Method for validating and selecting models for condition monitoring of machines

Abstract

The invention relates to a method for validating at least one model for monitoring a state of a machine and selecting at least one model for monitoring a state of a machine from a first set of models, wherein the models are algorithms trained by means of machine learning, wherein the models are trained by means of reference data records of operating parameters and characterized by different parameter configurations for different models, the method comprising: generating one or more test data records from the reference data records, wherein at least one value of an operating parameter in the reference data records is replaced by a corresponding impermissible operating parameter value; for each model of the set of models: applying the model to each test data record of the one or more test data records in order to determine an error value of the model with respect to each test data record, and determining a distance value of the model based on the error value of the model; selecting one or more models based on the size of the distance value.

Description

Technical Field

[0001] This invention relates to a method for validating at least one model for monitoring the state of a machine and selecting at least one model for monitoring the state of a machine from a first set of models. More particularly, this invention relates to validating models suitable for machine monitoring and selecting models suitable for machine monitoring from one or more models. Background Technology

[0002] Machine learning-based mathematical models can be used within the framework of predictive maintenance to classify the operating states of machines in industrial facilities and, in particular, to identify abnormal operating states that indicate machine malfunction. To train these mathematical models or algorithms, sample or reference data recorded during machine operation can be used as training data. During training, the model learns to identify the machine's operating states as represented by the training data. After training, by applying the trained model to newly recorded operating data of the machine, conclusions about the machine's operating state can be drawn.

[0003] The sample data used during training is typically recorded during normal machine operation. Since production in modern industrial facilities proceeds without failure in most cases, this sample data contains little or no data on failure states. Correspondingly, since abnormal or failure states are not included in the training data, the model obtained using machine learning cannot "learn" to identify these abnormal or failure states. Testbed data is also unsuitable because machines in real-world use are constructed and operate differently than those on a testbed. That is, data recorded on a testbed, for example, where failure states are intentionally generated, cannot be used.

[0004] Since there is no data on fault states and correspondingly no error metric for the machine's operating state, in practice, machine learning-based algorithms trained through unsupervised learning are used to encode and classify operating states. However, due to the lack of data on abnormal states, no training has been conducted on these abnormal states. Summary of the Invention

[0005] According to the present invention, a method for verifying and selecting at least one model for monitoring the state of a machine, a corresponding computing unit, a corresponding computer program, and a corresponding storage medium are proposed, having the features of the independent patent claims. Advantageous designs are the subject of the dependent claims and the description thereafter.

[0006] This invention enables the validation and selection of machine learning-based models that can identify abnormal states without prior explicit training or learning of the machine's abnormal operating states.

[0007] A method for validating at least one model for monitoring the state of a machine using unacceptable operating parameter values ​​and selecting at least one model for monitoring the state of a machine from a first model set comprising at least one model involves models that are algorithms trained using machine learning, which model the operating state of a machine based on operating parameters characterizing the machine's operating state, wherein the models are trained using reference data records of these operating parameters, and wherein the models are characterized by different parameter configurations for different models. The method includes: a) generating one or more test data records based on the reference data records, wherein at least one value of at least one operating parameter in the reference data records is replaced by a corresponding unacceptable operating parameter value; b) for each model in the model set, applying the model to each of the one or more test data records to determine an error value of the model with respect to each test data record, and determining a distance value of the model based on the error value; and c) selecting one or more models based on the magnitude of the distance value.

[0008] The distance value indicates how far the model assigns the test data record from the reference data record. Correspondingly, by selecting a model based on the magnitude of the distance value, a particularly suitable model can be determined.

[0009] The "operating parameters" of a machine should be understood as data recorded to monitor the machine's operation. This recording may, for example, be done at regular intervals. The corresponding specific values ​​of the operating parameters are then called operating parameter values. These values ​​are contained in the relevant data records. These operating parameter values ​​are, in particular, sensor values ​​measured by sensors placed on the machine. However, they can also involve pre-defined parameters, such as those pre-defined by the machine's controller. It is also possible that the operating data includes pre-processed data, such as the Fast Fourier Transform (FFT) values ​​of periodically measured values. The machine to be monitored may, for example, be a hydraulic system; in this case, sensor data can be selected, for example, from the pressure and volumetric flow of the hydraulic fluid, the temperature of the hydraulic fluid or machine components, the position of the hydraulic cylinders, or even the torque and power output of the engine driving the pump.

[0010] The term "unacceptable operating parameter value" refers to a value of an operating parameter that does not occur during the normal, trouble-free operation of the machine. Therefore, it is an unacceptable or unachievable operating parameter value. This value is known through domain knowledge. Such a value can be understood not only as, for example, a physically impossible value, but also as a value that represents an extreme state that is normally impossible. In particular, this includes extreme values ​​of operating parameters outside the range of values ​​that occur during the normal, trouble-free operation of the machine. This especially includes values ​​on the boundary between impossible and just achievable. In the case of hydraulic machines, a known example of an extreme state is a tank temperature of 100°C. Furthermore, this term should include not only individual operating parameter values ​​that are unacceptable on their own, but also combinations of operating parameter values ​​where the individual values ​​are permissible, but the combination is not. That is, the term "unacceptable operating parameter value" should include "unacceptable combinations of operating parameter values." A simple example of this is the readings of adjacent temperature sensors that are strongly correlated during normal operation. For these sensors, correspondingly large differences in temperature readings represent an unacceptable combination of operating parameter values. Unacceptable operating parameter values ​​(and unacceptable combinations thereof) are included in the fault data log.

[0011] The term "model" and machine monitoring methods based on these models are briefly described below as used within the scope of this invention. A model should be understood as a modeling of an operating state or measured operating parameters of such operating states through algorithms based on artificial intelligence or machine learning. In a process known as training or learning, the machine learning algorithm generates a mathematical model based on sample data (so-called training data) to make predictions about the process (here, the operating state of the machine) without being explicitly programmed for this. A model can be viewed as a mapping or function that maps input values, here, the operating parameters of the machine to be monitored, to output values.

[0012] The specific type and internal structure of the model or algorithm are not necessary for implementing the method according to the invention; it is only necessary to be a model that maps a set of input values ​​(vectors)—here, the values ​​of the operating parameters—to an output from which an error (error value) describing the difference from a reference operating state encoded in the model can be derived, wherein this difference is measured according to a measure of the model's error. The reference operating state is encoded in the model through machine learning, wherein the model is trained based on training data or reference data characterizing the reference operating state. This training is typically performed to minimize the error values ​​recorded in the training data with respect to the error measure. The training data includes, for example, operating parameter values ​​recorded during a reference run of the machine.

[0013] For machine fault monitoring, models that model the operating state in a way that allows input values, i.e., operating parameter values, to be reconstructed as the model's output values ​​are particularly useful. In a sense, the model reconstructs the input; the corresponding model might be called a "reconstructed model" or "model undergoing reconstruction." The error metric (error measure) is then obtained by comparing the output values ​​with the input operating parameters (which are the model's input values), for example, as the mean squared error. Machine monitoring methods used for anomaly detection typically choose models that reduce the dimensionality of the parameters within the model (i.e., make it impossible to replicate the same mapping). The operating state is then, in a sense, encoded by the model in a reduced-dimensional space. The model generates output values ​​from this reduced-dimensional space, which can be considered decoding. If an anomaly or malfunction occurs in the monitored, model-modeled machine, this results in deviations in the operating parameters that are no longer included in the encoding, causing the error metric used as an anomaly measure to obtain a higher value. If the error determined by the error metric exceeds the pre-defined error limit of the model, the machine is considered to be in an abnormal or faulty operating state. The use of such a model for anomaly identification and for monitoring the machine's operating status is generally known. Typically, the model is implemented as an (artificial) neural network, such as an autoencoder.

[0014] In one implementation, multiple models with the maximum distance value can be selected. The distance value of a model indicates how far the test data record is from the reference state encoded in the model (according to the model's error measure); that is, according to this implementation, the model that indicates the maximum distance for the test data record is selected. More preferably, the number is equal to one, that is, the model with the maximum distance value is selected.

[0015] Advantageously, the distance value for the corresponding model is determined such that it corresponds to a value selected from the following group: the minimum error value among the model's error values, the maximum error value among the model's error values, and the average error value of the model. Using these different feasible methods, different aspects of the error values ​​can be emphasized. More preferably, the same method is used to determine the distance value for all models, thereby ensuring better comparability of the distance values.

[0016] According to one implementation, step b) may further include: if at least one of these error values ​​is less than a predetermined maximum error limit of the model, then the model is discarded. Thus, models that cannot distinguish between test data records that should be characterized as abnormal states according to the design and data records that characterize normal operating states can be excluded from the selection.

[0017] According to the design scheme of this embodiment, when all models are discarded, the method may further include: defining a second model set including at least one model, wherein different models are defined using different parameter configurations, and the parameter configurations of the models in the second model set are different from the parameter configurations of the models in the first model set; training the models in the second model set using reference data records; and performing steps b) and c) for the second model set.

[0018] According to this design, when no valid model can be found in the first, i.e., the original first model set, the method can be applied to the modified second model set. If no valid model can be found in the second model set either, then this can be performed on other modified model sets.

[0019] Advantageously, the model is implemented as a neural network, particularly an autoencoder, wherein the parameter configuration includes the number of layers, the number of neurons for each of these layers, and / or the number of epochs. Neural networks and their training are well known to those skilled in the art, and the use of neural networks ensures the effective implementation of this method.

[0020] In one embodiment, the method may further include storing the parameter configuration of at least one selected model. Thus, the selected model can also be used on other similar machines or at least serve as a starting point for training a model.

[0021] The computing unit according to the invention, such as the control device of a hydraulic system, is particularly configured in a programmatic manner to execute the method according to the invention.

[0022] Especially when the control equipment being implemented is also used for other tasks and thus always exists, the implementation of the method according to the invention in the form of a computer program or computer program product having program code for executing all method steps is also advantageous, as this results in particularly low costs. In particular, suitable data carriers for providing the computer program are magnetic storage, optical storage, and electrical storage, such as hard disks, flash memory, EEPROM, DVDs, and others. Downloading the program via computer networks (Internet, intranet, etc.) is also feasible.

[0023] This method can be executed, in particular, on a remote computer or computer system. Correspondingly, the computer program or computer program product can be specifically set up or programmed to execute on a remote computer or computer system. Here, "remote" should be understood as meaning that the computer or computer system is not integrated into the machine to be monitored and is generally not located near the machine, for example, not in the factory where the machine is located. This could involve a remote server, such as a so-called cloud server, which is connected to the machine or the sensor recording operating parameters, especially via a network, such as the Internet and / or an intranet.

[0024] Other advantages and designs of the present invention will become apparent from the description and the accompanying drawings.

[0025] It is readily understood that the features mentioned above and to be described below can be applied not only in the combinations described separately, but also in other combinations or individually, without departing from the scope of protection of this invention.

[0026] The present invention is schematically illustrated in the accompanying drawings according to embodiments and is described in detail below with reference to the accompanying drawings. Attached Figure Description

[0027] Figure 1A and 1B Block diagrams and associated flowcharts of preferred embodiments of the method according to the invention are shown; and

[0028] Figure 2 A structure for an autoencoder that is particularly suitable for use as a model in the context of this invention is shown. Detailed Implementation

[0029] In the following text, based on Figure 1A and 1B This describes exemplary embodiments of a method according to the present invention for validating and selecting models for machine monitoring. Figure 1A The corresponding block diagram is shown in the figure, while Figure 1B The flowchart is shown below.

[0030] The verification and selection are performed based on a (first) model set, which includes at least one model, preferably multiple models. Figure 1AThe text symbolically depicts two models. These models are characterized by defining their parameter configurations. Here, the parameter configuration represents a description of the structure, the values ​​of specific parameters, training parameters (such as the learning rate or the number of iterations in an iterative learning process), the error metric of the algorithm on which the model is based, etc., which yields the determined model when the algorithm has this parameter configuration and is trained via machine learning using training data corresponding to the training parameters. That is, a specific model is obtained through training (using training data) based on an algorithm with a specific parameter configuration. In this sense, the model is defined by its parameter configuration and training data.

[0031] Reference data record 4 is used as the training data record, which includes, for example, operating parameters recorded during machine operation, such as sensor measurement data. With the aid of reference data 4, the model defined by parameter configuration 2 can be trained during training 6. Correspondingly, the method can begin by defining and training 52 a (first) model set including at least one model, wherein different parameter configurations are selected for different models under this definition.

[0032] Based on reference data 2 and fault data record 8 containing data with unacceptable operating parameter values, at least one test data record 10 is generated, step 54. A single test data record 10 is exemplarily shown in the figure; however, generating multiple test data records is equally possible. To generate test data record 10, it is advantageous to start with reference data record 4, and in the reference data record containing the recorded operating parameter values, at least one value of at least one operating parameter, preferably multiple values ​​of multiple operating parameters, is replaced by the corresponding unacceptable operating parameter value in fault data record 8. If the fault data record also contains unacceptable combinations of operating parameter values, the corresponding combination of operating parameter values ​​in the reference data record is entirely replaced.

[0033] In testing step 12, the model taught in training 6 is applied to the at least one test data record 10, step 56. This determines a series of error values ​​14, where each pair of model and test data record corresponds to a single error value. This error value is the error measured with respect to the model's error when it is applied to the test data record. In a sense, the error value reflects the deviation or distance between the corresponding test data record and the operating state trained by the model, more precisely as if that distance were measured by the model itself.

[0034] For each model, the distance or distance value between the model and the test data records can be determined based on the error value for all test data records, step 62. One or more models are selected based on the magnitude of the distance value, step 64. Preferably, multiple models with the largest distance value are selected, wherein the number is preferably equal to one in most cases, that is, the model with the largest distance value is selected.

[0035] Preferably, the selected one or more models are stored in database 16, step 66. Similarly, the parameter configurations of the selected one or more models can be stored in database 16, because these parameter configurations are promising and can be used as a starting point for monitoring models on similar or identical machines in another production cycle. That is, a new set of models, for example a new set of models for similar machines, can be obtained based on the stored models. Figure 1A (symbolically represented by arrow 18) and the same new model set is used to perform the method according to the invention.

[0036] The selected one or more models can be used to monitor the operating status of the machine and to identify abnormal states of the machine, such as hydraulic systems. That is, the method may include using the selected models for machine status monitoring and for identifying abnormal states of the machine, step 68.

[0037] The distance value of the model can be determined based on the model's error values ​​in different ways. One possibility is to determine the distance value such that it corresponds to the minimum error value among the model's error values, that is, equal to that minimum error value. In this case, it can be said that the error value indicates, to some extent, how far each anomalous state represented in the test data record is from the reference state modeled by the model. Alternatively, the maximum error value among the model's error values ​​can be chosen as the distance value. The corresponding distance value is a measure of how well the model has successfully separated at least one test data record from the reference data. Third, the distance value of the model can be determined as the average of the model's error values. Preferably, the arithmetic mean is used. The geometric mean can also be used, where smaller error values ​​are weighted more heavily than larger error values. More generally, an average with arbitrarily chosen weights can be used, thereby allowing for different weightings for different test data records, for example.

[0038] When determining the error value, it is possible that the model's error value or error with respect to at least one of the test data records is less than the model's error limit, which indicates the maximum permissible error when applying the model to the data record for that data record to still be classified as an acceptable operating state. In other words, the model's minimum error value is less than this error limit. In this case, the error value is measured by the model's error metric or error gauge; typically, the error with respect to this error metric is minimized during model training. In this case, the "model's error limit" should be understood as follows: even if the operating state is correctly classified, applying the model to data records of operating parameters recorded during machine operation will generally always result in small errors, because small fluctuations in the operating parameters characterizing that operating state always occur within the operating state. Operating states that result in only small errors when applying the model are considered acceptable (i.e., normal, fault-free) operating states. Training data records will also typically have non-zero errors. Therefore, the (maximum) error limit is typically associated with each model, which assumes that the machine is not in an abnormal state as long as the error when applying the model to the operating parameters corresponding to the operating state is less than the error limit. That is, the model's error limit indicates the maximum error that the model allows for when applying it to data records of operating parameters, such that the data record is evaluated as a valid, normal operating state.

[0039] If the error when applying the model to test data records describing an abnormal state of the machine according to the design is within the error limit, then the model cannot distinguish the abnormal state from the normal state of the machine. In principle, the method according to the invention can be applied even if one of the model's error values ​​is less than the model's error limit. If the model is selected, it is partially invalid, meaning the test data records are partially classified as permissible, that is, these test data records are incorrectly classified. That is, the term "verification" should not be understood as causing the corresponding model to inevitably identify all test data records as abnormal states, but rather more generally, verification includes the selection of models that are particularly suitable in terms of distance measurement. If necessary, the problem of partially invalid models can be eliminated by selecting (and later using) multiple models, because one of the other selected models may be able to correctly classify the corresponding test data records.

[0040] However, in this case, that is, when one of the model's error values ​​is less than the error limit, it is preferable to discard the corresponding model, step 60. Discarding a model means excluding that model from the selection. If all models are discarded, then the selection for one or more models with the maximum distance value is empty, meaning that in this preferred embodiment, no model has been verified as a suitable model.

[0041] Alternative locations, based on Figure 1B In the preferred embodiment shown, if all models are discarded, a modified model set can be defined, step 74, which will be referred to below as the second model set, to distinguish it from the original first model set. Like the first model set, the second model set contains at least one model.

[0042] Specifically, in step 60, it is first determined whether all models have been discarded (step 70); or whether not all models have been discarded (step 72). If not all models have been discarded (step 72), then proceed to step 62, that is, continue to determine the distance values. If all models have been discarded (step 70), then proceed to step 74, that is, continue to define the second model set.

[0043] It is also possible that (as in) Figure 1B The situation shown in the example is different: distance values ​​are determined first, and models are discarded later. In this case, it can be based on these distance values ​​to decide which models should be discarded; for example, models whose distance values ​​are less than a predetermined minimum distance value are discarded.

[0044] As described above in conjunction with the first model set, the models in the second model set are defined or represented through parameter configuration. Different models in the second model set have different parameter configurations. Similarly, the parameter configurations of the models in the second model set differ from those in the first model set.

[0045] The models of the second parameter set can be trained using reference data records based on their respective parameter configurations, step 76. Next, the models of the second model set can be applied to the test data records as has happened with the models of the first model set, step 56, to determine the error values ​​of the models of the second model set, i.e., returning to the corresponding steps of the method. Other steps (discarding models, step 58; determining the distance for each model, step 62; selecting models, step 64) are also performed as described above, except that the models of the second model set are now used instead of the models of the first model set. If all models in the second model set should also be discarded, a third model set can be defined, and so on.

[0046] Preferably, the first model set, as well as the second model set and other model sets, if necessary, each contain at least two models.

[0047] Figure 2A simplified diagram of the structure of an autoencoder is shown, which can be used, for example, as a model in the monitoring method upon which this invention is based. In principle, an autoencoder can be used to model a reference state of a machine. The autoencoder 200 has an input layer 202, an output layer 204, and multiple hidden layers 206 (in... Figure 2 The example shown contains three hidden layers, but other numbers are also possible. Each of layers 202, 204, and 206 includes one or more neurons (in...). Figure 2 Only a few neurons (220, 222, 230, 232) are shown in the diagram. The neurons in the output layer 204 have the same meaning as the neurons in the input layer 202; that is, the values ​​in the output layer should correspond to the values ​​input in the input layer. Layers 202, 204, and 206, or more precisely, the neurons contained in these layers, are connected to each other by (weighted) edges not shown further. The autoencoder encodes the input on the encoding side 208 and decodes the input on the decoding side 210 of the autoencoder 200. In one or more of these hidden layers 206, an encoding is formed that represents the input data in a reduced-dimensional form.

[0048] During training, the weights of the edges of the neural network's autoencoder are adapted to make the output values ​​at the neurons of the output layer 204 as close as possible to the input values ​​in terms of an error metric. To this end, these weights are modified to minimize the appropriate error metric. For example, the root mean square error of these individual values ​​can be used as the error metric.

[0049] exist Figure 2In the diagram, input neurons 220 and 222 are exemplarily shown in input layer 202, and output neurons 230 and 232 are shown in output layer 204. If the data record includes, for example, measurements of pressure and temperature (e.g., in a hydraulic system), the value of pressure p is input to one of the input neurons 220, symbolically represented by arrow 224, and the value of temperature T is input to the other input neuron 222, symbolically represented by arrow 226. The autoencoder (i.e., the model) maps these input values ​​to the input neurons, along with other input values ​​(not shown) corresponding to other operating parameters in the data record, to output values ​​on output neurons 230 and 232. In this case, output neuron 230 should output a value p' corresponding to the pressure, symbolically represented by arrow 234, and output neuron 232 should output a value T' corresponding to the temperature, symbolically represented by arrow 236. The autoencoder is trained to make the output values ​​p', T', and other output values ​​(not shown) as close as possible to their corresponding input values ​​p, T; that is, the reconstruction error should be as small as possible in terms of a suitable error metric. Here, these weights are modified during iterations to minimize the error metric. The extent of the weight modification in each iteration is determined by the learning rate, which may change between iterations, for example, decreasing with each iteration.

[0050] For an example of an autoencoder, the model's parameter configuration may include: the number of layers, the number of neurons in each of those layers, the description of the connections between these neurons (i.e., how the edges are directed), the initial values ​​of the weights, the learning rate, the maximum number of iterations (epochs) to be performed during training, and the error metric. Different values ​​of the model parameters included in the parameter configuration typically result in different models or algorithms after training. For example, even if the autoencoder structure (number of layers, number of neurons, connections / edges between these neurons) is the same and these autoencoders differ only in other parameters of the parameter configuration (such as initialization, learning rate, or error metric), different parameter configurations can result in autoencoders with different final weights.

[0051] It should be reiterated that the structural details of the model used are not necessary for the implementation of this invention. The autoencoder is merely one possibility for implementing a suitable model. Figure 2 It is appended primarily for the purpose of completeness and better understanding of the invention.

Claims

1. A method for validating at least one model for monitoring the state of a machine and selecting at least one model for monitoring the state of a machine from a first model set including at least one model, wherein the model is an algorithm trained by means of machine learning, the algorithm modeling the operating state of the machine based on operating parameters characterizing the operating state of the machine, wherein the model is trained by means of reference data records of the operating parameters and characterized by different parameter configurations for different models, and wherein the validation and selection are performed by means of unacceptable operating parameter values, the method comprising: a) Generate (54) one or more test data records based on the reference data record, wherein at least one value of at least one operating parameter in the reference data record is replaced by a corresponding unacceptable operating parameter value; b) For each model in the model set: The model is applied (56) to each of the one or more test data records to determine the error value of the model with respect to each test data record, and The distance value of the model (62) is determined based on the error value of the model; c) Select one or more models (64) based on the magnitude of the distance value. Step b) further includes: If at least one of the error values ​​is less than the model’s predetermined maximum error limit, the model is discarded (58).

2. The method of claim 1, wherein multiple models with the maximum distance value are selected.

3. The method of claim 2, wherein the model with the maximum distance value is selected.

4. The method according to any one of claims 1 to 3, wherein in step b), the distance value of the corresponding model is determined such that the distance value corresponds to a value selected from the following group: the minimum error value among the error values ​​of the model, the maximum error value among the error values ​​of the model, and the average error value of the model.

5. The method of claim 4, wherein the distance value is determined using the same method for all models.

6. The method according to any one of claims 1 to 3, further comprising: Definition and training (52) includes a first model set of at least one model, wherein different parameter configurations are selected for different models under the definition.

7. The method of any one of claims 1 to 3, further comprising: When all models are discarded: Definition (74) includes a second model set comprising at least one model, wherein different models are defined using different parameter configurations, and the parameter configurations of the models in the second model set are different from the parameter configurations of the models in the first model set; (76) The model of the second model set is trained using the reference data records; and Steps b) and c) are performed for the second model set.

8. The method according to any one of claims 1 to 3, wherein the parameter configuration of the model includes initialization parameters, learning rate and / or error metric.

9. The method according to any of claims 1 to 3, wherein the model is implemented as a neural network, wherein the parameter configuration comprises a number of layers, a number of neurons for each of the layers and / or a number of epochs.

10. The method according to claim 9, wherein the neural network is an autoencoder.

11. The method according to any of claims 1 to 3, the method further comprising: storing (66) the parameter configuration of the selected at least one model.

12. A computing unit, the computing unit being arranged to perform the method according to any of claims 1 to 11.

13. A computer program product comprising a computer program which, when executed on a computing unit, causes the computing unit to perform the method according to any of claims 1 to 11.

14. A machine-readable storage medium having stored thereon a computer program which, when executed on a computing unit, causes the computing unit to perform the method according to any of claims 1 to 11.

Images