Equipment component failure mode identification method based on multi-source uncertain sparse samples

Abstract

The present application discloses a method for equipment component failure mode identification for multi-source uncertain sparse samples, which relates to the technical field of equipment failure identification. The method comprises: considering that the equipment contains multi-source uncertain noise during actual service, applying model-independent meta-learning to an adaptive filter to filter the multi-source uncertain noise; then combining short tree transformation with a dual-stream spatiotemporal attention deep neural network to obtain time-frequency and multi-scale features from multi-channel data, thereby enhancing the ability to extract failure features and reducing the dependence of failure identification on data volume; finally, applying a least squares support vector machine for failure mode identification, and adopting a crested porcupine algorithm to determine the optimal hyperparameter combination, thereby improving the failure identification capability of the model under complex working conditions during actual service of equipment components, and effectively solving the problem in the prior art that failure modes of sparse data containing multi-source uncertain noise are difficult to identify.

Description

Technical Field

[0001] The present application relates to the technical field of equipment failure identification, and in particular to a method for identifying equipment component failure modes for multi-source uncertain sparse samples. Background Art

[0002] In modern complex equipment systems, failure mode identification of core components is an important part of ensuring system reliability and safety. However, existing research faces many challenges in the identification of equipment component failure modes. First, data sparsity and imbalance problems significantly affect the identification accuracy of the model, and traditional methods are difficult to effectively handle the situation where failure data is sparse. Secondly, the various uncertain noises in the actual service environment, such as environmental noise, electromagnetic interference, and mechanical vibration noise, are complex and diverse, which interfere with data quality and increase the difficulty of accurately extracting useful features. In addition, multi-source heterogeneous data fusion is also a major challenge. The monitoring data of core components in modern equipment systems comes from multiple sensors. The data types and sources are heterogeneous and complex, making it difficult to effectively fuse and extract comprehensive features.

[0003] In terms of feature extraction and data representation, existing methods have great limitations. It is difficult to fully capture the time-frequency domain features and their detailed and directional features in the data. The data representation method is single and cannot fully utilize the time dependency and multi-scale characteristics of the data. In addition, model optimization and real-time issues also plague researchers. Traditional optimization methods such as grid search and random search are inefficient and it is difficult to quickly find the optimal parameter combination, which affects the real-time and application effects of the model. Finally, the existing models are not robust enough under complex working conditions and changing environments. They are easily affected by noise and sparse data and it is difficult to maintain stable identification performance in changing actual environments. These problems urgently need new methods and technologies to overcome in order to improve the effectiveness and practicality of failure mode identification. Summary of the Invention

[0004] The purpose of this application is to provide an equipment component failure mode identification method for multi-source uncertain sparse samples, which can effectively improve the effect and practicality of failure mode identification of sparse data containing multi-source uncertain noise.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] This application provides a method for equipment component failure mode identification based on multi-source uncertain sparse samples, including:

[0007] Acquire multi-source sensor data to be tested; the multi-source sensor data to be tested is multi-channel working data of equipment components with unclear failure modes.

[0008] An adaptive active noise reduction filter is used to filter out multi-source uncertain noise in the multi-source sensor data to be measured, thereby obtaining filtered multi-source sensor data to be measured; the adaptive active noise reduction filter is a filter obtained by optimizing the filter coefficients of the adaptive filter through a model-independent meta-learning algorithm; the adaptive active noise reduction filter can adapt to different noise environments by adjusting and updating the filter parameters in real time.

[0009] The filtered multi-source sensor data to be tested is input into the short tree transform algorithm to extract a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix; each element in the first image represents the signal strength at a specific time and frequency; each element in the second image represents the signal feature at a specific scale and direction.

[0010] The first image and the second image are processed in parallel through a dual-stream spatiotemporal attention deep neural network, and the time-frequency domain features and detail and directional features are extracted and fused respectively to obtain enhanced fusion features of failure mode data; the dual-stream spatiotemporal attention deep neural network includes a first branch and a second branch in parallel; the first branch is used to process the first image and extract the time-frequency domain features; the second branch is used to process the second image and extract the detail and directional features.

[0011] The enhanced fusion features of the failure mode data are input into the failure mode identification model to obtain the corresponding failure mode identification results; the failure mode identification model is a least squares support vector machine whose hyperparameters have been optimized in advance using the crested porcupine algorithm.

[0012] Optionally, before using an adaptive active noise reduction filter to filter out the multi-source uncertain noise in the multi-source sensor data set to be measured to obtain a filtered multi-source sensor data set to be measured, the equipment component failure mode identification method for multi-source uncertain sparse samples further includes the following steps:

[0013] The optimal initial control filter is found through model-independent meta-learning; in the model-independent meta-learning algorithm, the hyperparameters in the model-independent meta-learning algorithm are optimized through the grid search algorithm.

[0014] The filter coefficients of the optimal initial control filter are optimized to obtain an adaptive active noise reduction filter.

[0015] Optionally, the filter coefficients of the optimal initial control filter are optimized according to the following formula:

[0016]

[0017] Where ω(p+1) is the filter coefficient at the p+1th step, ω(p) represents the filter coefficient at the pth step, E(p) is the input vector at the pth step, the superscript T represents the transpose operation, e(p) is the error between the expected output vector and the actual output vector at the pth step, β is the learning rate, γ is the time decay factor, and N is the historical step size. e(p) is calculated according to the following formula:

[0018] e(p)=Φ(p)-E T (p)ω(p).

[0019] Where Φ(p) is the expected output vector at step p.

[0020] The multi-source uncertain noise in the multi-source sensor data to be measured is filtered out according to the following formula:

[0021]

[0022] Among them, x(p) is the multi-source sensor data to be measured, is the multi-source sensor data to be measured after filtering.

[0023] Optionally, optimize the hyperparameters in the model-agnostic meta-learning algorithm according to the following formula:

[0024]

[0025] Among them, β* is the optimal learning rate, γ* is the optimal time decay factor, K is the number of cross-validation folds, β is the learning rate, γ is the time decay factor, is the loss function, The performance of the model-independent meta-learning algorithm under the hyperparameter combination (β, γ).

[0026] Optionally, the short tree transform algorithm includes a short-time Fourier transform and a dual-tree complex wavelet transform; inputting the filtered multi-source sensor data to be measured into the short tree transform algorithm, and extracting a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix, specifically includes the following steps:

[0027] The filtered multi-source sensor data to be measured is subjected to short-time Fourier transform to obtain a time-frequency feature matrix of the time series signal; each element in the time-frequency feature matrix represents the signal strength at a specific time and frequency.

[0028] The filtered multi-source sensor data to be measured are subjected to dual-tree complex wavelet transform to obtain a multi-scale feature matrix; the elements of each feature image represent the signal characteristics at a specific scale and in a specific direction.

[0029] The time-frequency feature matrix is ​​arranged according to the time and frequency axes to form a first image including the time-frequency feature matrix.

[0030] The multi-scale feature matrix is ​​arranged according to the scale and direction axes to form a second image including the multi-scale feature matrix.

[0031] Optionally, the structures of the first branch and the second branch are the same; the first branch includes a convolutional neural network, a bidirectional gated recurrent unit and a spatiotemporal attention mechanism; the first branch extracts spatial features through the convolutional neural network, the bidirectional gated recurrent unit extracts time-dependent features, and introduces the spatiotemporal attention mechanism to extract spatiotemporal coupling features; the spatial attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the convolutional neural network to enhance the region of interest in the spatial feature map; the temporal attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the bidirectional gated recurrent unit to enhance the region of interest in the temporal feature sequence.

[0032] Alternatively, the expression of the spatial attention mechanism is as follows:

[0033]

[0034] Among them, H, W and C represent the height, width and number of channels of the feature image respectively, W s and b s Denote the weight and bias of the spatial attention mechanism, X ijct represents the value of the feature image at position (i, j), channel c and time step t, S ijct is the spatial attention weight of the position.

[0035] The expression of the temporal attention mechanism is as follows:

[0036]

[0037] Among them, W t and b t Represent the weight and bias of the temporal attention mechanism, Y ijct Represents the value of the feature image at position (i, j), channel c and time step t, T ijct Represents the temporal attention weight of the position.

[0038] Optionally, before inputting the enhanced fusion features of the failure mode data into the failure mode identification model to obtain the corresponding failure mode identification results, the equipment component failure mode identification method for multi-source uncertain sparse samples further includes:

[0039] The crown porcupine algorithm is used to initialize the particle swarm, randomly generate the hyperparameter combination of the least squares support vector machine, and continuously update the speed and position of the particles through iterative optimization, finally finding the optimal hyperparameter combination.

[0040] Under the optimal hyperparameter combination, the enhanced fusion feature dataset is used to train the least squares support vector machine model to obtain the failure mode identification model; the enhanced fusion feature dataset includes several enhanced fusion features of failure mode data and corresponding failure mode labels.

[0041] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0042] The present application provides a method for identifying failure modes of equipment components for multi-source uncertain sparse samples, which obtains multi-channel working data of equipment components with unclear failure modes, applies model-independent meta-learning to an adaptive filter to filter out multi-source uncertain noise, and then inputs the filtered data into a short tree transform algorithm to extract a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix. The first image and the second image are processed in parallel by a dual-stream spatiotemporal attention deep neural network, and the time-frequency domain features and detail and directional features are respectively extracted and fused to obtain enhanced fusion features of failure mode data; the enhanced fusion features of the failure mode data are input into a least squares support vector machine whose hyperparameters have been optimized in advance using the crested porcupine algorithm to obtain corresponding failure mode identification results. This application takes into account the multi-source uncertain noise contained in the actual service process of equipment, and applies a meta-learning-based algorithm to filter the multi-source uncertain noise to reduce the interference of noise signals on sparse failure data; then combines the short tree transform with the dual-stream spatiotemporal attention deep neural network to obtain time-frequency and multi-scale features from multi-channel data, enhance the ability to extract failure features, and reduce the dependence of failure identification on data volume; finally, the least squares support vector machine is applied to failure mode identification, and the crown porcupine algorithm is used to determine the optimal hyperparameter combination, which improves the failure identification ability of the model under the complex working conditions of the actual service of equipment components and is suitable for engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 A flowchart of a component failure mode identification method for multi-source uncertain sparse samples provided in one embodiment of the present application.

[0045] Figure 2 This is a flowchart of step S3 in a method for identifying equipment component failure modes based on multi-source uncertain sparse samples provided in one embodiment of the present application.

[0046] Figure 3 An overall flow chart of an equipment component failure mode identification method for multi-source uncertain sparse samples provided in another embodiment of the present application.

[0047] Figure 4 A schematic diagram of a two-dimensional image obtained by short-time Fourier transform in a method for equipment component failure mode identification for multi-source uncertain sparse samples provided in another embodiment of the present application.

[0048] Figure 5 A schematic diagram of a two-dimensional image obtained by dual-tree complex wavelet transform in an equipment component failure mode identification method for multi-source uncertain sparse samples provided in another embodiment of the present application.

[0049] Figure 6 A schematic diagram of a confusion matrix for identification accuracy in an equipment component failure mode identification method for multi-source uncertain sparse samples provided in another embodiment of the present application. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0051] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0052] In an exemplary embodiment, Figure 1 As shown, a method for identifying equipment component failure modes based on multi-source uncertain sparse samples is provided, comprising the following steps S1 to S5:

[0053] S1. Acquire multi-source sensor data to be tested; the multi-source sensor data to be tested is multi-channel operating data of an equipment component with an unclear failure mode. Specifically, when the equipment component is a rolling bearing, the failure modes of the equipment component include inner race failure, outer race failure, and rolling element failure. The types of multi-channel operating data of the equipment component include drive-end data, fan-end data, and acceleration data.

[0054] S2. An adaptive active noise reduction filter is used to filter out the multi-source uncertain noise in the multi-source sensor data to be measured, so as to obtain the filtered multi-source sensor data to be measured; the adaptive active noise reduction filter is a filter obtained by optimizing the filter coefficients of the adaptive filter through a model-independent meta-learning algorithm; the adaptive active noise reduction filter can adapt to different noise environments by adjusting and updating the filter parameters in real time.

[0055] S3. Input the filtered multi-source sensor data to be measured into the short tree transform algorithm to extract a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix; each element in the first image represents the signal strength at a specific time and frequency; each element in the second image represents the signal feature at a specific scale and direction. The short tree transform algorithm includes short-time Fourier transform and dual-tree complex wavelet transform; Figure 2 In the flowchart shown, step S3 specifically includes the following steps:

[0056] S31. Perform short-time Fourier transform on the filtered multi-source sensor data to be measured to obtain a time-frequency feature matrix of the time series signal; each element in the time-frequency feature matrix represents the signal strength at a specific time and frequency.

[0057] S32. Perform a dual-tree complex wavelet transform on the filtered multi-source sensor data to obtain a multi-scale feature matrix; the elements of each feature image represent the signal characteristics at a specific scale and in a specific direction. The dual-tree complex wavelet transform in this step specifically constructs two parallel wavelet transform trees to process the real and imaginary parts of the input data respectively, with good translation invariance and directional selectivity; its expression is shown in the following formula:

[0058]

[0059] Among them, W x (e, f) are wavelet coefficients, e and f are complex wavelet parameters in scale and direction respectively, ψ is the wavelet mother function, * represents the complex conjugate, x(t) is the input data, and t is time.

[0060] S33. Arrange the time-frequency feature matrix according to the time and frequency axes to form a first image including the time-frequency feature matrix.

[0061] S34. Arrange the multi-scale feature matrix according to the scale and direction axis to form a second image including the multi-scale feature matrix.

[0062] S4. The first image and the second image are processed in parallel through a dual-stream spatiotemporal attention deep neural network, and the time-frequency domain features and detail and directional features are extracted and fused respectively to obtain enhanced fusion features of failure mode data; the dual-stream spatiotemporal attention deep neural network includes a first branch and a second branch in parallel; the first branch is used to process the first image and extract the time-frequency domain features; the second branch is used to process the second image and extract the detail and directional features.

[0063] In the dual-stream spatiotemporal attention deep neural network, the first branch and the second branch have the same structure; the first branch includes a convolutional neural network, a bidirectional gated recurrent unit, and a spatiotemporal attention mechanism; the first branch extracts spatial features through the convolutional neural network, the bidirectional gated recurrent unit extracts time-dependent features, and the spatiotemporal attention mechanism is introduced to extract spatiotemporal coupling features; the spatial attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the convolutional neural network to enhance the region of interest in the spatial feature map; the expression of the spatial attention mechanism is shown in the following formula:

[0064]

[0065] Among them, H, W and C represent the height, width and number of channels of the feature image respectively, W s and b s Denote the weight and bias of the spatial attention mechanism, X ijct represents the value of the feature image at position (i, j), channel c and time step t, S ijct is the spatial attention weight of the position.

[0066] The temporal attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the bidirectional gated recurrent unit to enhance the region of interest in the temporal feature sequence; the expression of the temporal attention mechanism is shown in the following formula:

[0067]

[0068] Among them, W t and b t Represent the weight and bias of the temporal attention mechanism, Y ijct Represents the value of the feature image at position (i, j), channel c and time step t, T ijct Represents the temporal attention weight of the position.

[0069] S5. Inputting the enhanced fusion features of the failure mode data into the failure mode identification model to obtain the corresponding failure mode identification results; the failure mode identification model is a least squares support vector machine whose hyperparameters have been optimized in advance using the crested porcupine algorithm.

[0070] In another exemplary embodiment of the present application, before step S2, the method provided in this embodiment further includes the following steps:

[0071] The optimal initial control filter is found through model-independent meta-learning. In the model-independent meta-learning algorithm, the hyperparameters in the model-independent meta-learning algorithm are optimized through a grid search algorithm. Specifically, the hyperparameters in the model-independent meta-learning algorithm are optimized according to the following formula:

[0072]

[0073] Among them, β* is the optimal learning rate, γ* is the optimal time decay factor, K is the number of cross-validation folds, β is the learning rate, γ is the time decay factor, is the loss function, The performance of the model-independent meta-learning algorithm under the hyperparameter combination (β, γ).

[0074] The filter coefficients of the optimal initial control filter are optimized to obtain an adaptive active noise reduction filter. Specifically, the filter coefficients of the optimal initial control filter are optimized according to the following formula:

[0075]

[0076] Where ω(p+1) is the filter coefficient at the p+1th step, ω(p) represents the filter coefficient at the pth step, E(p) is the input vector at the pth step, the superscript T represents the transpose operation, e(p) is the error between the expected output vector and the actual output vector at the pth step, β is the learning rate, γ is the time decay factor, and N is the historical step size. e(p) is calculated according to the following formula:

[0077] e(p)=Φ(p)-E T (p)ω(p).

[0078] Where Φ(p) is the expected output vector at step p.

[0079] The multi-source uncertain noise in the multi-source sensor data to be measured is filtered out according to the following formula:

[0080]

[0081] Among them, x(p) is the multi-source sensor data to be measured, is the multi-source sensor data to be measured after filtering.

[0082] In another exemplary embodiment of the present application, before step S5, the method provided in this embodiment further includes the following steps:

[0083] The crown porcupine algorithm is used to initialize the particle swarm, randomly generate the hyperparameter combination of the least squares support vector machine, and continuously update the speed and position of the particles through iterative optimization, finally finding the optimal hyperparameter combination.

[0084] Under the optimal hyperparameter combination, the enhanced fusion feature dataset is used to train the least squares support vector machine model to obtain the failure mode identification model; the enhanced fusion feature dataset includes several enhanced fusion features of failure mode data and corresponding failure mode labels.

[0085] The above embodiment illustrates a method for identifying equipment component failure modes based on multi-source uncertain sparse samples from the perspective of actual use. In the following embodiment, further explanation is given from the perspective of training. The overall process is as follows: Figure 3 As shown in the figure, the upper dotted box is the process of the model training stage, and the lower dotted box is the process of the actual failure identification stage; Figure 3 It can be seen that before using the above-mentioned equipment component failure mode identification method for multi-source uncertain sparse samples to perform actual failure mode identification, it is necessary to first design and train the algorithm model, which specifically includes the following steps:

[0086] A1. Perform accelerated failure tests on the equipment, collect multi-channel operating data of the core components throughout their life cycle, label the data according to failure categories, and obtain a sparse sample label dataset. In step A1: simulate extreme operating conditions of the bearings through accelerated failure tests, and collect multi-channel operating data throughout their life cycle. Data types include drive-end vibration data, fan-end vibration data, etc. The collected data are labeled according to known failure modes to construct an accelerated failure dataset. Data of bearings with diameters of 0.007, 0.014, and 0.021 are collected and processed respectively. Under the same data type, 30 groups of data with different failure modes are randomly selected, including 10 groups with different bearing diameters under each type of failure mode. The data under this data type are divided into 80 training groups and 10 testing groups.

[0087] A2. Preprocess the data using a meta-detection algorithm to address the multi-source uncertain noise contained in accelerated failure testing of equipment, obtaining a sparse sample dataset after adaptively filtering the multi-source uncertain noise. In this embodiment, step A2 specifically involves preprocessing the data using a meta-detection algorithm to address the multi-source uncertain noise contained in accelerated failure testing of equipment. First, model-independent meta-learning is used to find the optimal initial control filter, which accelerates convergence and reduces the resources consumed by manual adjustments. By adjusting the filter parameters in real time, the multi-source uncertain noise in the data is removed, resulting in a filtered sparse sample dataset, providing a high-quality data foundation for subsequent failure mode identification.

[0088] Model-agnostic meta-learning is an algorithm that optimizes initial parameters to enable the model to quickly adapt to different tasks. Its goal is to find good initial parameters so that the model can perform well on new tasks after a small number of gradient updates. The update control filter is defined as:

[0089] e(p)=Φ(p)-E T (p)ω(p).

[0090] Update initial value is defined as:

[0091]

[0092] The grid search algorithm selects the best hyperparameters by exhaustively enumerating all possible hyperparameter combinations and performing cross-validation on each combination. In the model-independent meta-learning algorithm, grid search is used to select hyperparameters in the model-independent meta-learning algorithm, which can be specifically expressed as:

[0093]

[0094] Among them, (β*, γ*) is the optimal hyperparameter combination, (β, γ) is the hyperparameter combination to be optimized, K is the number of cross-validation folds, is the performance of the model under the hyperparameter combination (β, γ).

[0095] The adaptive filter coefficients are optimized through a model-independent meta-learning algorithm to obtain an adaptive active noise reduction filter that is adaptive to multiple types of uncertain noise. The adaptive active noise reduction filter adjusts and updates the filter parameters in real time to adapt to different noise environments. The meta-identification algorithm is applied to filter the multi-source sensor data set containing multi-source uncertain noise. Specifically:

[0096]

[0097] in, represents the output vector after filtering, ω T (p) represents the current weight parameter of the filter, and x(p) represents the input signal vector.

[0098] A3. Process and transform the multi-channel data, input the filtered data into a short tree transform, and convert it into a two-dimensional image containing time features, thereby constructing two independent sparse sample image datasets with time dependence. Step A3 is specifically as follows:

[0099] The filtered multi-channel data is input into the short tree transform, and the time-frequency features and multi-scale features of the signal are extracted using short-time Fourier transform and dual-tree complex wavelet transform, and converted into a two-dimensional image containing time features. Finally, two independent sparse sample image data sets with time dependence are constructed to provide data feature information for subsequent feature extraction.

[0100] The short-tree algorithm includes short-time Fourier transform and dual-tree complex wavelet transform. The dual-tree complex wavelet transform constructs two parallel wavelet transform trees to process the real and imaginary signals respectively, and has good translation invariance and directional selectivity. Its expression is:

[0101]

[0102] Among them, W x (e, f) are wavelet coefficients, e and f are complex wavelet parameters in scale and position respectively, x(t) is the input data, t is time, ψ is the wavelet mother function, and * represents the complex conjugate.

[0103] Perform a short-time Fourier transform on the data in A2 to obtain the time-frequency feature matrix of the time series signal. Each element of the time-frequency feature matrix represents the signal strength at a specific time and frequency. Simultaneously, perform a dual-tree complex wavelet transform to obtain a multi-scale feature image. Each element of the feature image represents the signal characteristics at a specific scale and direction.

[0104] The time-frequency feature matrix of each sample subjected to short-time Fourier transform is arranged according to the time and frequency axis to form a two-dimensional image, and the multi-scale feature matrix of each sample subjected to dual-tree complex wavelet transform is arranged according to the scale and direction axis to form a two-dimensional image. Figures 4 and 5 As shown in Figure 2, two-dimensional images obtained by short-time Fourier transform and dual-tree complex wavelet transform are shown, respectively, which are two independent sparse sample image data sets with time dependence.

[0105] A4: Feature extraction is performed on the sparse failure data. A dual-stream spatiotemporal attention deep neural network is constructed to process two different types of image sets in parallel. Time-frequency domain features and detail and directional features are extracted respectively. The two types of features are then fused to obtain a failure data feature-enhanced fusion dataset. Step A4 is as follows:

[0106] To extract features from sparse failure data, a two-stream spatiotemporal attention deep neural network is constructed to process two different image sets in parallel. The first branch processes the image set based on the short-time Fourier transform, extracting time-frequency features; the second branch processes the image set based on the dual-tree complex wavelet transform, extracting detail and directional features. The branch extracts spatial features using a convolutional neural network, temporally dependent features using a bidirectional gated recurrent unit, and spatiotemporal attention mechanisms to extract spatiotemporal coupling features. Finally, the features of the two branches are fused to produce a feature-enhanced fused dataset, providing rich fused feature information for failure mode identification.

[0107] The two-stream spatiotemporal attention deep neural network is a two-line parallel convolutional neural network coupled with a bidirectional gated recurrent unit, and adds a spatiotemporal attention mechanism. The spatial attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the convolutional neural network to enhance the region of interest in the spatial feature map. Its expression is:

[0108]

[0109] Where H, W, and C represent the height, width, and number of channels of the feature map, respectively. s and b s Denote the weight and bias of spatial attention respectively, X ijct represents the value of the feature map at position (i, j), channel c and time step t, S ijct Represents the spatial attention weight of this position.

[0110] The temporal attention mechanism acts on the second layer of the bidirectional gated recurrent unit to enhance the interval of interest in the temporal feature sequence. Its expression is:

[0111]

[0112] Where W t and b t Represent the weight and bias of temporal attention, Y ijct Represents the value of the feature map at position (i, j), channel c and time step t, T ijct Represents the temporal attention weight of the position.

[0113] Based on the extraction and fusion of multi-source features, the computing efficiency is improved and the trained model parameters are saved.

[0114] A5: Apply the least squares support vector machine to the feature-enhanced fusion data set for identification, use the crested porcupine algorithm to obtain the parameter combination of the matching data, input the feature fusion set obtained in A4 into the training, and obtain a parameter-adaptive failure mode identification model; A5 is specifically:

[0115] For the feature-enhanced fusion dataset, a least-squares support vector machine (LSSVM) is applied for failure mode identification, and the crested porcupine algorithm is used to determine the optimal hyperparameter combination for the LSSVM. First, the swarm is initialized using the crested porcupine algorithm to randomly generate LSSVM hyperparameter combinations. Through iterative optimization, the particle speed and position are continuously updated to ultimately find the optimal hyperparameter combination. The feature fusion set obtained by A4 is then input, and the LSSVM model is trained using the optimized hyperparameter combination, resulting in a parameter-adaptive failure mode identification model.

[0116] The data after feature extraction and fusion are input into the least squares support vector machine, and the hyperparameters are automatically adapted by the crested porcupine algorithm. The parameters of the crested porcupine algorithm are set as follows: the number of particle swarms is 100, the number of iterations is 100, the upper bound of the kernel parameter σ and the penalty parameter C is 500, and the lower bound is 0.0001.

[0117] A6: Multi-channel data collection is performed on the core components of the actual in-service equipment to be identified. Adaptive filtering is used to process multi-source uncertain noise, and feature fusion is used to enhance the characteristic information of the sparse data. This data is then input into the model generated in A5 to predict the failure mode category of the component. In step A6, if the failure mode category of the core component of the actual in-service equipment to be identified contains multi-source uncertain noise and the sample size is sparse, it is consistent with the actual failure category of the core component.

[0118] The hyperparameters obtained by optimizing in step A5 are used to define the least squares support vector machine and perform failure mode identification. The identification results are as follows: Figure 6 As shown, Figure 6 The model identification accuracy confusion matrix is ​​shown. Figure 6 It can be seen from the figure that the above method provided by the present application can well perform failure mode identification on sparse sample data containing multi-source uncertain noise, and the identification accuracy can reach 97.875%.

[0119] This embodiment proposes a method for identifying failure modes of equipment components for sparse samples containing uncertain noise, which is used to effectively solve the problem of failure identification of sparse data containing multi-source uncertain noise. Multi-channel working data of the entire life cycle is collected by accelerating failure testing and annotated according to known failure modes. For multi-source uncertain noise, the meta-detection algorithm is used for adaptive filtering preprocessing. Combining the short tree transform with the dual-stream spatiotemporal attention deep neural network, time-frequency and multi-scale features are obtained from multi-channel data, and parallel feature extraction and fusion are performed. The least squares support vector machine is applied for failure mode identification, and the crown porcupine algorithm is used to determine the optimal hyperparameter combination. Finally, the above-mentioned method after training and optimization is used to identify sparse multi-channel data of actual service equipment components, accurately obtain the failure mode category of the components, and is applicable to actual engineering.

[0120] Compared with the prior art, this application has the following beneficial effects:

[0121] 1) This application takes into account the fact that there are multi-source uncertain noises in the actual service process of the equipment, and applies the meta-detection algorithm for filtering processing to reduce the interference of noise signals on sparse failure data.

[0122] 2) This application targets the sparsity of failure samples and adopts a method combining short tree transformation with a dual-stream spatiotemporal attention deep neural network to enhance the ability to extract failure features and reduce the dependence of failure identification on data volume.

[0123] 3) This application performs failure mode identification based on the least squares support vector machine, adopts the crested porcupine algorithm to determine the optimal hyperparameter combination, and quickly converges to the optimal hyperparameter combination, thereby improving the failure identification capability of the model under complex working conditions in actual service.

[0124] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for equipment component failure mode identification based on multi-source uncertain sparse samples, characterized by: include: Acquire multi-source sensor data to be tested; The multi-source sensor data to be tested is multi-channel working data of equipment components with unclear failure modes; An adaptive active noise reduction filter is used to filter out multi-source uncertain noise in the multi-source sensor data to be measured, thereby obtaining filtered multi-source sensor data to be measured; the adaptive active noise reduction filter is a filter obtained by optimizing the filter coefficients of the adaptive filter through a model-independent meta-learning algorithm; the adaptive active noise reduction filter can adapt to different noise environments by adjusting and updating the filter parameters in real time; Inputting the filtered multi-source sensor data to be measured into a short tree transform algorithm to extract a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix; each element in the first image represents the signal strength at a specific time and frequency; each element in the second image represents the signal feature at a specific scale and direction; the short tree transform algorithm includes a short-time Fourier transform and a dual-tree complex wavelet transform; The first image and the second image are processed in parallel by a dual-stream spatiotemporal attention deep neural network, and time-frequency domain features and detail and direction features are extracted and fused respectively to obtain enhanced fusion features of failure mode data; the dual-stream spatiotemporal attention deep neural network includes a first branch and a second branch in parallel; the first branch is used to process the first image and extract time-frequency domain features; The second branch is used to process the second image and extract details and direction features; Inputting the enhanced fusion features of the failure mode data into a failure mode identification model to obtain corresponding failure mode identification results; The failure mode identification model is a least squares support vector machine whose hyperparameters have been optimized in advance using the crested porcupine algorithm; Before using an adaptive active noise reduction filter to filter out the multi-source uncertain noise in the multi-source sensor data to be measured to obtain the filtered multi-source sensor data to be measured, the equipment component failure mode identification method based on the multi-source uncertain sparse samples further includes: Finding an optimal initial control filter through model-independent meta-learning; in the model-independent meta-learning algorithm, optimizing hyperparameters in the model-independent meta-learning algorithm through a grid search algorithm; The filter coefficients of the optimal initial control filter are optimized to obtain an adaptive active noise reduction filter; the filter coefficients of the optimal initial control filter are optimized according to the following formula: ； in, For the The filter coefficients at step , Indicates the The filter coefficients at step , For the The input vector at step time, superscript T is the transpose operation, For the The error value between the expected output vector and the actual output vector at step , is the learning rate, is the time decay factor, is the historical step length; calculated according to the following formula : ； in, For the The expected output vector at step ; The multi-source uncertain noise in the multi-source sensor data to be measured is filtered out according to the following formula: ； in, is the multi-source sensor data to be measured, is the multi-source sensor data to be measured after filtering.

2. The equipment component failure mode identification method for multi-source uncertain sparse samples according to claim 1 is characterized in that: The hyperparameters in the model-agnostic meta-learning algorithm are optimized according to the following formula: ； in, is the optimal learning rate, is the optimal time decay factor, is the number of cross-validation folds, β is the learning rate ， is the time decay factor, is the loss function, Hyperparameter combination for model-independent meta-learning algorithms The performance below.

3. The equipment component failure mode identification method for multi-source uncertain sparse samples according to claim 1 is characterized in that: The filtered multi-source sensor data to be measured is input into a short tree transform algorithm to extract a first image including a time-frequency feature matrix and a second image including a multi-scale feature matrix, specifically including: Performing a short-time Fourier transform on the filtered multi-source sensor data to be measured to obtain a time-frequency feature matrix of the time series signal; each element in the time-frequency feature matrix represents the signal strength at a specific time and frequency; Performing dual-tree complex wavelet transform on the filtered multi-source sensor data to obtain a multi-scale feature matrix; the elements of each feature image represent the signal characteristics at a specific scale and in a specific direction; Arranging the time-frequency feature matrix according to the time and frequency axes to form a first image including the time-frequency feature matrix; The multi-scale feature matrix is ​​arranged according to the scale and direction axes to form a second image including the multi-scale feature matrix.

4. The equipment component failure mode identification method for multi-source uncertain sparse samples according to claim 1 is characterized in that: The structures of the first branch and the second branch are the same; the first branch includes a convolutional neural network, a bidirectional gated recurrent unit and a spatiotemporal attention mechanism; the first branch extracts spatial features through a convolutional neural network, extracts time-dependent features through a bidirectional gated recurrent unit, and introduces a spatiotemporal attention mechanism to extract spatiotemporal coupling features; the spatial attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the convolutional neural network to enhance the region of interest in the spatial feature map; the temporal attention mechanism of the spatiotemporal attention mechanism acts on the second layer of the bidirectional gated recurrent unit to enhance the region of interest in the temporal feature sequence.

5. The equipment component failure mode identification method for multi-source uncertain sparse samples according to claim 4 is characterized in that: The expression of the spatial attention mechanism is shown as follows: ； in, and Represent the height, width and number of channels of the feature image respectively, Indicates that the feature image is at position ,aisle and time steps The value at and denote the weight and bias of the spatial attention mechanism, respectively, is the spatial attention weight of the position; The expression of the temporal attention mechanism is as follows: ； in, and denote the weight and bias of the temporal attention mechanism, Indicates that the feature image is at position ,aisle and time steps The value at Represents the temporal attention weight of the position.

6. The equipment component failure mode identification method for multi-source uncertain sparse samples according to claim 1 is characterized in that: Before inputting the enhanced fusion features of the failure mode data into a failure mode identification model to obtain a corresponding failure mode identification result, the equipment component failure mode identification method based on multi-source uncertain sparse samples further includes: The crown porcupine algorithm is used to initialize the particle swarm, randomly generate the hyperparameter combination of the least squares support vector machine, and continuously update the speed and position of the particles through iterative optimization to finally find the optimal hyperparameter combination; Under the optimal hyperparameter combination, the enhanced fusion feature dataset is used to train the least squares support vector machine model to obtain a failure mode identification model; the enhanced fusion feature dataset includes several failure mode data enhanced fusion features and corresponding failure mode labels.

Images