A rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN

Abstract

A method for fault diagnosis of rotating machinery based on multi-sensor fusion and ECA-CNN, 1) Data acquisition: Collect multi-sensor vibration signals of key transmission components of rotating machinery in various health states; 2) Data preprocessing: Sampling, converting and channel fusion of multi-sensor vibration signals to obtain multi-channel images with rich and comprehensive features; 3) Model building: Introducing efficient channel attention into convolutional neural networks to construct ECA-CNN models; 4) Model training: Training the model using multi-channel image sample sets; 5) Fault diagnosis: Using the trained model to perform fault diagnosis on rotating machinery. The present invention realizes channel fusion of multi-sensor vibration signals; introduces efficient channel attention into convolutional neural networks to improve feature learning capabilities and focus attention on more sensitive features; constructs an ECA-CNN model with high computational efficiency to learn representative fault features from multi-sensor fusion data, thereby realizing fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to a rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN. Background Art

[0002] With the vigorous development of modern industrial systems, rotating machinery plays an irreplaceable role in intelligent equipment, and its safety has received extensive attention from academia and industry. Key transmission components of rotating machinery (such as bearings and gears) will inevitably suffer from wear, corrosion, deformation, cracks and other faults under complex operating environments such as high speed and heavy load. The failure of transmission components directly affects the operating reliability of rotating machinery, which may cause major accidents, huge economic losses and even casualties. Therefore, it is of great research value to carry out fault diagnosis and predictive maintenance of rotating machinery.

[0003] In recent years, deep learning (DL) methods have gradually become a hot research topic in the field of rotating machinery fault diagnosis, thanks to their powerful fault feature learning capabilities (without the need for artificially designed features) and end-to-end diagnostic features (mapping raw data to machine health status). DL methods have only two steps: the first step is to sample the raw data to obtain a set containing samples and labels; the second step is to classify the samples using a DL model that integrates feature design, extraction, and classification.

[0004] For the first step of DL, with the rapid development of sensing and communication technologies, a large number of low-cost vibration sensors can be deployed at different positions and directions of rotating machinery to monitor a large amount of operating status data of the equipment. Vibration data is usually a waveform signal (such as acceleration signals and acoustic signals), and there are differences in the characteristics of faults between multi-sensor signals. DL models require a large amount of data with rich features to participate in training. Therefore, multi-sensor signals not only bring new vitality to DL-based fault diagnosis, but also pose many application challenges, such as data idleness and data abuse.

[0005] For the second step of DL, various DL models, such as multi-layer perceptron (MLP), deep belief network (DBN), autoencoder (AE), recurrent neural network (RNN) and convolutional neural network (CNN) have been widely studied and applied to rotating machinery fault diagnosis. In particular, CNN stands out from these algorithms because it is more suitable for processing periodic signals and learning the characteristics of mechanical vibration signals. Although the fault diagnosis method based on CNN performs well in many tasks, it still has the following shortcomings.

[0006] (1) Most existing methods have insufficient data utilization. They only use sensor signals from one position or direction for fault diagnosis, ignoring the differences in fault characteristics between multiple sensor signals. This makes it impossible for the model to thoroughly learn representative fault characteristics, limiting the further improvement of diagnostic performance.

[0007] (2) In CNN, each convolution kernel performs independent convolution operations on the corresponding input channel and generates a feature map (output channel). This means that CNN only processes different channels independently, ignoring the importance of different channel features in the diagnosis results.

[0008] (3) In order to maximize the diagnostic performance, most newly developed models often require complex structures and a large number of parameters, resulting in computational inefficiency and waste of computing resources. Summary of the invention

[0009] In view of the deficiencies in the above-mentioned background technology, the present invention proposes a rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN. The proposed method has three main objectives: (1) to make full use of multi-sensor data; (2) to improve the feature learning ability of CNN; and (3) to improve computational efficiency.

[0010] To achieve the above object, the technical solution adopted by the present invention is:

[0011] A rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN includes the following steps:

[0012] Step 1: Data collection:

[0013] Collect multi-sensor vibration signals of key transmission components of rotating machinery under various health conditions;

[0014] Step 2: Data preprocessing:

[0015] Sampling, converting and channel fusion of multi-sensor vibration signals to obtain multi-channel images with rich and comprehensive features;

[0016] Step 3: Model building:

[0017] Introduce efficient channel attention into convolutional neural network and build ECA-CNN model;

[0018] Step 4: Model training:

[0019] Train the model using a multi-channel image sample set;

[0020] Step 5: Fault diagnosis:

[0021] Use the trained model to perform fault diagnosis on rotating machinery.

[0022] As a further improvement of the present invention, in step 1, multiple sensors are installed at different positions or directions near the rotating mechanical transmission component to be diagnosed, and a data acquisition device is used to collect multi-sensor vibration signals in various health states;

[0023] Assuming that the transmission components of the rotating machinery have c health states, including normal state and c-1 fault states, under certain working conditions, the vibration signal collected by m sensors is defined as follows:

[0024] {X,y}={(X i ,y i )}, i∈[1,c] (1)

[0025]

[0026]

[0027] Among them, {X, y} is defined as the signal and label of c health states, X i ,y i Represents the signal and label of state i, X i j represents the signal of the jth sensor in state i, X i j (N) represents X i j The Nth signal point.

[0028] As a further improvement of the present invention, in step 2, sliding window sampling is performed on the collected multi-sensor signals according to the sample length L and the sliding step S, and the obtained multi-sensor signal sample set is defined as follows:

[0029] {x,y}={(x i ,y i )}, i∈[1,c] (4)

[0030]

[0031]

[0032]

[0033] Among them, {x, y} is defined as the signal samples and labels of c health states, x i ,y i Represent the signal samples and labels of state i respectively, represents the signal sample of the jth sensor in state i, express The kth sample of

[0034] The sampled sensor signal samples are converted into a two-dimensional grayscale image. The continuous signal segments of length L are sequentially filled into the rows of the constructed image. The pixel values ​​of the constructed image are normalized to integers between 0 and 255, corresponding to the pixel intensity of the grayscale image. The pixel value of the constructed image is defined as follows:

[0035]

[0036] in, express The p-th point of , max(·) and min(·) represent the maximum and minimum functions respectively, and round(·) represents the rounding function;

[0037] The converted single-channel images of each sensor signal sample are fused into a multi-channel image in the channel dimension, and the obtained multi-sensor image sample set is defined as follows:

[0038]

[0039]

[0040]

[0041] Where {x,y} fusion Defined as multi-channel image samples and labels of c health states, y i Represent the multi-channel image samples and labels of state i respectively, express The kth sample of Represents the k-th single-channel image of the m-th sensor in state i.

[0042] As a further improvement of the present invention, in step 3, the ECA-CNN model is composed of multiple CNN-ECA blocks, a global average pooling layer GAP, and a classification layer, i.e., a fully connected layer FC+Softmax function, which are connected in sequence, and the number of CNN-ECA blocks is adjusted according to different task requirements;

[0043] The CNN-ECA block is composed of a two-dimensional convolutional layer Conv2d, a batch normalization layer BN, an activation layer ReLU, a two-dimensional maximum pooling layer MaxPool2d and an ECA module connected in sequence;

[0044] The two-dimensional convolution layer Conv2d is shown in the following formula:

[0045]

[0046] Among them, * represents the convolution operation, x il-1 represents the i-th input feature map of the l-1th layer, represents the weight of the jth convolution kernel in the lth layer, represents the deviation of the lth layer, represents the j-th output feature map of the l-th layer;

[0047] The batch normalization layer BN is shown in the following formula:

[0048]

[0049]

[0050]

[0051]

[0052] Among them, h l ={h l(1) ,…,h l(N)} represents the input feature map of the lth layer with a batch size of N, h l(n) ={h1 l (n) ,…,h k l(n)}, represents the j-th output feature map of the l-th layer, u j , σ j 2 Respectively represent h j l The mean and variance of , ε represents a small constant used to prevent invalid calculations when the variance is 0, They represent the scale and translation parameters that need to be learned respectively;

[0053] The activation layer ReLU is shown in the following formula:

[0054]

[0055] The two-dimensional maximum pooling layer MaxPool2d is shown in the following formula:

[0056]

[0057] Among them, max(·) means taking the maximum function, represents the j-th input feature map of the l-th layer, s represents the pooling size, represents the j-th output feature map of the l-th layer;

[0058] The ECA module is shown in the following formula:

[0059]

[0060]

[0061] ω=σ(C1D k (g(X))) (21)

[0062] Where X∈R C×H×W represents the feature map after convolution, C, H and W are the number of channels, height and width of the feature map respectively, g(·) represents the global average pooling GAP operation, ψ(·) represents the function of adaptively calculating the number of cross-channels k, C1D k (·) represents a one-dimensional convolution operation, σ(·) represents the Sigmoid function, and b and γ are constants.

[0063] As a further improvement of the present invention, in step 4, the multi-channel image sample set is split into a training set, a validation set and a test set according to the time series, the training set and the validation set are used to train the model, and the trained model is used to predict the fault type on the test set to test the model performance;

[0064] Multi-channel image samples are first input into CNN for feature learning. Then the ECA module adaptively scores and weights the learned channel features, enhances features related to fault information, and ignores irrelevant features. In the ECA-CNN model, the above operations are performed multiple times, which means that the required feature information is learned layer by layer. Finally, the connection between features and results is established through the global average pooling layer GAP, and the fault classification results are output through the classification layer.

[0065] The ECA-CNN model is trained using a multi-classification cross entropy loss function to measure the performance of fault classification, as shown in the following formula:

[0066]

[0067] Where y=[y0,…,y c-1 ] represents the unique hot encoding of the sample label. When the sample belongs to category i, y i =1, otherwise y i =0; p = [p0,…,p c-1 ] represents the predicted probability of the model output, p i Represents the probability that the sample belongs to the i-th category;

[0068] The training of the ECA-CNN model uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following formula:

[0069] L(θ)=min θ L C(twenty three)

[0070]

[0071] Among them, θ represents the model parameters, α represents the learning rate of the Adam optimization algorithm, Represents the optimal value of the model parameters found by the error back-propagation derivative method.

[0072] Compared with the prior art, the advantages of the present invention are: (1) channel fusion of multi-sensor vibration signals is realized, so that the multi-channel fusion data has rich and comprehensive fault features; (2) efficient channel attention (ECA) is introduced into the convolutional neural network (CNN) to improve the feature learning ability and focus on more sensitive features; (3) an ECA-CNN model with high computational efficiency is constructed to learn representative fault features from multi-sensor fusion data, thereby realizing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is a flow chart of the method of the present invention;

[0074] Figure 2 The sampling process of multi-sensor signals of the method of the present invention;

[0075] Figure 3 The signal-to-image conversion process of the method of the present invention;

[0076] Figure 4 The channel fusion process of the multi-sensor image of the method of the present invention;

[0077] Figure 5 It is a schematic diagram of the structure of the ECA-CNN model of the method of the present invention;

[0078] Figure 6 It is a schematic diagram of the structure of the CNN-ECA block of the method of the present invention;

[0079] Figure 7 It is a schematic diagram of the structure of the ECA module of the method of the present invention. DETAILED DESCRIPTION

[0080] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] In this embodiment, if Figure 1As shown in the figure, a rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN mainly includes: step 1, data acquisition: collecting multi-sensor vibration signals of key transmission components (such as bearings and gears) of rotating machinery in various health states; step 2, data preprocessing: sampling, converting and channel fusion of multi-sensor vibration signals to obtain multi-channel images with rich and comprehensive features; step 3, model building: introducing efficient channel attention (ECA) into convolutional neural network (CNN) to construct an ECA-CNN model; step 4, model training: using a multi-channel image sample set to train the model; step 5, fault diagnosis: using the trained model to diagnose the fault of rotating machinery.

[0082] In step one, multiple sensors are installed at different positions or directions near the rotating mechanical transmission parts (such as bearings and gears) that need to be diagnosed, and a data acquisition device is used to collect multi-sensor vibration signals under various health conditions.

[0083] Assume that the transmission components of the rotating machinery have c health states (normal state and c-1 fault states). Under certain working conditions, the vibration signal collected by m sensors is defined as follows:

[0084] {X,y}={(X i ,y i )}, i∈[1,c] (1)

[0085]

[0086]

[0087] Among them, {X, y} is defined as the signal and label of c health states, X i ,y i Represents the signal and label of state i, X i j represents the signal of the jth sensor in state i, X i j (N) represents X i j The Nth signal point.

[0088] In this embodiment, acceleration sensors are installed in the X, Y, and Z directions of the planetary gearbox housing to collect multi-sensor vibration signals of bearings and gears, respectively, including two working conditions, namely, speed-load 20Hz-0V and 30Hz-2V respectively. Bearings and gears have five health states under different working conditions, as shown in Table 1 below:

[0089]

[0090] like Figure 2As shown, in step 2, the collected multi-sensor signals are sampled by sliding window according to the sample length L and the sliding step S, and the obtained multi-sensor signal sample set is defined as follows:

[0091] {x,y}={(x i ,y i )}, i∈[1,c] (4)

[0092]

[0093]

[0094]

[0095] Among them, {x, y} is defined as the signal samples and labels of c health states, x i ,y i Represent the signal samples and labels of state i respectively, represents the signal sample of the jth sensor in state i, express The kth sample of .

[0096] like Figure 3 As shown, the sampled sensor signal samples are converted into a two-dimensional grayscale image, and the continuous signal segments of length L fill the rows of the constructed image in sequence. The pixel values ​​of the constructed image are normalized to integers between 0 and 255, corresponding to the pixel intensity of the grayscale image. The pixel value of the constructed image is defined as follows:

[0097]

[0098] in, express The p-th point of , max(·) and min(·) represent the maximum and minimum functions respectively, and round(·) represents the integer function.

[0099] In this embodiment, the sample length L is 1024, which means that the side length a of the constructed image is 32.

[0100] like Figure 4 As shown, the single-channel images of each sensor signal sample obtained by conversion are fused into a multi-channel image in the channel dimension, and the obtained multi-sensor image sample set is defined as follows:

[0101]

[0102]

[0103]

[0104] Where {x,y} fusion Defined as multi-channel image samples and labels of c health states, y i Represent the multi-channel image samples and labels of state i respectively, express The kth sample of Represents the k-th single-channel image of the m-th sensor in state i.

[0105] In this embodiment, if Figure 5 As shown, in step 3, the ECA-CNN model is composed of three CNN-ECA blocks, a global average pooling layer (GAP), and a classification layer (fully connected layer (FC) + Softmax function) connected in sequence. The detailed parameter configuration of the ECA-CNN model is shown in Table 2 below:

[0106]

[0107]

[0108] like Figure 6 As shown, the CNN-ECA block is composed of a two-dimensional convolutional layer (Conv2d), a batch normalization layer (BN), an activation layer (ReLU), a two-dimensional maximum pooling layer (MaxPool2d) and an ECA module connected in sequence.

[0109] The two-dimensional convolutional layer (Conv2d) is shown in the following formula:

[0110]

[0111] Among them, * represents the convolution operation, x i l-1 represents the i-th input feature map of the l-1th layer, represents the weight of the jth convolution kernel in the lth layer, represents the deviation of the lth layer, represents the j-th output feature map of the l-th layer.

[0112] The batch normalization layer (BN) is shown in the following formula:

[0113]

[0114]

[0115]

[0116]

[0117] Among them, h l ={hl(1) ,…,h l(N)} represents the input feature map of the lth layer with a batch size of N, h l(n) ={h1 l (n) ,…,h k l(n)},y j l(n) represents the j-th output feature map of the l-th layer, u j , σ j 2 Respectively represent h j l The mean and variance of , ε represents a small constant used to prevent invalid calculations when the variance is 0, γ l j , β l j They represent the scale and translation parameters that need to be learned respectively.

[0118] The activation layer (ReLU) is shown in the following formula:

[0119]

[0120] The two-dimensional maximum pooling layer (MaxPool2d) is shown in the following formula:

[0121]

[0122] Among them, max(·) means taking the maximum function, represents the j-th input feature map of the l-th layer, s represents the pooling size, represents the j-th output feature map of the l-th layer.

[0123] like Figure 7 As shown, the ECA module is as shown in the following formula:

[0124]

[0125]

[0126]

[0127] Where X∈R C×H×W represents the feature map after convolution, C, H and W are the number of channels, height and width of the feature map respectively, g(·) represents the global average pooling (GAP) operation, ψ(·) represents the function of adaptively calculating the number of channels k across channels, C1D k (·) represents a one-dimensional convolution operation, σ(·) represents the Sigmoid function, and b and γ are constants.

[0128] In this embodiment, b=1 and γ=2.

[0129] In step 4, the multi-channel image sample set is split into a training set, a validation set, and a test set according to the time series. The training set and the validation set are used to train the model, and the trained model is used to predict the fault type on the test set to test the model performance.

[0130] In this embodiment, 1022 multi-channel image samples are obtained for each health state under the two working conditions, of which the first 614 are training samples, the middle 204 are verification samples, and the last 204 are test samples.

[0131] The multi-channel image samples are first input into CNN for feature learning, and then the ECA module adaptively scores and weights the learned channel features, enhancing the features related to the fault information and ignoring irrelevant features. In the ECA-CNN model, the above operations are performed multiple times, which means that the required feature information is learned layer by layer. Finally, the connection between the features and the results is established through the global average pooling layer (GAP), and the fault classification results are output through the classification layer (fully connected layer (FC) + Softmax function).

[0132] The ECA-CNN model is trained using a multi-classification cross entropy loss function to measure the performance of fault classification, as shown in the following formula:

[0133]

[0134] Where y=[y0,…,y c-1 ] represents the unique hot encoding of the sample label. When the sample belongs to category i, y i =1, otherwise y i =0; p = [p0,…,p c-1 ] represents the predicted probability of the model output, p i It represents the probability that the sample belongs to the i-th category.

[0135] The training of the ECA-CNN model uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following formula:

[0136] L(θ)=min θ L C (twenty three)

[0137]

[0138] Among them, θ represents the model parameters, α represents the learning rate of the Adam optimization algorithm, Represents the optimal value of the model parameters found by the error back-propagation derivative method.

[0139] In this embodiment, the training hyperparameters of the ECA-CNN model are configured as follows: the learning rate α of the Adam optimization algorithm is 0.001, the batch size is 64, and the number of iterations is 100.

[0140] In this embodiment, different sub-datasets are used to train and perform performance tests on the ECA-CNN model, including the same data set under different working conditions, the same data set under the same working conditions, a mixed data set under the same working conditions, and a complete data set, and four existing classic models based on CNN (2D-CNN, 2D-LeNet, 2D-ResNet18, 2D-BiLSTM) are selected for comparative analysis. It should be noted that all models use multi-sensor fusion data and the same training strategy. In addition, in order to reduce the impact of randomness, each experiment is repeated five times, and the mean and standard deviation are analyzed as the final experimental results. The results of the comparative experiment are shown in Table 3 below:

[0141]

[0142] As can be seen from the table, ECA-CNN outperforms the four comparison models on each sub-dataset, with an average accuracy of more than 99% and a small standard deviation. This shows that ECA-CNN has excellent fault diagnosis performance and has good adaptability and stability. It can also be seen that the diagnosis results of all models on the bearing dataset are generally better than those on the gear dataset. In particular, the average accuracy of ECA-CNN on the bearing dataset has reached 100%. In addition, although the diagnosis results of 2D-CNN, 2D-ResNet18 and 2D-BiLSTM are not as good as ECA-CNN, their performance on each sub-dataset is still acceptable with the help of multi-sensor fusion data.

[0143] In this embodiment, the number of parameters (Params) and the number of floating-point operations (FLOPs) are used as evaluation indicators of the model computational complexity. The number of parameters (Params) refers to the total number of parameters of the model, and the number of floating-point operations (FLOPs) refers to the number of floating-point operations in the forward propagation process of the model. The computational efficiency of the model is reflected by diagnostic performance and computational complexity. The number of parameters (Params) and floating-point operations (FLOPs) of ECA-CNN and the four comparison models on the complete data set are shown in Table 4 below:

[0144]

[0145]

[0146] The structure of NoECA-CNN is the same as that of ECA-CNN, except that the ECA module is missing. Although each CNN-ECA block of ECA-CNN contains an ECA module, the computational complexity is almost not increased, which shows that the ECA module has a very low computational complexity. The number of parameters (Params) of ECA-CNN and NoECA-CNN is only 0.024M, and the number of floating-point operations (FLOPs) is only 2.9M, which is much smaller than the four comparison models. ECA-CNN has both low computational complexity and excellent fault diagnosis performance, which shows that ECA-CNN has high computational efficiency. In addition, among them, the diagnostic performance of 2D-CNN is slightly worse than that of ECA-CNN, but the number of parameters (Params) and floating-point operations (FLOPs) are much larger than those of ECA-CNN.

[0147] It can be seen from the above embodiments that the present invention is a rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN. Compared with the prior art, the advantages of the present invention are: (1) channel fusion of multi-sensor vibration signals is realized, so that the multi-channel fusion data has rich and comprehensive fault characteristics; (2) efficient channel attention (ECA) is introduced into the convolutional neural network (CNN) to improve the feature learning ability and focus on more sensitive features; (3) an ECA-CNN model with high computational efficiency is constructed to learn representative fault features from multi-sensor fusion data, thereby realizing fault diagnosis.

[0148] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A rotating machinery fault diagnosis method based on multi-sensor fusion and ECA-CNN, characterized in that: The steps include: Step 1: Data collection: Collect multi-sensor vibration signals of key transmission components of rotating machinery under various health conditions; In step 1, multiple sensors are installed at different positions or directions near the rotating mechanical transmission component to be diagnosed, and multi-sensor vibration signals under various health states are collected using a data acquisition device; Assuming that the transmission components of the rotating machinery have c health states, including normal state and c-1 fault states, under certain working conditions, the vibration signal collected by m sensors is defined as follows: {X,y}={(X i ,and i )},i∈[1,c] (1) Among them, {X, y} is defined as the signal and label of c health states, X i ,y i Represents the signal and label of state i, X i j represents the signal of the jth sensor in state i, X i j (N) represents X i j The Nth signal point; Step 2: Data preprocessing: Sampling, converting and channel fusion of multi-sensor vibration signals to obtain multi-channel images with rich and comprehensive features; In step 2, the collected multi-sensor signals are sampled using a sliding window according to the sample length L and the sliding step S. The obtained multi-sensor signal sample set is defined as follows: {x,y}={(x i ,and i )},i∈[1,c] (4) Among them, {x, y} is defined as the signal samples and labels of c health states, x i ,y i Represent the signal samples and labels of state i respectively, represents the signal sample of the jth sensor in state i, express The kth sample of The sampled sensor signal samples are converted into a two-dimensional grayscale image. The continuous signal segments of length L are sequentially filled into the rows of the constructed image. The pixel values ​​of the constructed image are normalized to integers between 0 and 255, corresponding to the pixel intensity of the grayscale image. The pixel value of the constructed image is defined as follows: in, express The p-th point of , max(·) and min(·) represent the maximum and minimum functions respectively, and round(·) represents the rounding function; The converted single-channel images of each sensor signal sample are fused into a multi-channel image in the channel dimension, and the obtained multi-sensor image sample set is defined as follows: Where {x,y} fusion Defined as multi-channel image samples and labels of c health states, y i Represent the multi-channel image samples and labels of state i respectively, express The kth sample of Represents the kth single-channel image of the mth sensor in state i; Step 3: Model building: Introduce efficient channel attention into convolutional neural network and build ECA-CNN model; In step 3, the ECA-CNN model is composed of multiple CNN-ECA blocks, a global average pooling layer GAP, and a classification layer, i.e., a fully connected layer FC+Softmax function, connected in sequence, and the number of CNN-ECA blocks is adjusted according to the needs of different tasks; The CNN-ECA block is composed of a two-dimensional convolutional layer Conv2d, a batch normalization layer BN, an activation layer ReLU, a two-dimensional maximum pooling layer MaxPool2d and an ECA module connected in sequence; The two-dimensional convolution layer Conv2d is shown in the following formula: Among them, * represents the convolution operation, x i l-1 represents the i-th input feature map of the l-1th layer, w l ij represents the weight of the jth convolution kernel in the lth layer, represents the deviation of the lth layer, represents the j-th output feature map of the l-th layer; The batch normalization layer BN is shown in the following formula: Among them, h l ={h l(1) ,…,h l(N) } represents the input feature map of the lth layer with a batch size of N, h l(n) ={h1 l(n) ,…,h k l (n) },y j l(n) represents the j-th output feature map of the l-th layer, u j , σ j 2 Respectively represent h j l The mean and variance of , ε represents a small constant used to prevent invalid calculations when the variance is 0, They represent the scale and translation parameters that need to be learned respectively; The activation layer ReLU is shown in the following formula: The two-dimensional maximum pooling layer MaxPool2d is shown in the following formula: Among them, max(·) means taking the maximum function, represents the j-th input feature map of the l-th layer, s represents the pooling size, represents the j-th output feature map of the l-th layer; The ECA module is shown in the following formula: ω=σ(C1D k (g(X))) (21) Where X∈R C×H×W represents the feature map after convolution, C, H and W are the number of channels, height and width of the feature map respectively, g(·) represents the global average pooling GAP operation, ψ(·) represents the function of adaptively calculating the number of channels k across channels, C1D k (·) represents a one-dimensional convolution operation, σ(·) represents the Sigmoid function, and b and γ are constants; Step 4: Model training: Train the model using a multi-channel image sample set; In step 4, the multi-channel image sample set is split into a training set, a validation set, and a test set according to the time series. The training set and the validation set are used to train the model, and the trained model is used to predict the fault type on the test set to test the model performance. Multi-channel image samples are first input into CNN for feature learning. Then the ECA module adaptively scores and weights the learned channel features, enhances features related to fault information, and ignores irrelevant features. In the ECA-CNN model, the above operations are performed multiple times, which means that the required feature information is learned layer by layer. Finally, the connection between features and results is established through the global average pooling layer GAP, and the fault classification results are output through the classification layer. The ECA-CNN model is trained using a multi-classification cross entropy loss function to measure the performance of fault classification, as shown in the following formula: Where y=[y0,…,y c-1 ] represents the unique hot encoding of the sample label. When the sample belongs to category i, y i =1, otherwise y i =0; p = [p0,…,p c-1 ] represents the predicted probability of the model output, p i Represents the probability that the sample belongs to the i-th category; The training of the ECA-CNN model uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following formula: L(θ)=min θ L C (23) θ←Adam(▽ θ L C ,i,a)(24) Among them, θ represents the model parameters, α represents the learning rate of the Adam optimization algorithm, ▽ θ L C Represents the optimal value of the model parameters found by the error back-propagation derivative method; Step 5: Fault diagnosis: Use the trained model to perform fault diagnosis on rotating machinery.

Images