Fault Diagnosis Method for Rotating Machinery Based on Multi-Sensor Correlation Feature Fusion

Abstract

The present invention proposes a fault diagnosis method for rotating machinery based on multi-sensor related feature fusion, and the implementation steps are as follows: obtaining a training sample set and a test sample set; building a fault diagnosis model based on a multi-sensor related feature fusion network; initializing parameters; training the fault diagnosis model; obtaining the trained fault diagnosis model; obtaining the fault diagnosis target classification result. In the process of training the fault diagnosis model and obtaining the diagnosis result in the present invention, since the multi-sensor related feature fusion network can fuse the intra-sensor label correlation matrix and the inter-sensor label correlation matrix constructed from the output results of the multi-sensor feature extraction network, the complementarity of fault information between different sensors is realized, and the accuracy of fault diagnosis is effectively improved. By fusing the fault information in two modalities of the time domain and the time-frequency domain, the defects of the prior art are avoided, and the accuracy of fault diagnosis is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and relates to a fault diagnosis method for rotating mechanical equipment, in particular to a fault diagnosis method for rotating mechanical equipment based on multi-sensor correlation feature fusion. Background Art

[0002] In modern production, the complexity of various large-scale mechanical equipment is increasing, and the requirements for fault diagnosis of rotating mechanical equipment such as bearings, gears, and rotors affected by potential process abnormalities are also increasing. Early fault detection can not only eliminate faults before they occur, avoid causing considerable economic losses, but also prevent the occurrence of major safety accidents.

[0003] The advent of the industrial big data era has brought new opportunities for the health monitoring of mechanical equipment. More and more machine learning-based methods are applied to the field of rotating machinery fault diagnosis. Signal analysis methods are usually adopted, and their technical idea is to analyze the time domain or frequency domain of sensor signals, extract the fault information hidden in the data, and then perform fault classification. However, the research in this area mainly focuses on solving the fault diagnosis problem of single-sensor data, and still has certain limitations in dealing with multi-sensor high-dimensional data.

[0004] Generally, the operating conditions of rotating mechanical equipment are monitored by multiple types and quantities of sensors, which can provide more comprehensive information than single sensors. To address the above problems, the article "Intelligent Mechanical Fault Diagnosis Using Multisensor Fusion and Convolution Neural Network" published by Xie et al. in the journal "IEEE Transactions on Industrial Informatics" in 2021 proposed a multi-sensor fault diagnosis network. However, this method simply reduces the dimension and splices and fuses the multi-sensor time domain data directly, ignoring the complementary information between different sensors. Therefore, how to effectively extract and fuse the fault information obtained by multi-sensors, improve the fault diagnosis accuracy, and comprehensively evaluate the health status of rotating mechanical equipment is of great significance. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects existing in the above-mentioned prior art, and provide a fault diagnosis method for rotating mechanical equipment based on multi-sensor correlation feature fusion, which is used to solve the technical problem of low fault diagnosis accuracy existing in the prior art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain the training sample set and the test sample set:

[0008] Obtain multiple pieces of original signal data containing C fault categories collected by M sensors and their corresponding fault labels, as well as multiple pieces of real-time signal data of the rotating mechanical equipment to be diagnosed, and preprocess each piece of original signal data and each piece of real-time signal data. Then, form the training sample set X with the I pieces of preprocessed original signal data and their corresponding fault labels train , and form the test sample set X with the J pieces of preprocessed real-time signal data test , where M ≥ 2 and C ≥ 2;

[0009] (2) Build a fault diagnosis model based on a multi-sensor related feature fusion network:

[0010] Build a fault diagnosis model O including a multi-sensor feature extraction network, a multi-sensor related feature fusion network, and a classifier C connected in series f , where the multi-sensor feature extraction network includes M network branches arranged in parallel. The m-th network branch includes a global-local time encoder and a time-frequency encoder arranged in parallel, and a fusion module and a classifier C cascaded in sequence at the output ends of the two encoders m ; The multi-sensor related feature fusion network is used to fuse the intra-sensor label correlation matrix and the inter-sensor label correlation matrix constructed from the output results of the multi-sensor feature extraction network

[0011] (3) Initialize the parameters:

[0012] Initialize the training iteration number as t, the maximum iteration number as T. The parameters of the multi-sensor feature extraction network and the classifier C f in the t-th iteration are φ t and μ t , respectively, and let t = 0;

[0013] (4) Train the fault diagnosis model:

[0014] Use the training sample set X train as the input of the fault diagnosis model O for forward propagation to obtain I prediction probabilities P f ;

[0015] (5) Obtain the trained fault diagnosis model:

[0016] Update the parameters φ t of the feature extraction network and the parameters μ f of the classifier C t to obtain the fault diagnosis model O of this iteration t, and determine whether t≥T holds. If so, obtain the trained fault diagnosis model O * , otherwise, let t = t + 1, O t = O, and execute step (4);

[0017] (6) Obtain the fault diagnosis target classification result:

[0018] Take the test sample set X test as the input of the trained fault diagnosis model O * to perform forward propagation, and obtain J classification results.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] (1) In the process of training the fault diagnosis model and obtaining the diagnosis result in the present invention, since the multi-sensor related feature fusion network can fuse the intra-sensor label correlation matrix and the inter-sensor label correlation matrix constructed from the output results of the multi-sensor feature extraction network, the complementarity of fault information between different sensors is realized, and the accuracy of fault diagnosis is effectively improved.

[0021] (2) Each network branch in the multi-sensor feature extraction network of the present invention can obtain fault information in two modalities of time domain and time-frequency domain through the globally-local time encoder and time-frequency encoder arranged in parallel, avoiding the defect that the prior art does not consider the fault information in the time-frequency domain modality, and further improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is the implementation flowchart of the present invention;

[0023] Figure 2 is the structural schematic diagram of the fault diagnosis model of the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0024] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0025] Referring to Figure 1 , the present invention includes the following steps:

[0026] Step 1) Obtain the training sample set and the test sample set.

[0027] Obtain multiple pieces of original signal data including C fault categories collected by M sensors and their corresponding fault labels, as well as multiple pieces of real-time signal data of the rotating mechanical equipment to be diagnosed, and preprocess each piece of original signal data and each piece of real-time signal data. Then, form the training sample set X from the I pieces of preprocessed original signal data and their corresponding fault labels train, the preprocessed J real-time signal data are used to form the test sample set X test . In this embodiment, the data uses the CWRU public dataset, which includes accelerometer sensors at three different positions: the drive end, the fan end, and the base. It contains 9 fault conditions of inner race fault, outer race fault, and rolling element fault with 3 damage degrees. Specifically, M = 3, C = 9, I = 90, J = 900;

[0028] The implementation steps for preprocessing each piece of original signal data and each piece of real-time signal data are as follows:

[0029] (1a) Each piece of original signal data and each piece of real-time signal data are segmented into data segments of length l by means of sliding window sampling, and each is normalized to obtain I preprocessed original signal data segments and J preprocessed real-time signal data segments. In this embodiment, l = 2048, and the normalization method uses min-max normalization, and the calculation is as follows:

[0030]

[0031] (1b) The I preprocessed original signal data segments and the fault labels corresponding to the multiple pieces of original signal data are combined to form the training sample set And the J preprocessed real-time signal data segments are used as the test sample set Among them, x i represents the i-th preprocessed original signal data segment containing M sensor data, y i represents the fault label corresponding to x i , and x j represents the j-th test sample containing M sensor data,

[0032] Step 2) Build a fault diagnosis model based on a multi-sensor related feature fusion network.

[0033] Build a fault diagnosis model O as Figure 2 shown, which is used to better extract and fuse the complementary information between different sensors and further improve the fault diagnosis accuracy. It includes a multi-sensor feature extraction network, a multi-sensor related feature fusion network, and a classifier C in cascade f . Among them, the multi-sensor feature extraction network includes M parallel network branches. The m-th network branch includes a global-local time encoder and a time-frequency encoder arranged in parallel, and a fusion module and a classification module C cascaded in sequence at the output ends of the two encoders m ; The multi-sensor related feature fusion network is used to fuse the within-sensor label correlation matrix and the between-sensor label correlation matrix constructed from the output results of the multi-sensor feature extraction network;

[0034] The global-local time encoder includes a globally arranged global encoder and a local encoder arranged in parallel, and a feature fusion module cascaded at the output ends of the global encoder and the local encoder. The global encoder includes a cascaded S-layer LSTM structure and a fully connected layer. The local encoder includes a cascaded T convolutional structures and a fully connected layer. The convolutional structure includes stacked one-dimensional convolutional layers, non-linear activation layers, and max pooling layers. In this embodiment, S = 4, T = 2. The convolutional kernel sizes of the first and second layers of the local encoder are 128×1 and 5×1 respectively. The non-linear activation layer uses the ReLU activation function;

[0035] The time-frequency encoder includes a cascaded K convolutional structures and a fully connected layer. The convolutional structure includes stacked two-dimensional convolutional layers, non-linear activation layers, and max pooling layers. In this embodiment, K = 2. The convolutional kernel sizes of the two layers of the time-frequency encoder are both 3×3. The non-linear activation layer uses the ReLU activation function;

[0036] In this embodiment, the output feature dimensions of the fully connected layers of the global encoder, the local encoder, and the time-frequency encoder are all 256;

[0037] Step 3) Initialize the parameters.

[0038] Initialize the number of training iterations as t, the maximum number of iterations as T, and the parameters of the multi-sensor feature extraction network and the classifier C f at the t-th iteration are φ t and μ t respectively, and let t = 0. In this embodiment, T = 50;

[0039] Step 4) Train the fault diagnosis model.

[0040] Use the training sample set X train as the input of the fault diagnosis model O for forward propagation to obtain I predicted probabilities P f . The specific implementation steps are as follows:

[0041] (4a) The multi-sensor feature extraction network respectively extracts features and fuses features for each training sample to obtain M×I initial predicted probabilities P = {P 1 , P 2 ,..., P m ,..., P M};

[0042] (4a1) The global-local time encoder extracts features for each training sample to obtain I time-domain features. Specifically, it includes the global time-domain features extracted by the global encoder and the local time-domain features extracted by the local encoder. Then, through the feature fusion module, a concat operation is performed on each global time-domain feature and its corresponding local time-domain feature to achieve feature fusion, resulting in time-domain features;

[0043] (4a2) The time-frequency encoder extracts features for each training sample after wavelet transform to obtain I time-frequency domain features. Among them, the calculation method of wavelet transform is as follows:

[0044]

[0045] In this embodiment, the wavelet transform adopts the Morlet wavelet basis function, α = 256, β = 1;

[0046] (4a3) The fusion module performs concat on each time-domain feature and its corresponding time-frequency domain feature to achieve feature fusion; the classification module classifies the result of feature fusion to obtain I initial prediction probabilities P m ; the M×I initial prediction probabilities P output by the sensor feature extraction network are P = {P 1 , P 2 ,..., P m ,..., P M};

[0047] (4b) The multi-sensor correlation feature fusion network constructs the in-sensor label correlation matrix r intra and the inter-sensor label correlation matrix r inter for the M×I initial prediction probabilities P, and fuses these two matrices through a convolutional layer to obtain the fused label matrix r;

[0048] (4b1) The in-sensor and inter-sensor label correlation matrices are respectively constructed for the M initial prediction probabilities. The in-sensor label correlation matrix can be obtained from the following formula:

[0049]

[0050] r intra = [R 1,1 , R 2,2 ,..., R m,m ,..., R M,M

[0051] Among them, P m represents the prediction label matrix of the m-th sensor, is the corresponding transposed matrix, and R m,m ​Denote the intra-sensor label correlation matrix of the m-th sensor, r intra Denote the set of intra-sensor label correlation matrices of M sensors.

[0052] Similarly, the inter-sensor label correlation matrix for each pair of sensors can be expressed as follows:

[0053]

[0054] r inter = [R 1,2 , R 1,3 ,..., R u,w ,..., R M-1,M

[0055] where, P u denotes the predicted label matrix of the u-th sensor, is the transpose matrix of P w , R u,w denotes the inter-sensor label correlation matrix obtained by multiplying the transpose of the predicted label of the u-th sensor and the predicted label of the w-th sensor, r inter denotes the set of inter-sensor label correlation matrices of M sensors;

[0056] (4b2) Through a stacked convolutional layer and a non-linear activation layer, fuse the intra-sensor label correlation matrix r intra and the inter-sensor label correlation matrix r inter to obtain a fused label matrix r. In this embodiment, the size of the convolutional kernel is 1×1, and the non-linear activation layer uses the ReLU activation function;

[0057] (4c) Flatten the fused label matrix r output by the multi-sensor related feature fusion network, and the classifier C f classifies the flattened fused label matrix to obtain I predicted probabilities P f :

[0058] P f = C f (Flatten(r))

[0059] where, Flatten represents the flattening process.

[0060] Step 5) Obtain the trained model.

[0061] (5a) Adopt the cross-entropy loss function, and calculate the loss L f of each classification module and the loss L m of the classifier C f through the initial predicted probability P, the predicted probability P f and the true label y i ​, obtain the loss value L of the fault diagnosis model:

[0062]

[0063]

[0064]

[0065] where λ is the weight coefficient.

[0066] (5b) Through the partial derivatives of the loss value L of the fault diagnosis model with respect to the parameter φ t and the partial derivative of the parameter μ with respect to φ t and the partial derivative of μ update φ t and μ t to obtain the fault diagnosis model O t for this iteration:

[0067]

[0068]

[0069] where φ t+1 and μ t+1 represent the updated results of φ t and μ t , η represents the learning rate, and in this embodiment, η = 0.0001.

[0070] Step 6) Obtain the fault diagnosis target classification result:

[0071] Use the test sample set X test as the input of the trained fault diagnosis model O * for forward propagation to obtain J classification results.

[0072] Next, in combination with the simulation results, the technical effects of the present invention will be further described:

[0073] 1. Experimental conditions and content:

[0074] The hardware platform used in the simulation is: the CPU is AMD Ryzen R9 5900X, the memory is 64GB, and the GPU is NVIDIA GeForce RTX 3080Ti 12GB; the operating system is ubuntu 20.04; the software platform is python 3.8.11 and torch1.10.0.

[0075] 2. Analysis of experimental results:

[0076] Through simulation experiments, the average accuracy rates of 10 experiments obtained by the prior art and the present invention are 80.33% and 98.71% respectively. The classification accuracy rate of the prior art is 80.33%, and the classification performance is relatively low. This is because the prior art directly performs simple dimensionality reduction and splicing fusion on multi-sensor time-domain data, ignoring the complementary information between different sensors. This method first performs dimensionality reduction on multi-sensor data, then converts the sensor data obtained after dimensionality reduction into RGB images, and obtains the classification result by training a fault diagnosis model in a convolutional neural network. Compared with the comparative method, the classification accuracy rate of the present invention has increased by 18.38%. Thus, it can be seen that the present invention can better extract and fuse the complementary information between different sensors, and further improve the fault diagnosis accuracy rate.

[0077] The above prior method is the method proposed in the article "Intelligent Mechanical Fault Diagnosis Using Multisensor Fusion and Convolution Neural Network" published in the journal "IEEE Transactions on Industrial Informatics";

[0078] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in the field, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention. However, these corrections and changes based on the idea of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A fault diagnosis method for rotating mechanical equipment based on multi-sensor related feature fusion, characterized in that, It includes the following steps: (1) Obtain a training sample set and a test sample set: Obtain multiple pieces of original signal data containing C fault categories collected by M sensors and their corresponding fault labels, as well as multiple pieces of real-time signal data of the rotating machinery to be diagnosed, and preprocess each piece of original signal data and each piece of real-time signal data. Then, form a training sample set X with the preprocessed I pieces of original signal data and their corresponding fault labels train , and form a test sample set X with the preprocessed J pieces of real-time signal data test , where M≥2 and C≥2; (2) Build a fault diagnosis model based on a multi-sensor related feature fusion network: Build a fault diagnosis model O including a multi-sensor feature extraction network, a multi-sensor related feature fusion network, and a classifier C connected in sequence, where the multi-sensor feature extraction network includes M network branches arranged in parallel. The m-th network branch includes a global-local time encoder and a time-frequency encoder arranged in parallel, and a fusion module and a classification module C connected in sequence to the output ends of the two encoders. f ; The multi-sensor related feature fusion network is used to fuse the intra-sensor label correlation matrix and the inter-sensor label correlation matrix constructed from the output results of the multi-sensor feature extraction network; m ​ (3) Initialize the parameters: Initialize the number of training iterations as t, the maximum number of iterations as T, and the parameters of the multi-sensor feature extraction network and classifier C at the t-th iteration f are φ t and μ t , and set t = 0; (4) Train the fault diagnosis model: Take the training sample set X train as the input of the fault diagnosis model O for forward propagation to obtain I predicted probabilities P f ; (5) Obtain the trained fault diagnosis model: The parameter φ of the feature extraction network t and classifier C f The parameter μ t Update and obtain the fault diagnosis model O of this iteration t , and judge whether t≥T is established. If so, the trained fault diagnosis model O is obtained. * , otherwise, let t = t + 1, O t =0, and execute step (4); (6) Obtain the fault diagnosis target classification result: Take the test sample set X test as the input of the trained fault diagnosis model O * to perform forward propagation and obtain J classification results.

2. The method according to claim 1, wherein The preprocessing of each piece of original signal data and each piece of real-time signal data described in step (1) is realized as follows: (1a) Divide each piece of original signal data and each piece of real-time signal data into data segments of length l by means of sliding window sampling, and perform normalization processing on each of them to obtain I preprocessed original signal data segments and J real-time signal data segments; (1b) Compose a training sample set from the preprocessed I segments of original signal data and the fault labels corresponding to multiple original signal data segments And use the preprocessed J segments of real-time signal data as the test sample set Among them, x i represents the i-th preprocessed segment of original signal data, and y i represents the fault label corresponding to x i and x j represents the j-th test sample.

3. The method according to claim 1, characterized in that The fault diagnosis model described in step (2), where: The global-local time encoder includes a globally arranged global encoder and a local encoder in parallel, and a feature fusion module cascaded at the output ends of the global encoder and the local encoder; The time-frequency encoder includes K cascaded convolutional structures and a fully connected layer, where the convolutional structure includes stacked convolutional layers, non-linear activation layers, and max pooling layers.

4. The method according to claim 1, wherein The training of the fault diagnosis model described in step (4) is realized as follows: (4a) The global-local time encoder in the m-th network branch extracts features from each training sample to obtain I time domain features; the time-frequency encoder extracts features from each training sample after wavelet transform to obtain I time-frequency domain features; The fusion module performs feature fusion on each time domain feature and its corresponding time-frequency domain feature; The classification module classifies the result of feature fusion to obtain I initial prediction probabilities P m , then the initial prediction probabilities output by the multi-sensor feature extraction network are P = {P 1 , P 2 ,..., P m ,..., P M}; (4b) The multi-sensor related feature fusion network calculates the intra-sensor label correlation matrix of the m-th sensor through the initial prediction probability P Obtain the intra-sensor label correlation matrices r of M sensors intra =[R 1,1 ,R 2,2 ,...,R m,m ,...,R M,M , calculate the inter-sensor label correlation matrix between the u-th sensor and the w-th sensor Obtain the inter-sensor label correlation matrices r of M sensors inter =[R 1,2 ,R 1,3 ,...,R u,w ,...,R M-1,M , and fuse r intra and r inter to obtain the fused label matrix r, where is the transpose matrix of P m , P u and P w respectively represent the initial prediction probabilities of the u-th sensor and the w-th sensor, is the transpose matrix of P w ; (4c) Classifier C f Classify the flattened fusion label matrix to obtain I predicted probabilities P f .

5. The method according to claim 1, wherein The parameters φ of the feature extraction network described in step (5) t and the parameters μ f of the classifier C t are updated, and the implementation steps are as follows: (5a) The cross-entropy loss function is adopted to calculate the loss L of each classification module and the loss L of the classifier C by using the initial prediction probability P and the prediction probability P f respectively with the true label y i to obtain the loss value L of the fault diagnosis model: m and the classifier C f of the loss L f ​ where λ is the weight coefficient; (5b) Update φ and μ by taking the partial derivatives of the loss value L of the fault diagnosis model with respect to parameter φ t and parameter μ to obtain the fault diagnosis model O of this iteration t : Update φ t and μ t : t : where φ t+1 , μ t+1 denote the updated results of φ t , μ t , and η represents the learning rate.

Images