; d = days). It can be seen from this comparison, that the previous trained model was indeed capable of predicting the outside temperature data set T _n (t) in the region of t = -10d to t = 0 (d = days) quite well.

In a further prediction feature providing step (180) for providing a prediction feature data set the same M = 20 features are provided for the future (from t = 0 to t = +20d; d = days). These features are taken based on a production plan of the rotary kiln (1 ).

In a temperature prediction step (190) the outside temperature data set (T _n (t), t > 0) for the slice n of the rotary kiln (1 ) is predicted based on the prediction feature data set by using the prediction model Figure 4 shows the output of a (newly) trained algorithm (trained on 1.12.2020, t = 0; dotted line marked “p” for “prediction”), which additionally to the previously trained model uses the obtained outside temperature data set T _n (t) in the region of t = -10d to t = 0 (d = days) as the (additional) training outside temperature data set (T _n (t), t < 0). This updated trained algorithm is used to predict the future outside temperature data set (T _n (t), t = 0 to t = 20d; d = days). After a certain time, e.g., at 11 .12.2020 (thus at t = 10d), there might be another update to this trained model by using the additional outside temperature data set T _n (t) in the region of t = Od to t = 10d (d = days) as the (additional) training outside temperature data set (T _n (t), t < 0), and so on.

A certain predefined temperature range (in this specific example between Tmin = 340°C and T _m ax = 370°C) can be set. In case the predicted outside temperature data set (T _n (t), t > 0) lies outside this temperature range, an alarm can be triggered, allowing the operator to react in a timely manner. In a situation where excessive outside temperatures are predicted, the operator may change operational parameters of the kiln. Alternatively the operator may plan a repair or renewal of the refractory lining inside the kiln, well before a critical situation due to high temperatures arises. Such a situation is shown in Figure 4, where at t = 5d (marked with ® in Figure 4) a first situation occurs, where the predicted outside temperature data set (T _n (t), t > 0) drops below the lower threshold of Tmin = 340°C by a few °C. In this situation where a peak down in temperature is predicted, the operator may change operational conditions to prevent such a peak down in temperature, which will prevent or at least reduce wear of the refractories based on thermoshock, so that an increased refractory lifetime can be achieved.

A second situation is shown in Figure 4, where at t = 12d (marked with ® in Figure 4), where the predicted outside temperature data set (T _n (t), t > 0) exceeds the upper threshold of T _m ax = 370°C by a few °C. The operator may react by changing the operational parameters of the kiln to prevent this second situation. In an even more severe situation, the operator might have to react in order to prevent damage of the rotary kiln shell (3). The reaction by the operator can be planned well before such a situation arises, which either allows preventing the situation at all or allows scheduling a planned repair with a reduced downtime. In contrast, if such a situation occurs without advance notice, it may potentially lead to an emergency shutdown and unscheduled maintenance with significant downtime and production loss.

In this regard, Figure 2 also shows a specific example of prediction system (40) with a prediction unit (30) for predicting the outside temperature (T _n (t), t > 0) of a rotary kiln shell (3) for slice n of the rotary kiln (1 ). The prediction unit (30) of this specific example is a computer.

The prediction unit (30) comprises a training feature interface (310) for receiving a training feature data set ( of the rotary kiln (1 ) from a feature providing unit (10), which provides a feature data set 0) comprising values for M* = 150 features of the rotary kiln (1 ) received from the production information management system and I or laboratory information management system of the rotary kiln (1 ).

The prediction unit (30) comprises a training temperature interface (320) for receiving a training outside temperature data set (T _n (t), t < 0) for the slice n of the rotary kiln (1 ), in this example a training outside temperature data set (T _n (t), t < 0) for the slice n of the rotary kiln (1) is obtained by least one temperature surveillance unit (20), which is here an infrared camera (25). Here the training temperature interface (320) collects the received outside temperature data set (T _n (t), t < 0) over time.

The prediction unit (30) comprises a temperature prediction interface (390) for providing an outside temperature data set (Tn(t), t > 0) for the slice n of the rotary kiln (1 ), this temperature prediction interface (390) is connected to a temperature prediction displaying device (490) for displaying information based on the outside temperature data set (Tn(t), t > 0) for the slice n of the rotary kiln, which in this case is a display.

The prediction unit is programmed to perform the steps as discussed above in relation to the first aspect of the invention.

Shown in Figure 2 is a rotary kiln system (50) which comprises a rotary kiln (1 ) with a rotary kiln shell (3), comprising a production information management system and a laboratory information management system and the abovedescribed prediction unit (30). Here the feature providing unit (10) receives a feature data set comprising values for M* = 150 features of the rotary kiln (1 ) from the production information management system and laboratory information management system of the rotary kiln (1) and the temperature surveillance unit (20) provides an outside temperature data set (T _n (t), t < 0) comprising outside temperature data of the rotary kiln shell (3) for the slice n of the rotary kiln. Also shown in Figure 2 is the longitudinal axis L of the rotary kiln (1 ) and the z-direction running along the direction of the longitudinal axis L, and the azimuth angle of the rotary kiln (1 ), and the slice n (n = 1 ,N) of the rotary kiln (1 ).

The temperature prediction displaying device (490) of this specific example shows a graph like the one of Figure 4 to an operator. Also shown will be alarm messages in case the outside temperature data set (T _n (t), t > 0) lies outside a predefined temperature range.

In this example the method according to the first aspect of the invention is performed for a number of N = 3000 slices in parallel.

List of reference numerals and abbreviations:

1 rotary kiln

2 (outside) surface of rotary kiln shell

3 rotary kiln shell

10 feature providing unit

20 temperature surveillance unit

25 infrared camera

30 prediction unit

40 prediction system

50 rotary kiln system

100 Method

110 training feature providing step

120 training temperature providing step

125 pre-processing step

130 model setup step

140 scoring function setup step 150 algorithm selection step

160 algorithm feature selection step

170 algorithm training step

180 prediction feature providing step

190 temperature prediction step

310 training feature interface

320 training temperature interface

390 temperature prediction interface

490 temperature prediction displaying device

L Longitudinal axis of rotary kiln 1 t time t = tmin Starting time for training data set t = 0 End time for training data set I Starting time for prediction t > 0 Prediction Period n Index of slice, n = 1 m Index of a feature, m = 1 q Index of a segment in a slice, q = 1 ,...,Q Outside temperature of q-th segment in n-th slice of the rotary kiln shell at time t

T _n (t) Temperature data set comprising temperature data for each of the N slices, n = 1 ,...,N, of the rotary kiln shell

An Algorithm of n-th slice Training feature data set comprising M features (m = 1 for each of the N slices (n = 1 , ... , N) of the rotary kiln at time t set of K model parameters for the slice n; k = 1 ,...,K Optimized set of K model parameters for the slice n; k = 1 ,...,K Set of J hyperparameters for the slice n; j = 1 E _n (t) Scoring function Azimuth angle of the cylindrical coordinate system of the rotary kiln