Method and device for mechanical equipment fault diagnosis based on slow feature analysis

Abstract

The present invention relates to the technical field of mechanical equipment fault detection, and particularly to a mechanical equipment fault diagnosis method and device based on slow feature analysis. The method includes the following steps: using a time embedding method to reconstruct a one-dimensional vibration signal into a two-dimensional signal sequence; using slow feature analysis to extract a two-dimensional slow feature sequence from the two-dimensional signal sequence and intercepting the two-dimensional slow feature sequence with a square window to obtain a slow feature square matrix diagram; inputting the intercepted slow feature square matrix diagram data into a processing platform to perform feature extraction through a fault diagnosis model and obtaining a diagnosis result. By adding slow feature analysis, the present invention fully considers feature pre-extraction and the possibility of fusing high-dimensional features and low-dimensional features, significantly improving the effectiveness and robustness of the diagnosis model.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault detection, and particularly to a mechanical equipment fault diagnosis method and device based on slow feature analysis. Background Art

[0002] Driven by industrial big data and Industry 4.0, mechanical equipment is becoming more and more automated and intelligent. During the continuous operation of mechanical equipment, various fault situations will inevitably occur. Once these faults occur, they will lead to unplanned shutdown for maintenance, economic losses, and even catastrophic accidents. Therefore, it is necessary and of great significance to conduct fault diagnosis of mechanical equipment. However, in the actual industrial environment, mechanical equipment inevitably has to operate under different working conditions. Therefore, recent methods focus on using transfer learning methods to achieve fault diagnosis of mechanical equipment under different working conditions.

[0003] The goal of transfer learning is to reuse the diagnostic knowledge learned in one diagnostic task (source domain) to a related but different new task (target domain). In the field of mechanical equipment fault diagnosis, common transfer learning methods include parameter-based transfer, difference metric-based transfer, and adversarial learning-based transfer. These methods have achieved good results in fault diagnosis. Specifically, parameter-based transfer is to pre-train a diagnostic model on the source domain and then fine-tune the parameters of the diagnostic model on the target domain to make it adapt to the diagnostic task of the target domain. Difference metric-based transfer mainly maps the features extracted by the model on the source domain and the target domain to the same subspace and calculates their distribution differences, and minimizes this distribution difference during the training process to make the model adapt to the diagnostic task on the target domain. Adversarial learning-based transfer constructs a game method to learn transferable features. The model usually includes a feature extractor, a label classifier, and a domain discriminator. The domain discriminator distinguishes whether the feature is from the source domain or the target domain, while the feature extractor tries to deceive the discriminator to reduce the feature differences between the source domain and the target domain, so as to adapt to the diagnostic task on the target domain.

[0004] However, these methods do not fully consider the possibility of pre-extracting features and fusing high-dimensional features and low-dimensional features, and there are deficiencies in the effectiveness and robustness of the diagnostic model, which need to be further improved. In addition, in the context of industrial big data, the huge amount of data makes the model calculation and reasoning consume a large amount of time, which is not conducive to the application of the diagnostic system. Summary of the Invention

[0005] The purpose of the present invention is to solve at least one of the problems raised in the background art. To achieve the purpose of the present invention, the following technical solutions are adopted:

[0006] The first aspect of the present invention proposes a mechanical equipment fault diagnosis method based on slow feature analysis. The method includes the following steps:

[0007] Use the time embedding method to reconstruct the one-dimensional vibration signal into a two-dimensional signal sequence;

[0008] Use slow feature analysis to extract a two-dimensional slow feature sequence from the two-dimensional signal sequence, and intercept the two-dimensional slow feature sequence with a square window to obtain a slow feature square matrix diagram;

[0009] Input the intercepted slow feature square matrix diagram data into the processing platform, and after feature extraction through the fault diagnosis model, obtain the diagnosis result.

[0010] A further improvement lies in that the time embedding method is as follows: Assume that the one-dimensional time series signal is x i , and let the two-dimensional signal sequence X i =(x i , x i+1*s , x i+2*s , x i+3*s ,…), where s is the time translation step size.

[0011] A further improvement lies in that the steps of using slow feature analysis to extract a two-dimensional slow feature sequence from the two-dimensional signal sequence include:

[0012] Use the transformation matrix to map the two-dimensional signal sequence to obtain a two-dimensional slow feature sequence arranged according to slow speed, and its expression is:

[0013] Q = g(X) = WX (1)

[0014] In formula (1), Q represents the two-dimensional slow feature sequence, X represents the input two-dimensional signal sequence, g(*) is the SFA transformation function, and W is the linear SFA transformation matrix;

[0015] Specifically, solve the transformation matrix W by the method of solving the generalized eigenvalue decomposition:

[0016] AW = BWΛ (2)

[0017] In formula (2), is the first derivative of x, is 's covariance matrix, B = <xx T > represents the covariance matrix of the weighted key variable x, Λ is the diagonal matrix of the generalized eigenvalues, and W is a matrix of the generalized eigenvectors.

[0018] A further improvement lies in that the fault diagnosis model is a residual convolutional network model, and the residual convolutional network model includes a first convolutional layer and two residual convolutional layers. The input slow feature matrix diagram data first passes through the first convolutional layer and then respectively passes through the two residual convolutional layers for feature extraction.

[0019] A further improvement lies in that the fault diagnosis model further includes a bypass branch structure. The bypass branch structure is respectively connected to the input slow feature matrix diagram data and the output of the last residual convolutional layer to weight them together, obtain an aggregated output that aggregates low-dimensional features and high-dimensional features, and obtain a feature output through average pooling.

[0020] A further improvement lies in that the specific method for feature extraction includes:

[0021] The input slow feature matrix diagram data first passes through a convolutional layer, and the expression is:

[0022] C f1 =σ(W1X + b1) (3)

[0023] In formula (3), C f1 represents the convolutional output, σ(*) is a non-linear activation function, W1 is the convolutional kernel of the first convolutional layer, and b1 is the bias term of the first convolutional layer;

[0024] Then the output of the first convolutional layer respectively passes through two residual convolutional layers, and the expressions are as follows:

[0025]

[0026]

[0027] In formulas (4) and (5), C f2 is the output of the first layer of the residual convolutional layer, h(C f1 ) is the direct mapping part of the second layer of the residual convolutional layer, W2 is the convolutional kernel of the first layer of the residual convolutional layer, is the residual part of the first layer of the residual convolutional layer, is the output of the second layer of the residual convolution, is the direct mapping part of the second layer of the residual convolution, W3 is the convolutional kernel of the second layer of the residual convolutional layer, is the residual part of the second layer of the residual convolutional layer;

[0028] The working process of the bypass branch structure is as follows: The branch structure convolutional layer is used for dimension matching, and the convolutional output expression is as follows:

[0029] C f4 =σ(W4X + b4) (6)

[0030] In formula (6), Cf4 is the convolution output of the bypass branch structure, σ(*) is the non-linear activation function, W4 is the convolution kernel of the convolution layer of the branch structure, and b4 is the bias term of the convolution layer of the branch structure;

[0031] The expression for obtaining the total output of feature extraction is:

[0032]

[0033] In Equation (7), C represents the aggregated output of the high-dimensional and low-dimensional features, and α is the weighting coefficient of the weighted mapping;

[0034] Then, through mean pooling processing, the expression is:

[0035]

[0036] In Equation (8), is the output result of the c-th channel of mean pooling, N is the number of features in each channel, represents the i-th feature of the c-th channel.

[0037] A further improvement is that the processing platform is a supercomputer platform. On the supercomputer platform, the MPI parallel programming communication protocol is used to distribute the slow feature square matrix data to multiple computing nodes. Each computing node independently runs a fault diagnosis model, performs parallel computing and inference on the data assigned to it, and outputs the diagnosis result through the softmax function and the fully connected layer. The diagnosis results of each computing node are recovered to the master node to obtain the diagnosis results of all data.

[0038] A further improvement is that the expression for outputting the diagnosis result through the softmax function and the fully connected layer is:

[0039]

[0040] In Equation (9), k is the number of categories, θ i (1 ≤ i ≤ k) are the parameters of the classification layer, exp(*) represents the exponential function with base e, y i is the output result of the i-th classification layer, max(*) represents the maximum value function, and y out is the corresponding classification result of the input data.

[0041] A further improvement is that the specific method for recovering the diagnosis results of each computing node to the master node to obtain the diagnosis results of all data includes:

[0042] Aggregate communication is performed using the MPI interface. Through the Gather aggregation function, the inference results of each computing node are aggregated onto the master node. The diagnostic results of all data are recovered on the master node, and then through the TCP communication protocol of the Socket interface, the diagnostic results are sent to the local system over the network, where the local system receives and displays the diagnostic results.

[0043] A second aspect of the present invention proposes a mechanical equipment fault diagnosis device based on slow feature analysis. The device includes:

[0044] A recombination module for recombining one-dimensional vibration signals into two-dimensional signal sequences using time embedding.

[0045] A two-dimensional slow feature sequence extraction module for extracting two-dimensional slow feature sequences from the two-dimensional signal sequences using slow feature analysis and intercepting the two-dimensional slow feature sequences with a square window to obtain a slow feature square matrix diagram.

[0046] A model processing module for inputting the intercepted slow feature square matrix diagram data into a processing platform to perform feature extraction and diagnosis through a fault diagnosis model to obtain a diagnostic result.

[0047] A third aspect of the present invention proposes an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the above-mentioned mechanical equipment fault diagnosis methods based on slow feature analysis in the invention embodiments.

[0048] A fourth aspect of the present invention proposes a computer-readable storage medium. The computer-readable storage medium includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute any one of the above-mentioned mechanical equipment fault diagnosis methods based on slow feature analysis in the invention embodiments.

[0049] The beneficial effects of the present invention are:

[0050] By adding slow feature analysis, the present invention fully considers feature pre-extraction and the possibility of fusing high-dimensional features and low-dimensional features, significantly improving the effectiveness and robustness of the diagnostic model.

[0051] The present invention is processed through a supercomputer platform, distributing data to multiple computing nodes of the supercomputer platform. Each computing node independently runs a residual convolutional network, parallel computing and inferring the data assigned to it, with an obvious acceleration effect, which can significantly improve the data processing efficiency and cope with a large amount of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1Flow chart of a mechanical equipment fault diagnosis method based on slow feature analysis according to the present invention;

[0053] Figure 2 Schematic diagram of slow feature matrix diagram extraction in the present invention;

[0054] Figure 3 Schematic diagram of the convolutional residual network model in the present invention;

[0055] Figure 4 Visualization diagram of the feature distribution of the convolutional residual network model in the present invention on the target domain dataset;

[0056] Figure 5 Schematic diagram of the local system of the present invention receiving and displaying the diagnosis result;

[0057] Figure 6 Schematic diagram of the structure of a mechanical equipment fault diagnosis device based on slow feature analysis according to the present invention;

[0058] Figure 7 Schematic diagram of an electronic device according to the present invention. Detailed implementation manners

[0059] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0061] Please refer to the atta Figure 1 - atta Figure 7 ched drawings, a mechanical equipment fault diagnosis method based on slow feature analysis is proposed in the first aspect of the embodiments of the present invention. As Figure 1 shown, the method includes the following steps:

[0062] Step S1: Use the time embedding method to reconstruct the one-dimensional vibration signal into a two-dimensional signal sequence.

[0063] Step S2: Use slow feature analysis to extract a two-dimensional slow feature sequence from the two-dimensional signal sequence.

[0064] Step S3: Use a square window to intercept the two-dimensional slow feature sequence to obtain a slow feature square matrix diagram.

[0065] Step S4: Input the data of the intercepted slow feature square matrix diagram into the processing platform, and after feature extraction through the fault diagnosis model, obtain the diagnosis result.

[0066] Specifically, in this embodiment, the time embedding method is as follows: Assume that the one-dimensional time series signal is x i , and let the two-dimensional signal sequence X i =(x i , x i+1*S , x i+2*S , x i+3*S ,…), where s is the time translation step size. By reconstructing the one-dimensional time series signal into a two-dimensional signal sequence, it is convenient for slow feature analysis.

[0067] Using the slow feature analysis technique (SFA), from the rapidly changing input signal, extract the slowly changing output signal, that is, in the present invention, use slow feature analysis to extract a two-dimensional slow feature sequence from the input two-dimensional signal sequence.

[0068] Linear SFA is used to obtain a set of slow features arranged from slow to fast. Specifically, the steps of using slow feature analysis to extract a two-dimensional slow feature sequence from the two-dimensional signal sequence include:

[0069] Use the transformation matrix to map the two-dimensional signal sequence to obtain a two-dimensional slow feature sequence arranged according to the slow speed, and its expression is:

[0070] Q = g(X) = WX (1)

[0071] In formula (1), Q represents the two-dimensional slow feature sequence, X represents the input two-dimensional signal sequence, g(*) is the SFA transformation function, and W is the linear SFA transformation matrix.

[0072] The key to solving the SFA problem is to solve the transformation matrix W. Specifically, the transformation matrix W can be solved by the method of solving the generalized eigenvalue decomposition (GED):

[0073] AW = BWΛ (2)

[0074] In formula (2), is the first derivative of x, is The covariance matrix, B = <xx T > represents the covariance matrix of the weighted key variable x, Λ is a diagonal matrix of generalized eigenvalues, and W is a matrix of generalized eigenvectors.

[0075] Then, the two-dimensional signal sequence is mapped using the transformation matrix W to obtain a two-dimensional slow feature sequence arranged according to the slow speed. A square window is used to intercept the two-dimensional slow feature sequence, and a slow feature square matrix diagram can be obtained.

[0076] In this embodiment, the fault diagnosis model is a residual convolutional network model. The residual convolutional network model includes a first convolutional layer and two residual convolutional layers. The input slow feature square matrix diagram data first passes through the first convolutional layer and then respectively passes through the two residual convolutional layers for feature extraction. The model uses an Adam optimizer with a learning rate of 0.001 and a batch size of 50.

[0077] The fault diagnosis model further includes a bypass branch structure. The bypass branch structure is respectively connected to the input slow feature square matrix diagram data and the output of the last residual convolutional layer to weight them together to obtain an aggregated output of aggregated low-dimensional features and high-dimensional features, and the feature output is obtained through average pooling.

[0078] As Figure 3 is a schematic diagram of the convolutional residual network model in the present invention. The specific method for feature extraction by the residual convolutional network model in the present invention includes:

[0079] The input slow feature square matrix diagram data first passes through the convolutional layer, and the expression is:

[0080] C f1 = σ(W1X + b1) (3)

[0081] In formula (3), C f1 represents the convolutional output, σ(*) is a non-linear activation function, W1 is the convolutional kernel of the first convolutional layer, and b1 is the bias term of the first convolutional layer.

[0082] After passing through the first convolutional layer, the convolutional output is sent to the residual block. Since the number of feature maps of the input x L and the output x L+1 of the residual convolution is different, a 1X1 convolution is required to reduce the dimension of the direct mapping part, and the direct mapping is expressed as h(x L ).

[0083] Therefore, the output of the first convolutional layer passes through the two residual convolutional layers respectively, and the expressions are as follows:

[0084]

[0085]

[0086] In equations (4) and (5), C f2 is the output of the first-layer residual convolutional layer, h(C f1 ) is the direct mapping part of the second-layer residual convolutional layer, W2 is the convolutional kernel of the first-layer residual convolutional layer, is the residual part of the first-layer residual convolutional layer, is the output of the second-layer residual convolution, is the direct mapping part of the second-layer residual convolution, W3 is the convolutional kernel of the second-layer residual convolutional layer, is the residual part of the second-layer residual convolutional layer.

[0087] Meanwhile, a bypass branch structure mainly used for feature extraction is added to the network to solve the problem that the deep abstraction of high-dimensional features will lose more low-dimensional information closely related to the original signal.

[0088] Specifically, the working process of the bypass branch structure is as follows: The branch structure convolutional layer is used for dimension matching, and the convolutional output expression is as follows:

[0089] C f4 = σ(W4X + b4) (6)

[0090] In equation (6), Cf4 is the convolutional output of the bypass branch structure, σ(*) is the non-linear activation function, W4 is the convolutional kernel of the branch structure convolutional layer, and b4 is the bias term of the branch structure convolutional layer.

[0091] The expression for obtaining the total output of feature extraction is:

[0092]

[0093] In equation (7), C represents the aggregated output of high-dimensional and low-dimensional features, and α is the weighting coefficient of the weighted mapping.

[0094] Then, mean pooling processing is performed, and the expression is:

[0095]

[0096] In equation (8), is the output result of the c-th channel of mean pooling, N is the number of features in each channel, represents the i-th feature of the c-th channel.

[0097] As a preferred solution of an embodiment of the present invention, the processing platform is a supercomputer platform. On the supercomputer platform, the MPI parallel programming communication protocol is used to distribute the slow feature matrix diagram data to multiple computing nodes. Specifically, collective communication is carried out using the MPI interface, and the data is evenly distributed to multiple computing nodes through the Scatter function. Each computing node independently runs a fault diagnosis model, performs parallel computing and inference on the data assigned to it, outputs the diagnosis result through the softmax function and the fully connected layer, and the diagnosis results of each computing node are recovered to the master node to obtain the diagnosis results of all data.

[0098] Specifically, the expression for outputting the diagnosis result through the softmax function and the fully connected layer is:

[0099]

[0100] In Equation (9), k is the number of categories, θ i (1 ≤ i ≤ k) are the parameters of the classification layer, exp(*) represents the exponential function with base e, y i is the output result of the i-th classification layer, max(*) represents the maximum value function, and y out is the corresponding classification result of the input data.

[0101] Specifically, the specific method for recovering the diagnosis results of each computing node to the master node to obtain the diagnosis results of all data includes:

[0102] Use the MPI interface for collective communication. Through the Gather function, the inference results of each computing node are gathered to the master node. The diagnosis results of all data are recovered on the master node, and then through the TCP communication protocol of the Socket interface, the diagnosis results are sent to the local system through the network, and the local system receives and displays the diagnosis results. Figure 5 This is a schematic diagram of the local system of the present invention receiving and displaying the diagnosis results.

[0103] Experimental demonstration is carried out for the method proposed in the embodiment of the present invention, as follows:

[0104] Experimental dataset collection:

[0105] In this embodiment, the bearing fault dataset open-sourced by Case Western Reserve University (CWRU) is selected to test and verify the proposed residual convolutional network model. Rolling element bearing fault is a common fault in rotating machinery, and the CWRU bearing dataset is a standard dataset for verifying the quality of a diagnosis model in bearing datasets.

[0106] The CWRU bearing fault dataset contains the vibration acceleration signals of the bearing seats at the motor fan end and the drive end when the experimental platform is working under normal conditions and faulty bearing conditions. The experimental platform consists of a motor, a torque sensor, a dynamometer, and a control part. By means of single-point electrical discharge machining, three types of faults were introduced on the bearing: outer race fault (OF), inner race fault (IF), and rolling ball fault (BF). Each fault type has four different damage diameters.

[0107] In the experiment of this embodiment, the fault data from the drive end (DE) is used, and three fault types, namely ball fault, inner race fault, and outer race fault, are selected. Each fault type has damage diameters of 0.007 in, 0.014 in, and 0.021 in respectively. Therefore, the operating states include one normal state and nine fault states. Each operating state includes vibration acceleration signals collected at a sampling frequency of 12 kHz under four motor load conditions of 0 - 3 HP.

[0108] Next, preprocess the experimental data:

[0109] Select the dataset obtained under the 0 HP motor load working condition as the source domain dataset S0, and the datasets under the other three load conditions as the target domain datasets T1, T2, and T3 respectively. For the source domain dataset, after the vibration signals in each operating state are subjected to slow feature analysis, the data splitting technique is used to obtain 4000 training samples and 1000 test samples. Therefore, a total of 40000 training samples and 10000 test samples are obtained for the ten operating states. For the target domain datasets, similarly, the data splitting technique is used to obtain 4000 test samples for each healthy state, and a total of 40000 test samples are obtained for the ten operating states.

[0110] This invention uses the pytorch library to build a network structure model. The main parameter settings of the model are described as follows: The residual convolutional network model is set to 3 layers, including the first convolutional layer and two residual convolutional layers. At the same time, there is also a branch structure that aggregates high-dimensional features and low-dimensional features. For details, please refer to the previous text of this embodiment. The model uses the Adam optimizer, the learning rate is 0.001, and the batch size is 50. We will use the classification accuracy of the test results as the evaluation index of the model.

[0111] To evaluate the performance of the model of this invention, this invention is compared with a series of state-of-the-art baseline models, including:

[0112] TSDL (Wang et al., 2021): The purpose of the first stage is to generate different types of fault data in the source domain, and then train a cross-domain classifier in the second stage.

[0113] DFA (Wang et al., 2021): Adopt a latent feature alignment method guided by Gaussian prior to maximize the difference of the classifier.

[0114] EFACNN (Tang et al., 2021): This model applies an enhanced attention feature (EAF) module and a two-branch CNN architecture with shared feature weights, uses MMD with an edge probability distribution alignment function for cross-domain feature alignment, and regularizes the weights using the L2 norm.

[0115] FCDNN (Huang and Carley, 2019): This model includes modules for feature learning and metric learning. The feature learning module extracts features from sample pairs. The metric learning module predicts the similarity of sample pairs. The classification task is completed using the similarity of sample pairs that combine the test sample with each labeled sample.

[0116] MJDCNN (Liang et al., 2022): An end-to-end joint adaptive transfer learning framework based on multi-layer feature fusion. The baseline model is pre-trained on the source domain, and then the pre-trained weight initialization model is transferred to related but different tasks through a joint adaptive fine-tuning strategy.

[0117] The experimental results are shown in Table 1 below:

[0118] Methods S0→T1 S0→T2 S0→T3 Average TSDL 97.81 96.02 94..24 96.02 DFA 99.81 99.61 89.23 96.22 EFACNN 95.4 86.4 84.7 88.83 FCDNN 99.61 99.03 96.80 98.48 MJDCNN 98.38 99.04 99.15 98.86 SFRCN(ours) 99.45 99.31 98.95 99.24

[0119] Table 1

[0120] As can be seen from Table 1, the method proposed in the present invention (SFRCN (ours)) is superior to other methods. In the three transfer diagnosis tasks, the present invention has good diagnostic performance, is superior to most algorithms, and has the highest average diagnostic accuracy. Among these methods, the method of the present invention is the only method with an average accuracy rate higher than 99%, indicating that the method of the present invention has good diagnostic performance and robustness and can achieve accurate diagnosis under different working conditions.

[0121] Compared with MJDCNN, the method of the present invention can be directly extended to new fields after training without the need for fine-tuning on the target domain, which is more convenient. Compared with TSDL, DFA, EFACNN, and FCDNN, by introducing manifold feature analysis in the present invention, the network structure of our model is more concise, but still maintains excellent diagnostic performance and robustness, and reduces the parameters and the time for network training.

[0122] Feature distribution visualization is used to reveal the changes in the feature distribution during the training process. By Figure 4It can be clearly seen that (a)-(f) are the changes in the feature distribution during the training process of the present invention. At the initial stage of training, the features of different categories are mixed with each other and are difficult to distinguish. However, as the training progresses, the features of different categories tend to be distributed dispersedly, and the features of the same category tend to gather together. Finally, after the training is completed, a certain distance is maintained between the features of different categories, while the same features are clustered together. The feature distribution has good separability, and the extracted features are discriminative and are easily classified by the classifier. Therefore, the method of the present invention has good feature mining and feature expression capabilities.

[0123] Parallel acceleration effect using the Tianhe-2 supercomputer platform:

[0124] Using the CWRU bearing fault data set, the data volume of the test set is expanded to 100,000 data. A 64-bit computer with a GTX960M GPU, the main node of Tianhe-2, and 4 computing nodes of Tianhe-2 are respectively used to run the network and calculate and infer these 100,000 data. On a 64-bit computer with a GTX960M GPU, it takes 406 s for the network model to diagnose these 100,000 data. When diagnosing on the main node of Tianhe-2, it takes 98 s, and the speedup ratio is 4.14. When performing parallel diagnosis on 4 computing nodes of Tianhe-2, it takes 35 s, and the speedup ratio is 11.6. Thus, it can be seen that using the Tianhe-2 supercomputer platform for parallel fault diagnosis in the present invention has an obvious acceleration effect, can significantly improve the data processing efficiency, and cope with a large amount of data.

[0125] A second aspect of the embodiment of the present invention proposes a mechanical equipment fault diagnosis device based on slow feature analysis. Refer to Figure 6 As shown, it is a schematic structural diagram of a mechanical equipment fault diagnosis device based on slow feature analysis provided corresponding to an embodiment of the present invention. Corresponding to the method for diagnosing mechanical equipment faults based on slow feature analysis provided in the above embodiment of the present invention, since the mechanical equipment fault diagnosis device based on slow feature analysis provided in the embodiment of the present invention corresponds to the method for diagnosing mechanical equipment faults based on slow feature analysis provided in the above embodiment of the present invention, the implementation manner of the foregoing method for diagnosing mechanical equipment faults based on slow feature analysis is also applicable to the mechanical equipment fault diagnosis device based on slow feature analysis provided in this embodiment.

[0126] Specifically, the mechanical equipment fault diagnosis device based on slow feature analysis includes:

[0127] A recombination module 10, configured to recombine a one-dimensional vibration signal into a two-dimensional signal sequence by using a time embedding method.

[0128] The two-dimensional slow feature sequence extraction module 20 enables the user to extract a two-dimensional slow feature sequence from a two-dimensional signal sequence using slow feature analysis and to obtain a slow feature square matrix diagram by intercepting the two-dimensional slow feature sequence with a square window.

[0129] The model processing module 30 is used to input the data of the intercepted slow feature square matrix diagram into a processing platform to perform feature extraction and diagnosis through a fault diagnosis model to obtain a diagnosis result.

[0130] The display module 40 is used to display the diagnosis result.

[0131] In addition, the specific functions of the modules of the mechanical equipment fault diagnosis device based on slow feature analysis may be referred to the above-mentioned mechanical equipment fault diagnosis method based on slow feature analysis, and will not be repeated in this embodiment.

[0132] See Figure 7 , and the embodiment of the present invention also correspondingly provides an electronic device and a computer-readable storage medium.

[0133] As Figure 7 shown is a schematic diagram of an electronic device provided by an embodiment of the present invention. The electronic device of this embodiment includes: a processor 11, a memory 12, and a computer program stored in the memory and executable on the processor 11. When the processor 11 executes the computer program, it implements the steps in the embodiment of the above-mentioned mechanical equipment fault diagnosis method based on slow feature analysis. Alternatively, when the processor 11 executes the computer program, it implements the functions of each module / unit in the above-mentioned device embodiments.

[0134] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor 11 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0135] The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the electronic device, and does not constitute a limitation on the electronic device. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0136] The so-called processor 11 may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the electronic device and connects various parts of the entire electronic device using various interfaces and lines.

[0137] The memory 12 can be used to store the computer programs and / or modules. The processor realizes various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by invoking the data stored in the memory. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system 121, application programs 122 required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0138] Among them, if the modules / units integrated in the electronic device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0139] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0140] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0141] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included within the protection scope of the present application.

Claims

1. A mechanical equipment fault diagnosis method based on slow feature analysis, characterized in that, It includes the following steps: Reconstruct the one-dimensional vibration signal into a two-dimensional signal sequence by using the time embedding method; Extract a two-dimensional slow feature sequence from the two-dimensional signal sequence by using slow feature analysis and intercept the two-dimensional slow feature sequence with a square window to obtain a slow feature square matrix diagram. Extracting the two-dimensional slow feature sequence from the two-dimensional signal sequence by using slow feature analysis includes: Map the two-dimensional signal sequence by using a transformation matrix to obtain a two-dimensional slow feature sequence arranged according to slow speed, and its expression is: Q = g(X) = WX (1) In formula (1), Q represents the two-dimensional slow feature sequence, X represents the input two-dimensional signal sequence, g(*) is the SFA transformation function, and W is the linear SFA transformation matrix; Specifically, solve the transformation matrix W by solving the generalized eigenvalue decomposition method: AW = BWΛ (2) In formula (2), is the first-order differential of x, is 's covariance matrix, B = xx T > represents the covariance matrix of the weighted key variable x, Λ is a diagonal matrix of generalized eigenvalues, and W is a matrix of generalized eigenvectors; Input the data of the intercepted slow feature square matrix diagram into the processing platform, perform feature extraction through a fault diagnosis model, and obtain a diagnosis result. The fault diagnosis model is a residual convolutional network model. The residual convolutional network model includes a first convolutional layer and two residual convolutional layers. The input data of the slow feature square matrix diagram first passes through the first convolutional layer and then passes through the two residual convolutional layers respectively for feature extraction. The fault diagnosis model also includes a bypass branch structure. The bypass branch structure is respectively connected to the input data of the slow feature square matrix diagram and the output of the last residual convolutional layer to weight them together to obtain an aggregated output of aggregated low-dimensional features and high-dimensional features, and obtain a feature output through average pooling.

2. The mechanical equipment fault diagnosis method based on slow feature analysis according to claim 1, characterized in that The time embedding method is as follows: Assume that the one-dimensional time series signal is x i , and let the two-dimensional signal sequence X i =(x i , x i+1*s , x i+2*s , x i+3*s ,...), where s is the time translation step size.

3. A mechanical equipment fault diagnosis method based on slow feature analysis according to claim 1, characterized in that The specific method for feature extraction includes: The input data of the slow feature square matrix diagram first passes through a convolutional layer, and the expression is: C f1 = σ(W1X + b1) (3) In formula (3), C f1 represents the convolution output, σ(*) is a non-linear activation function, W1 is the convolution kernel of the first convolutional layer, and b1 is the bias term of the first convolutional layer; Then the output of the first convolutional layer passes through the two residual convolutional layers respectively, and the expressions are as follows: In Formula (4) and Formula (5), C f2 is the output of the first-layer residual convolutional layer, h(C f1 ) is the direct mapping part of the second-layer residual convolutional layer, W2 is the convolutional kernel of the first-layer residual convolutional layer, is the residual part of the first-layer residual convolutional layer, C f3 is the output of the second-layer residual convolution, h(C f2 ) is the direct mapping part of the second-layer residual convolution, W3 is the convolutional kernel of the second-layer residual convolutional layer, is the residual part of the second-layer residual convolutional layer; The working process of the bypass branch structure is as follows: Use a branch structure convolutional layer for dimension matching, and the convolutional output expression is as follows: C f4 = σ(W4X + b4) (6) In formula (6), C f4 is the convolution output of the bypass branch structure, σ(*) is the non-linear activation function, W4 is the convolution kernel of the convolution layer of the branch structure, and b4 is the bias term of the convolution layer of the branch structure; The expression for obtaining the total output of feature extraction is: C=C f3 +α*C f4 (7) In formula (7), C represents the aggregated output of high-dimensional features and low-dimensional features, and α is the weighting coefficient of the weighted mapping; Then perform average pooling processing, and the expression is: In Equation (8), is the output result of the c-th channel of average pooling, N is the number of features in each channel, represents the i-th feature of the c-th channel.

4. A mechanical equipment fault diagnosis method based on slow feature analysis according to claim 1, characterized in that, The processing platform is a supercomputer platform. Use the MPI parallel programming communication protocol on the supercomputer platform to distribute the data of the slow feature square matrix diagram to multiple computing nodes. Each computing node independently runs a fault diagnosis model, performs parallel computing and inference on the data assigned to it, and outputs the diagnosis result through the softmax function and the fully connected layer. Recover the diagnosis results of each computing node to the master node to obtain the diagnosis results of all data.

5. The mechanical equipment fault diagnosis method based on slow feature analysis according to claim 4, wherein The expression for outputting the diagnosis result through the softmax function and the fully connected layer is: In Equation (9), k is the number of categories, and θ i (1 ≤ i ≤ k) are the parameters of the classification layer, exp(*) represents the exponential function with base e, and y i is the output result of the i-th classification layer, max(*) represents the maximum value function, and y out is the corresponding classification result of the input data.

6. A fault diagnosis method for mechanical equipment based on slow feature analysis according to claim 4, characterized in that, The specific method for recovering the diagnosis results of each computing node to the master node to obtain the diagnosis results of all data includes: Use the MPI interface for aggregated communication. Through the Gather aggregation function, aggregate the inference results of each computing node to the master node. Recover the diagnosis results of all data on the master node, and then through the TCP communication protocol of the Socket interface, send the diagnosis result to the local system through the network, and the local system receives and displays the diagnosis result.

7. A mechanical equipment fault diagnosis device based on slow feature analysis, which is used to execute a mechanical equipment fault diagnosis method based on slow feature analysis according to any one of claims 1-6, characterized in that, The device includes: A recombination module for recombining one-dimensional vibration signals into a two-dimensional signal sequence using time embedding; A two-dimensional slow feature sequence extraction module for extracting a two-dimensional slow feature sequence from the two-dimensional signal sequence using slow feature analysis and intercepting the two-dimensional slow feature sequence with a square window to obtain a slow feature square matrix diagram; A model processing module for inputting the intercepted slow feature square matrix diagram data into a processing platform to perform feature extraction and diagnosis through a fault diagnosis model to obtain a diagnosis result.

Images