Method and apparatus for optimizing diagnostics of rotating equipment

The present invention relates generally to a method and an apparatus for optimizing diagnostics of rotating equipment, in particular a gas turbine. BACKROUND

Remote monitoring and diagnostics of rotating equipment is indispensable in practice. Remote diagnostics of gas turbines is a complex task which can be divided into three steps: (1) Detection, (2) Isolation, and (3) Diagnosis. Recently, there has been an increased demand for a systematic approach to plant process safety, increased reliability and availability, lower maintenance cost, and continuous awareness about the equipment health status. This demand challenges the existing tool landscape which typically builds on an adoption of con ^¬ dition monitoring solutions and expert systems. Specifically, fault detection, fault isolation, failure mechanism definition and diagnosis definition as part of the systematic diag ^¬ nostics are fundamental functionality to support engineers in their decision making process, until the corrective action recommendation. However, due to the technical complexity caused by the large number of subsystems and process flows, diagnosis for industrial gas turbines is non-trivial, and re ^¬ quires multi-disciplinary expertise of various engineers from domains such as system mechanics, aerodynamics, and thermody ^¬ namics, to name only a few.

Only recently, the growth of computational power gave autono ^¬ mous decision making methods from the area of artificial in ^¬ telligence a second wind, making available new methods and tools to tackle the challenges outlined before. One such ex ^¬ ample is Deep Learning, a powerful method that makes use of GPU hardware to build models with unseen capabilities to au ^¬ tomatically construct relevant features from data. During the analysis phase, the expert at the remote diagnos ^¬ tics center (RDC) normally enriches the sensor data available in the above mentioned step (1) with his findings and hypoth ^¬ eses about failure modes and solutions, all of which are doc- umented in a ticketing system (e.g. Salesforce or STM-RMS) as free text in natural language. While this unstructured (or semi-structured) way of documentation is convenient for the technician, it makes it very hard to share the knowledge ex ^¬ pressed in these annotations with other colleagues. It is to propose solutions based on similar cases from the past. The challenge is therefore to provide a system which can automat ^¬ ically propose relevant historic cases to the technician dur ^¬ ing diagnosis, where both sensor data as well as (intermedi ^¬ ary) human-generated content, mostly textual information, is taken into account. Furthermore, it is not practically feasi ^¬ ble to have a solution that needs extensive manual tuning of parameters to perform well. Up to now, the diagnostic process in the remote diagnostic center (RDC) is largely manual and lacks support through software tools.

The above mentioned ticketing system Salesforce has integrat ^¬ ed functionality for the discovery of tickets that are simi ^¬ lar to the one currently opened. It is very likely that standard measures such as TF/IDF over bag-of-words are used in this system. Term frequency-inverse document frequency

(TF/IDF) is a numerical statistic that is intended to reflect how important a word is to a document. Moreover, no sensor data is included. One further possible approach is case-based reasoning over spectral decompositions of sensor data to be used for identi ^¬ fying vibration situations. The underlying the feature vector computation with respect to spectral decomposition is completely different approach than learning methods.

An aspect of classic case-based reasoning application is to manually define weights for similarity comparisons.

Firstly, such weights are typically not known and also not intuitive for experts to define. Secondly, the ef ^¬ fort for collecting such "estimations" of similarity is con ^¬ siderable, taking into account that the experts need to de ^¬ fine both "local" similarities between different manifesta- tions as a feature, and a global combination function

that combines the local similarities to a global value.

Common downsides of most of previous mentioned potential working solution that are not able to automatically predict relevant historic cases taking into account both sensor and textual information, nor can avoid extensive manual parame ^¬ terization .

It is an object of the present invention to provide an ap- proach that integrates textual information into learning based approaches to optimize gas turbines diagnostics.

SUMMARY The above-mentioned object is achieved by a method and one or more apparatus for optimizing diagnostics of rotating equip ^¬ ment comprising the features of the independent claims. Pre ^¬ ferred embodiments of the invention are described in the de ^¬ pendent claims.

An Aspect of the invention is (dynamic) integration textual information into learning based approaches to optimize gas turbines diagnostics. The inventive approach supports engineers in the Remote Diag ^¬ nostic Centers. It is based on a combination of Natural Lan ^¬ guage Processing (NLP) technologies that allow us to build on the vast amount of diagnostic knowledge written down by the engineers with Deep Learning to include information about the actual turbine status derived from the available sensors.

This approach is embedded into an overall systematic workflow building on physics-based, rule-based and data-driven meth ^¬ ods. This framework supports the engineer in identifying rel- evant information, thereby reducing trouble shooting time significantly, increasing both Technical Responsiveness capa ^¬ bility and capacity.

The proposed invention claims a method for optimizing rotat ^¬ ing equipment diagnostics, in particular gas turbine diagnos ^¬ tics comprises the method steps of: a) Providing status information about status of the rotat ^¬ ing equipment over a series of time windows whereby sta ^¬ tus is derived from sensor features of at least one available sensor taking measurements during a

predefinable time period,

b) Using deep learning which combines provided historic

sensor information with sequence of events data indicat ^¬ ing warnings and/or alerts of the rotating equipment, c) whereby status information is supplemented with via deep learning predicted probabilities whether a warning and/or an alert has occurred within a time window, d) providing an amount of textual diagnostic knowledge cas-

& S

e) extracting semantic information on text features from the textual diagnostic knowledge cases , and

f) combining status information and semantic information into a unified representation enabling optimization of the diagnostics .

Such optimized diagnostics lead to adj ust operation of rota ^¬ tion equipment and/or to maintain the rotation equipment .

Deep learning can use a case based reasoning learning method . A natural language training method can be used for extracting said semantic information . Different weights can be applied to different types of text features .

The result of this extraction is preferably a set of text feature vectors , one said vector for each textual diagnostic knowledge case . One text feature vector can be determined by classifying the case against trained resulting in dif ^¬ ferent clusters brought about case similarity computation whereby the vector contains as many cluster membership degrees as clusters exist . Different types of text features can be affected parts and/or observed symptoms .

Deep learning can automatically identify latent structure that makes two said time windows similar or dissimilar in or ^¬ der to predict said probabilities.

Status information from the predefinable time period e.g. 24 hours with said probabilities are represented by a vector . Predefinable may mean that a user can enter or select a time period or the time period is set by a default value e.g. 24 hours .

Said unified representation leads into one unified sensor and text feature vector .

A further aspect of the invention is an apparatus for opti ^¬ mizing gas turbine diagnostics comprising: a) Means (11) for providing status information about status of the rotating equipment over a series of time windows whereby status can be derived from sensor features (10) of at least one available sensor taking measurements during a predefinable time period,

b) means (20) for using deep learning which combines pro ^¬ vided historic sensor information with sequence of events data indicating warnings and/or alerts of the ro ^¬ tating equipment,

c) whereby status information is supplemented with via deep learning predicted probabilities whether a warning and/or an alert has occurred within a time window, d) means (24 ) for providing an amount of textual diagnostic knowledge cases , e) means ( 17 ) for extracting semantic information on text features from the textual diagnostic knowledge cases , and

f) means (19) for combining status information and semantic information into a unified representation enabling optimization of the diagnostics .

A further aspect of the invention is a computer program

(product) directly loadable into the internal memory of a computer, comprising software code portions for performing the steps of the above mentioned method when said computer program (product ) is running on a computer or on one of the above mentioned apparatus . Such a framework for context-aware analytics within flexible manufacturing systems, motivated by the need for accurate processing time estimates, is a benefit of this inventive ap ^¬ proach. It can be successfully applied and commits less pre ^¬ diction errors compared to state-of-the-art adaptive learning models. More accurate estimates of processing times directly influence reliability of the manufacturing system's through ^¬ put times and cycle times, which are the basis for optimized production planning and scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

The foregoing and other aspects of the present invention are best understood from the following detailed description when read in connection with the accompanying drawings. For the purpose of illustrating the invention, there are shown in the drawings embodiments that are presently preferred, it being understood, however, that the invention is not limited to the specific instrumentalities disclosed. Included in the draw ^¬ ings are the following Figures:

FIG. 1 schematically shows a concept using CBR and NLP based clustering and ranking, FIG. 2 schematically shows deep learning used for classifying time windows of events and extracting compact latent feature representations,

FIG. 3 depicts combining the feature vectors from CBR and NLP into one unified feature vector, and

FIG. 4 depicts an example of Similarity Computation

DETAILED DESCRIPTION FIG. 1 schematically shows a concept a concept using CBR and NLP based clustering and ranking.

Marked with 1, 2 and 3 the figure 1 shows the previous men ^¬ tioned steps (1) Detection 11, (2) Isolationl2, (3) Diagnosis 15 of sensor data 10 within one or several time windows (24h window) . In the step (2) and (3) a user may interacts with the system via a dashboard 13 which executes said steps. Pre ^¬ liminary Notification 14 following after step (2) and Customer Notification 16 after step (3) could be useful as shown in Fig. 1.

The proposed approach uses both the sensor data 10 as well as the natural language annotations for automatically identify ^¬ ing similar cases from the past. The result of the similarity analysis and computation can be presented to the engineer / user in order to facilitate his/her search for a solution. An integration of the recommendation mechanism with standard tools and/or with a system already in use at the remote diag ^¬ nostic centers (RDC) for gas turbines, such as Salesforce can be useful.

In a nutshell, historic data (both textual, that means his ^¬ torical cases 24 and historical sensor data 23, and inter ^¬ linked via case ID and time information) are used. A training model can be used to

(a) compress a 24h window of turbine sensor data into a com ^¬ pact latent feature vector (see "Sensor Feature Compression" 22 ) , and (b) to assign a new (preliminary) case in said tool/system to a set of pre-computed clusters based on the textual content (see "Text feature Extraction (NLP) " 17 and "Cluster Retrieval" 18, giving as result a vector of membership degrees for each of the clusters. So Ranking of the membership and the cluster Retrieval is integrated and can be presented to the Diagnosis 15 step (3) .

Both feature vectors are then combined and compared to the case base containing analogous representations for all his ^¬ toric cases, giving as a result a list of relevant historic cases along with their degree of similarity, ordered by de ^¬ creasing similarity. Said case similarities along with the cluster membership degrees of those related cases allow to compute straightforwardly which clusters are most relevant for a given case. The result is then displayed to the diag ^¬ nostic engineer within the GUI, e.g. Display, of the diagnos ^¬ tic system. Analysis of text features for extracting (un ^¬ structured) text information can be implemented in a NLP (software) module 17. Sensor feature Compression 22 can be implemented in a CBR (software module) . So a combination of the results of these two "uni-modal" modules into a "multi ^¬ modal" overall assessment is employed. In the following these three modules in detail:

- Case-based reasoning (CBR) :

It is useful to apply machine learning methods (more con ^¬ cretely deep learning methods) to automatically identify the "hidden (latent) structure" that makes two (time) windows of sensor data similar or dissimilar. The historic data 23 can be used to train a so-called convolutional neural network 40 (CNN) based on an auxiliary task that combines the historic sensor information with sequence-of-events (SoE) data indi- eating warnings and alerts of the gas turbine. Deep learning is part of a broader family of machine learning methods based on learning representations of data. It may be based on a set of algorithms that attempt to model high-level abstractions in data by using multiple processing layers with complex structures, or otherwise composed of multiple non-linear transformations. CNN is usually a type of feed-forward arti ^¬ ficial neural network where the individual neurons are tiled in such a way that they respond to overlapping regions in the receptive field.

Deep learning comprises the following steps:

Step I: Classification of time window, preferably 24h win- dows,

Step II: Feature extraction

Step III: Similarity Computation

Step IV: Generalization which can be used for solving future problems of the same type.

As a first step I, a CNN is trained that takes all sensor measurements from a 24h time window as inputs (the measure ^¬ ments can have a resolution of one minute, leading to 1440 values for each sensor) . The auxiliary tasks consists of pre- dieting the probability whether a warning or error has occurred in a prior given 24h time window within the corresponding sensor data time series. After training (CBR) 20 is completed, the output node of the network is removed and the last hidden layer is used as new output layer. The new output of the CNN will then output a set of latent features extract ^¬ ed from the complete set of sensor data in the 24h time win ^¬ dow (Step II) . The number of latent features extracted is configurable in the design phase of the CNN structure. The similarity computation (instead of discovery) (step III) of relevant features is done automatically by the CNN learning algorithm.

Finally the CBR learning model 21 learns features to differ ^¬ entiate the time series where an event (warning or alarm) has occurred. By looking at figure 2 on the left hand side there are 24h Sensor data depict in curves. Shown on the right hand side (see 40) local features are mapped to global features by shown steps of convolution, subsampling and their repetition as long as the output needs to be classified into normal event or abnormal event. In the Generalization step IV may be concluded more than one rotating equipment of the same type. - Natural language processing (NLP) :

NLP can be used for processing human-generated content which is fed into the diagnosis process by the diagnostic engi ^¬ neer/user. It is possible to extend the standard bag-of-words approach to representing a document as a feature vector by linguistic approaches. Based on a (language specific) dic ^¬ tionary and grammar 26 extracted from the historic cases 24 by an NLP expert, also "semantic" information on affected parts and observed symptoms from the text are extracted. This allows the NLP module to abstract from concrete (syntactic) formulations and to focus on the meaning of the content.

Training 25 of the NLP models is a two-step process:

• In the first step each document has its bag-of words and semantic features extracted as described above. Differ ^¬ ent weights can be applied to different types of fea ^¬ tures, for instance to set more importance to symptoms. Then TF-IDF scaling can be applied across all features. The result of this step is a set of document feature vectors, one for each document.

• In the second step the document feature vectors are

clustered using non-negative matrix factorization. This step has two outcomes:

1) an assignment of documents to clusters 27 and

2) a set of new latent feature vectors, one for each docu ^¬ ment .

The latent feature vector is characteristic for a single doc- ument in the context of the other documents and the clusters they belong to. Similar documents have similar latent feature vectors. Based on the availability of experts/users, the clustering can be refined in an iterative process by discuss- ing it with a domain expert and translating his or her feedback into configuration parameters for the clustering algorithm, such as feature weights. The feature vector for a giv ^¬ en case is then determined by classifying the case (fuzzily) against the trained clusters, resulting in a vector of clus ^¬ ter membership degrees having as many entries as clusters. This "detour" enables to integrate expert feedback on the "similarity" of textual information e.g. textual case de ^¬ scriptions in a very straightforward way.

Integrated Ranking and Retrieval 19:

As depict in figure 3 the proposed solution combines the fea ^¬ ture vectors from CBR and NLP into one unified feature vector 32 which is in the end used for similarity computation. Uni- fied representation more specifically the unified feature vector enables feedback to the optimization process in diagnostics. A possible implementation of similarity computation is to separately compute cosine distance in the CBR-portions 30 and on the NLP portions 31 of two cases, and combine these distances into a unified measure by taking the weighted mean of the two ones 30, 31. The benefit in using both CBR and NLP information for similarity calculation despite the fact that NLP-based information have already been taken into account for cluster formation is that similarities on case level 33 are usually more consistent with than similarities on clus ^¬ ters 34.

By computing distances within the NLP- and CBR- modules/components prior to only combining their results in a second step, it needs to be taken into account that the dimensionalities of these two vectors differ significantly.

Figure 3 depicts for example these two vectors. This could otherwise lead to a bias towards the CBR-based assessment. The overall solution can be presented to the diagnostic engi ^¬ neer. Also, a technical implementation of the

propose solution may include some Ul-element (UI = User In ^¬ terface) that allows the engineer to rate and/or weight the results provided by the method. This enables continuous eval ^¬ uation of the approach as well as providing training data which can be used for further optimizing the used models. Such optimi zed diagnostics can lead to adj ust operation of rotation equipment and/or to maintain the rotation equipment .

Referring to the above mentioned Step III figure 4 depicts a simplified example of Similarity Computation. In the rows on the top there are case numbers and textual findings. Similar- ity computation as previously described brings about similar ^¬ ity degrees. E.g. Similarity degree related to the text fea ^¬ ture "Bleed Valve" between case with case number 301142 and case 297987 is 0.44 and the similarity degree related to text feature "Pressure compressor" between case 301142 and case 299647 is 0.24.

It is to be understood that the elements and features recited in the appended claims may be combined in different ways to produce new claims that likewise fall within the scope of the present invention. Thus, whereas the dependent claims append ^¬ ed below depend from only a single independent or dependent claim, it is to be understood that these dependent claims may, alternatively, be made to depend in the alternative from any preceding or following claim, whether independent or de- pendent, and that such new combinations are to be understood as forming a part of the present specification.