Feature selection method and apparatus for industrial equipment time series

Abstract

This invention relates to the field of industrial equipment data processing technology, and embodiments provide a feature selection method and apparatus for industrial equipment time series. The feature selection method for industrial equipment time series includes: constructing a corresponding target dataset based on prior knowledge of the industrial equipment time series and determined target variables; acquiring the original dataset of the industrial equipment time series and constructing it as the input set for model training; and inputting the input set and the target dataset into I... M The filter consists of inner selectors based on the LSTM-SHAP model. The process involves obtaining the importance of each data feature in one inner selector to the target data; and then selecting a feature set as the output of the inner selector based on these importance results. This invention not only achieves dimensionality reduction of the original device dataset to reflect device state, but also preserves the temporal dependencies and correlations between features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment data processing, in particular to a feature selection method of industrial equipment time series, a feature selection device of industrial equipment time series, an electronic device and a corresponding storage medium. BACKGROUND

[0002] There are two challenges in the data mining of industrial equipment time series: (1) Lack of enough class label samples, which may induce the data-driven model to fail to learn the implicit knowledge in the data, resulting in poor application effect of the model. For example, in the abnormal detection task, it is difficult to collect samples of abnormal class labels, so that the detection model cannot fully learn the abnormal features during training, and the focus of learning obviously shifts to normal samples, resulting in poor abnormal detection effect of the model; (2) Unknown abnormalities existing in high-dimensional nonlinear industrial data may lead to insufficient understanding of the model on potential feature correlation. In the training and learning of the model, some implicit abnormal phenomena will imitate the benign behavior of normal samples, thereby misleading the judgment of the model. Therefore, in actual engineering, engineers need more suitable data processing means for industrial time series before carrying out data mining tasks to screen important features related to equipment production conditions.

[0003] In the face of the demand for feature selection, relevant operation personnel are accustomed to using expert experience to screen a feature subset highly related to an abnormal class. The data collected by industrial sensors usually contains physical meaning, so engineers can focus on the state of important physical indicators according to engineering experience to identify the running state of industrial equipment. However, this method ignores the potential relevance between indicators. For some large industrial equipment with complex structure and not fully transparent working principle, it is easy to ignore the implicit knowledge, and the result of feature selection is also unreliable. In addition, with the development of feature engineering technology, there are many feature selection methods for data classification, such as the classic recursive feature elimination algorithm (RFE), Boruta algorithm, or methods based on such algorithms combined with machine learning models such as SVM-RFE, XGBOOST-RFE, LGBM-Boruta, etc. These methods usually screen class-related features according to class labels. However, because the running records of industrial equipment generally lack abnormal labels, these label-based feature selection methods are limited when applied to small sample data sets, so they cannot well meet the feature selection needs in the field of industrial equipment time series anomaly detection. In addition, some researchers have proposed unsupervised feature selection methods based on no labels, including clustering-based methods and bio-inspired algorithms (such as genetic algorithms, ant colony algorithms), etc. However, these unsupervised feature selection methods either lack consideration of the temporal nature between the target variable and the classification-related features, or cannot handle a large amount of time flow data with cross-variable relationships, and are easy to ignore implicit abnormal points in the data set, so the feature data selected by them often cannot identify some anomalies or judge normal observation points as abnormal points in the detection model. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a feature selection method and device for industrial equipment time series, which addresses the problems of feature loss and the like in dimension reduction of industrial equipment data, and uses a feature selection method based on an LSTM-SHAP model to at least solve some of the problems in the background art.

[0005] To achieve the above-mentioned purpose, in the present application, a feature selection method for industrial equipment time series is provided, which comprises: constructing a corresponding target data set based on prior knowledge of industrial equipment time series and a determined target variable; obtaining an original data set of the industrial equipment time series and constructing it as an input set for model training; inputting the input set and the target data set into a filter composed of I M inner selectors based on LSTM-SHAP models, to obtain an importance result of each data feature in the inner selector to the target data; screening a feature set according to the importance result as the output result of the inner selector; and continuously executing I MAn inner layer selector obtains the output result of the current filter; and the iteration is performed I T The second filter calculates a selected coverage rate of the data features according to the recognition rate of the data features, and determines a data feature with a selected coverage rate higher than a preset threshold as a data feature of the time series of the industrial equipment.

[0006] Preferably, the target data set is in a matrix form, and the number of elements of the target data set is determined according to the number of observation samples and the number of target variables.

[0007] Preferably, the original data set of the time series of the industrial equipment is obtained and constructed as an input set for model training, including: performing random stretching expansion on the original data set to construct a shadow feature data set; and splicing the original data set and the shadow feature data set to construct the input set for model training.

[0008] Preferably, the importance result of each data feature to the target data in an inner layer selector is obtained, including: training a time series fitting device based on a multi-layer LSTM network layer by using the input set and the target data set; and calculating the importance score of the original feature to the training result by using a SHAP deep learning model interpreter, and the importance score calculation formula of the observation sample of the i th data feature to the target data is as follows:

[0009]

[0010] wherein, ξ i ∈R 1×n is the importance score matrix of the observation sample of the i th data feature, is the gradient approximation value matrix of the observation sample of the i th data feature of the constructed deep model network, F is the deep network function of the model, x i is the observation sample value of the i th data feature, is the baseline standard value of the model input, and the value of F at is close to 0, and m is the degree of Taylor expansion of the approximation value.

[0011] Preferably, a feature set is screened according to the importance result as the output result of the inner layer selector, including: taking the maximum value in the importance score set of the shadow feature in the shadow feature data set as a reference to screen a feature set with high importance; and obtaining a feature to be deleted based on a binomial distribution test; and the obtained feature set is taken as the output result of the inner layer selector.

[0012] Preferably, the selected coverage rate of the data features is calculated according to the identified rate of the data features, including: counting the number of times that the data features are identified as high importance in each output result of the filter; and calculating the selected coverage rate of the data features according to the following formula:

[0013]

[0014] wherein θi is the selected coverage rate of the ith data feature; representing whether the ith data feature is marked as high importance in the result of the jth filter, or I T representing the number of times that the filter is iteratively executed.

[0015] Preferably, the method further includes: obtaining a feature subset related to the state in the industrial equipment data by executing the inner selector and the filter in the feature selection network based on the LSTM-SHAP model, and applying the data corresponding to the feature subset to the equipment time series anomaly detection model VAE-LSTM, and comparing the selection result with that of other feature selection methods.

[0016] In the present application, an industrial equipment time series feature selection device is also provided, which includes: a data set construction module for constructing a corresponding target data set based on the prior knowledge of the industrial equipment time series and a determined target variable; an input set construction module for obtaining an original data set of the industrial equipment time series and constructing it into an input set for model training; an importance calculation module for inputting the input set and the target data set into I M inner selectors based on the LSTM-SHAP model to obtain the importance result of each data feature in the target data in each inner selector; a feature selection module for selecting a feature set as the output result of the inner selector according to the importance result; an execution control module for continuously executing I M inner selectors to obtain the output result of the filter; and a feature determination module for iteratively executing I T times of the filter, calculating the selected coverage rate of the data features according to the identified rate of the data features, and determining the data features with the selected coverage rate higher than a preset threshold as the data features of the industrial equipment time series.

[0017] Preferably, the target data set is in the form of a matrix, and the number of elements of the target data set is determined according to the number of observation samples and the number of target variables.

[0018] Preferably, the original data set of the industrial equipment time series is obtained and constructed as an input set for model training, including: performing random stretching expansion on the original data set to construct a shadow feature data set; and splicing the original data set and the shadow feature data set to construct the input set for model training.

[0019] Preferably, the importance result of each data feature in the inner layer selector to the target data is obtained, including: training a time series fitter based on a multi-layer LSTM network layer by using the input set and the target data set; and calculating the importance score of the original feature to the training result by a SHAP deep learning model interpreter, and the importance score of the observation sample of the i th data feature to the target data is calculated according to the following formula:

[0020]

[0021] Wherein, ξ i ∈R 1×n is the importance score matrix of the observation sample of the i th data feature, is the gradient approximation value matrix of the observation sample of the i th data feature of the constructed deep model network, F is the deep network function of the model, x i is the observation sample value of the i th data feature, is the baseline standard value of the model input, and the value of F at is close to 0, and m is the degree of Taylor expansion of the approximation value.

[0022] Preferably, the feature set is screened according to the importance result as the output result of the inner layer selector, including: taking the maximum value in the importance score set of the shadow feature in the shadow feature data set as a reference to screen a high importance feature set; and obtaining a feature to be deleted based on a binomial distribution test; and the obtained feature set is the output result of the inner layer selector.

[0023] Preferably, the selected coverage rate of the data feature is calculated according to the recognition rate of the data feature, including: statistically obtaining a time matrix in which the data feature is recognized as a high importance by the output result of each filter; and calculating the selected coverage rate of the data feature according to the following formula:

[0024]

[0025] Wherein, θi is the selected coverage rate of the i th data feature; represents whether the i th data feature is marked as high importance in the result of the j th filter, or I T represents the number of times of iterative execution of the filter.

[0026] Preferably, the device further comprises: obtaining a feature subset related to the state in the industrial equipment data through the inner layer selector and the filter in the feature selection network based on the LSTM-SHAP model, and applying the data corresponding to the feature subset to the equipment time series anomaly detection model VAE-LSTM, and comparing it with the selection results of other feature selection methods.

[0027] In the present application, an electronic device is also provided, comprising: at least one processor; a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the steps of the feature selection method for industrial equipment time series by executing the instructions stored in the memory.

[0028] In the present application, a machine readable storage medium is also provided, which stores instructions, and the instructions, when executed by a processor, cause the processor to be configured to implement the steps of the feature selection method for industrial equipment time series.

[0029] In the present application, a computer program product is also provided, which comprises a computer program, and the computer program, when executed by a processor, implements the steps of the feature selection method for industrial equipment time series.

[0030] The above technical solution has the following beneficial effects: compared with the prior art, the feature selection method for industrial equipment time series based on the LSTM-SHAP model in the present application further captures the time information in the features capable of describing the state of the equipment according to the existing knowledge related to the industrial equipment, such as physical, chemical or calculation indicators in engineering, and removes the features that are not important in the running process. Therefore, the method not only realizes the dimension reduction of the original data set of the equipment facing the state of the equipment, but also considers the time dependence and correlation between the reserved features.

[0031] Other features and advantages of the embodiments of the present application will be described in detail in the following specific implementation part. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings are included to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used together with the following specific implementation to explain the embodiments of the present application, but do not constitute a limitation on the embodiments of the present application. In the drawings:

[0033] Figure 1 The steps of the feature selection method for industrial equipment time series according to the embodiments of the present application are schematically shown;

[0034] Figure 2A structural schematic diagram of a feature selection device of an industrial equipment time series according to an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0035] The specific embodiments of the embodiments of the present application are described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely intended to illustrate and explain the embodiments of the present application, and are not intended to limit the embodiments of the present application.

[0036] Figure 1 A step schematic diagram of a feature selection method of an industrial equipment time series according to an embodiment of the present application is schematically shown. As shown in Figure 1 , a feature selection method of an industrial equipment time series, the method comprising:

[0037] S01, constructing a corresponding target data set based on prior knowledge of the industrial equipment time series and a determined target variable; based on engineering prior knowledge, preliminarily summarizing state-related target variables, and establishing a corresponding target data set Y t ∈R k×n , where n is the nth observation sample, and k is the kth target variable;

[0038]

[0039] S02, obtaining an original data set of the industrial equipment time series, and constructing it as an input set for model training; constructing the original data set X o ∈R m×n , performing random stretching expansion to construct a "shadow" feature data set X s ∈R m×n , where n is the nth observation sample, m is the mth feature in the data set, and X o and X s are spliced to construct an input set X t ∈R (m+m)×n for model training;

[0040]

[0041] S03, inputting the input set and the target data set into a filter composed of I M inner selectors composed of LSTM-SHAP models to obtain the importance of each data feature in the inner selector to the target data; inputting X t and the target data set Y t into the established filter for training. A filter is composed of I M inner selectors (IFS) composed of LSTM-SHAP models in series. In an inner selector (IFS), X t and Yt The time series fitter based on the multi-layer LSTM network layer is trained, and the importance score of the original features on the training result is calculated by the SHAP deep learning model interpreter GradientExplainer. In order to comprehensively consider the accuracy of feature selection results and reduce the training time of the model, the number I of memory feature selectors (IFS) is set M Preferably between 10 and 20. For example, in the implementation process of feature selection for the VMD dataset and the SMD_1 dataset, the number I of memory feature selectors (IFS) is set to M Set to 20.

[0042] S04, according to the importance result, a feature set is screened out as the output result of the inner selector; the maximum value in the importance score set of the constructed "shadow" feature is taken as a reference, and compared with the importance score of the real feature, a high importance feature set H is preliminarily screened out, and a feature set D after deleting unimportant features is obtained based on the binomial distribution test as the output result of the inner feature selector.

[0043] S05, the output result of the I M th inner selector is obtained; the output result is recorded as

[0044] S06, the I T th selector is iteratively executed, the selected coverage rate of the data features is calculated according to the recognition rate of the data features, and the data features with a selected coverage rate higher than a preset threshold are determined as the data features of the industrial equipment time series. The number of times that each feature is recognized as high importance is obtained to obtain a feature recognition frequency matrix R, and the selected coverage rate θ of the feature is calculated, and the features with θ > w are selected into the final feature subset, and the data features in the feature subset are the data features of the industrial equipment time series.

[0045] Through the above implementation, the time dependence and correlation between data features are retained while realizing the dimensionality reduction of the industrial equipment time series data.

[0046] In the embodiments provided by the present application, the target data set is in matrix form, and the number of elements of the target data set is determined according to the number of observation samples and the number of target variables. The number of target variables selected based on prior knowledge is generally not more than 5, that is, k is a positive integer less than or equal to 5. For example, in experiments, for a vertical roller mill production data set (VMD) and a data set of machine-1 of a public data platform of a network server operation record (SMD_1), the variable ZD_P of the main motor power of the mill is determined as the target variable of VMD according to engineering experience, and col_2, col_9, col_13, col_14 and col_15 are extracted as the target variables of the data set SMD_1 of the network server based on the server exception trace record.

[0047] In the embodiments provided by the present application, the original data set of the time series of the industrial equipment is obtained, and is constructed as the input set for model training, including: randomly stretching and expanding the original data set to construct a shadow feature data set; and splicing the original data set and the shadow feature data set to construct the input set for model training.

[0048] In the embodiments provided by the present application, the importance of each data feature to the target data in an inner selector is obtained, including: inputting X t and the target data set Y t into the established filter for training. One filter is composed of a series of I M inner selectors (IFS) based on LSTM-SHAP models, and I M is preferably a positive integer of 10-20. In an inner selector (IFS), X t and Y t are used to train a time series fitter based on a multi-layer LSTM network layer, and the importance score of the original feature to the training result is calculated through a SHAP deep learning model interpreter GradientExplainer. The importance score of the observation sample of the i-th feature to the target data is calculated according to the following formula:

[0049]

[0050] wherein, ξ i ∈R 1×n is the importance score matrix of the observation sample of the i-th data feature, is the gradient approximation matrix of the observation sample of the i-th data feature of the constructed deep model network, F is the deep network function of the model, x i is the observation sample value of the i-th data feature, is the baseline standard value of the model input, and F is in The value of the feature importance score is close to 0, m is the degree of the Taylor expansion of the approximation, and m is a constant of 80-100 in order to balance the accuracy of the feature importance score and the calculation efficiency of the GradientExplainer interpreter.

[0051] In the embodiment provided by the application, the feature set is screened according to the importance result as the output result of the inner layer selector, including: taking the maximum value in the importance score set of the shadow feature in the shadow feature data set as a reference, and comparing it with the importance score of the real feature, to preliminarily screen out a high importance feature set H, and obtaining a feature set D after deleting unimportant features based on a binomial distribution test as the output result of the inner layer feature selector.

[0052] In the embodiment provided by the application, the selected coverage of the data feature is calculated according to the recognition rate of the data feature, including: counting a time matrix of the data feature being recognized as a high importance each time the output result of the selector is obtained; and calculating the selected coverage of the data feature according to the following formula:

[0053]

[0054] wherein θi is the selected coverage of the i th data feature; representing whether the i th data feature is marked as high importance in the result of the j th selector, or I T representing the number of times of iterative execution of the selector. In order to retain a sufficient number of features and eliminate truly irrelevant features during training of the anomaly detection model, the threshold value w of the selected coverage θ of the features is preferably between 50% and 85%.

[0055] In the embodiment provided by the application, the method further includes: obtaining a feature subset related to the state in the industrial equipment data by executing the inner layer selector and the selector in the feature selection network based on the LSTM-SHAP model, and applying the data corresponding to the feature subset to the equipment time series anomaly detection model VAE-LSTM, and comparing it with the selection results of other feature selection methods.

[0056] Exemplarily, the feature selection method for industrial equipment time series based on the LSTM-SHAP model (ILSFS) of the application has obvious improvement in the ranking of the selected coverage results of the features of the VMD data set and the SMD_1 data set compared with other methods such as SVM-RFE, LGBM-Boruta and genetic algorithm (GA). In the specific implementation process of the experiment, the threshold value w is set to 50%, and the final selected high importance feature subset result is shown in Table 1.

[0057] Table 1 Feature subsets after feature selection of VMD dataset and SMD_1 dataset by ILSFS of the present application and other methods SVM-RFE, LGBM-Boruta, genetic algorithm (GA)

[0058]

[0059]

[0060] In step S7, the feature data corresponding to the feature subset obtained by the above method is used in the VAE-LSTM anomaly detection model to verify the effectiveness of the ILSFS method of the present application, and the anomaly detection results of the VAE-LSTM model on the VMD and SMD_1 datasets after feature selection are shown in Table 2. As can be seen from the F1 score results in Table 2, the higher the F1 score, the better the effect of anomaly detection, that is, the feature selection method (ILSFS) of the industrial equipment anomaly detection based on the LSTM-SHAP model of the present application has better effect on time series anomaly detection than other feature selection methods after feature selection, and the method of the present application is verified in two real device datasets.

[0061] Table 2 Results of VAE-LSTM anomaly detection model on VMD and SMD_1 datasets processed by different feature selection methods

[0062]

[0063]

[0064] As can be seen from the above embodiments or examples, the present application not only realizes the dimensionality reduction of the original dataset of the industrial equipment for the equipment state, but also considers the preservation of the time dependence and correlation between features.

[0065] Based on the same inventive concept, the present application also provides an industrial equipment time series feature selection device. Figure 2 The structure of the industrial equipment time series feature selection device according to the embodiment of the present application is schematically shown. As shown in Figure 2 The device comprises: a dataset construction module for constructing a corresponding target dataset based on prior knowledge of the industrial equipment time series and a determined target variable; an input set construction module for obtaining an original dataset of the industrial equipment time series and constructing it into an input set for model training; an importance calculation module for inputting the input set and the target dataset into an I Man inner selector based on an LSTM-SHAP model, to obtain an importance result of each data feature in the inner selector on the target data; a feature screening module, configured to screen a feature set as an output result of the inner selector according to the importance result; an execution control module, configured to continuously execute the I M an inner selector to obtain an output result of the current filter; and a feature determination module, configured to iteratively execute the I T a filter, to calculate a selected coverage rate of the data feature according to the recognition rate of the data feature, and determine the data feature with a selected coverage rate higher than a preset threshold as a data feature of the industrial equipment time series.

[0066] In some optional embodiments, the target data set is in a matrix form, and the number of elements of the target data set is determined according to the number of observation samples and the number of target variables.

[0067] In some optional embodiments, the original data set of the industrial equipment time series is obtained and constructed as an input set for model training, including: performing random stretching expansion on the original data set to construct a shadow feature data set; and splicing the original data set and the shadow feature data set to construct the input set for model training.

[0068] In some optional embodiments, the importance result of each data feature in the inner selector on the target data is obtained, including: training a time series fitter based on a multi-layer LSTM network layer using the input set and the target data set; calculating the importance score of the original feature on the training result by a SHAP deep learning model interpreter, and the importance score of the i th data feature on the target data is calculated according to the following formula:

[0069]

[0070] wherein, ξ i ∈R 1×n is the importance score matrix of the i th data feature observation sample, is the gradient approximation value matrix of the i th data feature observation sample of the constructed deep model network, F is the deep network function of the model, x i is the i th data feature observation sample value, is the baseline standard value of the model input, and the value of F at is close to 0, and m is the degree of Taylor expansion of the approximation value.

[0071] In some optional embodiments, the feature set is screened according to the importance result as the output result of the inner selector, including: taking the maximum value in the importance score set of the shadow feature in the shadow feature data set as a benchmark, screening a feature set with high importance; and obtaining a feature with low importance based on a binomial distribution test; and the feature set obtained thereby is the output result of the inner selector.

[0072] In some optional embodiments, the selected coverage of the data feature is calculated according to the recognition rate of the data feature, including: counting the number of times that the data feature is recognized as a feature with high importance in each output result of the filter; and calculating the selected coverage of the data feature according to the following formula:

[0073]

[0074] wherein, θi is the selected coverage of the i th data feature; represents whether the i th data feature is marked as a feature with high importance in the j th output result of the filter, or I T represents the number of times of iterative execution of the filter.

[0075] In some optional embodiments, the device further includes: obtaining a feature subset related to the state in the industrial equipment data by executing the inner selector and the filter in the feature selection network based on the LSTM-SHAP model, applying the feature subset to the equipment time series anomaly detection model VAE-LSTM, and comparing the feature subset with the selection results of other feature selection methods.

[0076] The specific limitations of each functional module in the above industrial equipment time series feature selection device can be referred to the limitations of the industrial equipment time series feature selection method in the above, which will not be repeated here. Each module in the above device can be realized by software, hardware and their combinations in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0077] In some embodiments of the present application, an electronic device is also provided, comprising: at least one processor; a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor executes the steps of the aforementioned feature selection method for industrial equipment time series by executing the instructions. The control module or processor herein has the functions of numerical calculation and logical operation, and has at least a central processing unit (CPU) with data processing capability, a random access memory (RAM), a read-only memory (ROM), various I / O ports, an interrupt system, and the like. The processor includes a core, which retrieves corresponding program units from the memory. The core can be one or more, and the aforementioned method is realized by adjusting the core parameters. The memory can include non-permanent memory in a computer readable medium, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one memory chip.

[0078] In an embodiment provided by the present application, a machine readable storage medium is provided, and the machine readable storage medium stores instructions, which, when executed by a processor, cause the processor to be configured to execute the steps of the aforementioned feature selection method for industrial equipment time series.

[0079] In an embodiment provided by the present application, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the steps of the aforementioned feature selection method for industrial equipment time series.

[0080] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0081] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocksFigure 1 means for performing the function specified by the block or blocks.

[0082] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a Figure 1 one or more processes and / or blocks Figure 1 means for performing the function specified by the block or blocks.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 one or more processes and / or blocks Figure 1 means for performing the function specified by the block or blocks.

[0084] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0085] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory. The memory can also include non-volatile memory, such as read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other non-volatile memory. Memory is an example of computer readable media.

[0086] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile discs (DVDs) or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0087] It should also be noted that the terms "comprising", "comprises" or other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0088] The above embodiments are only used to illustrate the present application, but not to limit it. Instead of the above, various modifications and changes can be made to the application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall fall into the scope of the claims of the application.

Claims

1. A feature selection method of an industrial equipment time series, characterized by, The method comprises: constructing a corresponding target data set based on prior knowledge of the industrial equipment time series and a determined target variable; obtaining an original data set of the industrial equipment time series and constructing it as an input set for model training; inputting the input set and the target data set into an inner selector composed of an LSTM-SHAP model to obtain an importance result of each data feature in the inner selector to the target data M inputting the input set and the target data set into an inner selector composed of an LSTM-SHAP model to obtain an importance result of each data feature in the inner selector to the target data screening a feature set as an output result of the inner selector according to the importance result; Continuous execution I M An inner layer selector obtains the output result of the current filter. Iterative execution I T The secondary filter calculates the selected coverage rate of the data features according to the identified rate of the data features, and determines the data features with the selected coverage rate higher than a preset threshold as the data features of the industrial equipment time series.

2. The method of claim 1, wherein, The target data set is in matrix form, and the number of elements of the target data set is determined according to the number of observation samples and the number of target variables.

3. The method of claim 1, wherein, Obtaining an original data set of the industrial equipment time series and constructing it as an input set for model training comprises: randomly stretching and expanding the original data set to construct a shadow feature data set; splicing the original data set and the shadow feature data set to construct the input set for model training.

4. The method of claim 1, wherein, Obtaining an importance result of each data feature in an inner selector to target data comprises: training a time series fitting device composed of multiple LSTM network layers using the input set and the target data set; calculating the importance score of the original feature to the training result by a SHAP deep learning model interpreter, and the importance score calculation formula of the i th data feature to the target data is as follows: where ξ i ∈ 1×n is a matrix of importance scores of the observed samples of the i-th data feature, is a matrix of gradient approximations of the observed samples of the i-th data feature of the constructed deep model network, F is the deep network function of the model, x i is the observed sample value of the i-th data feature, is the baseline standard value of the model input, and F is close to 0 at m is the degree of the Taylor expansion of the approximation.

5. The method of claim 3, wherein, Screening a feature set as an output result of the inner selector according to the importance result comprises: taking the maximum value in the importance score set of the shadow feature in the shadow feature data set as a reference to screen a feature set with high importance; and obtaining a feature to be deleted based on a binomial distribution test; The feature set obtained is the output result of the inner selector.

6. The method of claim 1, wherein, Calculating the selected coverage rate of the data feature according to the recognition rate of the data feature comprises: statistically obtaining a time matrix in which the data feature is recognized as a high importance feature each time the output result of the selector is obtained; calculating the selected coverage rate of the data feature according to the following formula: where θi is the selected coverage of the ith data feature; represents whether the ith data feature is marked as high importance in the result of the jth filter, or I T represents the number of times the filter is iteratively executed.

7. The method of claim 1, wherein, The method further comprises: obtaining a feature subset related to the state of the industrial equipment data by executing the inner selector and the selector in the feature selection network based on the LSTM-SHAP model, applying the data corresponding to the feature subset to the equipment time series anomaly detection model VAE-LSTM, and comparing the selection result with that of other feature selection methods.

8. An industrial device time series feature selection apparatus characterized by comprising: The device comprises: a data set construction module configured to construct a corresponding target data set based on prior knowledge of the industrial equipment time series and a determined target variable; an input set construction module configured to obtain an original data set of the industrial equipment time series and construct it as an input set for model training; An importance calculation module is configured to input the input set and the target data set into an inner selector composed of an LSTM-SHAP model to obtain an importance result of each data feature in the inner selector on the target data. M An importance calculation module is configured to input the input set and the target data set into an inner selector composed of an LSTM-SHAP model to obtain an importance result of each data feature in the inner selector on the target data. a feature screening module configured to screen a feature set as an output result of the inner selector according to the importance result; An execution control module is configured to continuously execute the I M An inner layer selector obtains the output result of the current filter; and The feature determination module is configured to iteratively perform I T The secondary filter calculates a selected coverage rate of the data features according to the identified rate of the data features, and determines a data feature with a selected coverage rate higher than a preset threshold as a data feature of the industrial equipment time series.

9. An electronic device, comprising: comprise: at least one processor; a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the steps of the feature selection method of the industrial equipment time series according to any one of claims 1 to 7 by executing the instructions stored in the memory.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium has stored thereon instructions which, when executed by the processor, cause the processor to be configured to implement the feature selection method of the industrial device time series of any one of claims 1 to 7.

Images