Rotating machinery noise label fault diagnosis method based on adaptive symmetric loss

Abstract

The application provides a rotating machinery noise label fault diagnosis method based on adaptive symmetric loss, solves the problem that a rotating machinery fault diagnosis model based on deep learning under a noise label is vulnerable to error-labeled samples, which damages model training, and further leads to reduced diagnosis accuracy. Branch convolution is performed by using a wide convolution kernel, network depth is retained, and extraction of vibration signal features by the network is strengthened. Then, different scale feature information is dynamically weighted and combined, and fault features of the rotating machinery are accurately extracted. Finally, variance evaluation is performed on network output features, the variance evaluation is used as a weight parameter of an adaptive symmetric cross-entropy loss function, and the network parameters are updated through back propagation.

Description

[0001] The present application relates to the field of rotating machinery fault diagnosis, and particularly relates to a rotating machinery noise label fault diagnosis method based on adaptive symmetric loss.

[0002] Rotating machinery components are crucial in most modern mechanical equipment such as wind turbines, high-speed trains, and shield machines. However, they are prone to wear and damage during long-term operation, leading to mechanical equipment downtime and even causing significant economic losses and casualties. Prognostics Health Management (PHM) technology can monitor the state of key components of rotating machinery and diagnose faults, and develop reasonable operation and maintenance strategies based on this. Therefore, fault diagnosis of rotating machinery through state monitoring signals has practical value and research significance.

[0003] Deep learning-based fault diagnosis methods have attracted widespread attention due to their strong non-linear feature self-adaptive extraction capability, and are also favored in rotating machinery fault diagnosis. Current deep learning-based rotating machinery fault diagnosis methods rely heavily on high-quality and accurate fault classification labels for iterative optimization of network parameters. However, obtaining a high-quality and accurate fault dataset is often very challenging and expensive, and some samples may be incorrectly labeled due to subjective or objective reasons. Incorrect labeling will directly affect the optimization and updating of deep neural network parameters, leading to a decrease in fault diagnosis accuracy. This situation where samples are incorrectly labeled is called the noisy label problem. Noisy labels are extremely harmful to deep neural network-based fault diagnosis methods. Unlike signal noise, noisy labels are absolutely harmful to a sample. During the training phase of a deep neural network, it mainly measures the difference between predicted labels and true labels through a loss function to update network parameters through backpropagation. The presence of noisy labels will directly lead to errors in loss function calculation. As the number of deep neural network layers increases and the number of training rounds increases, neural networks that rely on labels to automatically extract high-dimensional features from data will tend to "overfit" to noisy labels, leading to a decrease in diagnosis accuracy.

[0004] ​​For the noise label problem in rotating machinery fault diagnosis, such as a noise label limited data driven variable working condition fan rolling bearing fault diagnosis method disclosed in CN111275024B, an improved rAC-GAN model and its corresponding samples are constructed, and the fault diagnosis accuracy of the rolling bearing in the limited data driven scene and the variable working condition scene is improved. However, in actual engineering, the construction of a special model makes its generalization ability insufficient, and the complex network is difficult to train and optimize. Therefore, how to overcome the shortcomings of the prior art and make the rotating machinery achieve high diagnosis accuracy in the noise label scene is a problem to be solved. SUMMARY

[0005] The present application is to solve the problem that the rotating machinery fault diagnosis model based on deep learning under noise label is easily damaged by error labeled samples, which further leads to the reduction of diagnosis accuracy. A rotating machinery noise label fault diagnosis method based on adaptive symmetric loss is proposed. The method first uses a wide convolution kernel for branch convolution, which strengthens the extraction of vibration signal features while preserving the depth of the network. Secondly, dynamically weighted combination of feature information of different scales can accurately extract the fault features of rotating machinery. Finally, the variance evaluation of the network output features is used as the weight parameter of the adaptive symmetric cross-entropy loss function, and the network parameters are optimized by back propagation.

[0006] To achieve the above purpose, the technical scheme provided by the present application is:

[0007] The rotating machinery noise label fault diagnosis method based on adaptive symmetric loss comprises a dynamic wide convolution residual network and an adaptive symmetric cross-entropy loss function. The adaptive symmetric cross-entropy loss function is adaptively weighted by network output evaluation, and the network parameters are optimized by back propagation. The method comprises the following steps:

[0008] Step 1: Collect rotating machinery fault data and preprocess them into training data set D train and test data set D test , wherein D train is a data set with noise label;

[0009] Step 2: Construct a dynamic wide convolution residual network, input the training data set according to batch Batch into the network, and obtain the features f extracted by the network;

[0010] Step 3: According to the features f extracted in step 2, calculate the weight parameter Var of the adaptive symmetric loss function, and calculate the loss function L ASCE of each batch according to the parameter. Then, the loss function is back propagated to optimize the network parameters to obtain a trained diagnosis model. The corresponding calculation process comprises the following steps.

[0011] Step 3-1, the feature f in step 2 is subjected to Softmax calculation to obtain the distribution probability of each element in the feature map:

[0012]

[0013] wherein f i is the value of the i-th element, and n is the number of fault types; the probability distribution is combined by the formula to form a probability distribution vector P = (p(f 1 ), p(f 2 ),..., p(f n ));

[0014] Step 3-2, the variance of the probability distribution vector P obtained in step 3-1 is calculated to obtain the variance weight Var of the cross-entropy loss function in the adaptive symmetric loss function:

[0015]

[0016] wherein is the mean of the probability distribution vector P;

[0017] Step 3-3, according to the variance weight Var in step 3-2, the adaptive symmetric cross-entropy loss function L ASCE is calculated:

[0018]

[0019] wherein L CCE represents the cross-entropy loss function, L RCE represents the reverse cross-entropy loss function; p(f i ) represents the predicted label probability value of the i-th classification, y i represents the label value corresponding to the sample, and p(f c ) represents the predicted label probability value of the c-th classification; when y i = 0, log(y i ) = A is defined, and A is a constant between -7 and -5;

[0020] Step 3-4, according to the adaptive loss function L ASCE in step 3-3, the loss function value of all samples in each Batch is calculated, and the dynamic wide convolutional network constructed is subjected to back propagation to optimize the network parameters until the training is completed;

[0021] Step 4: using the fault diagnosis model trained in step 3 to diagnose the rotating machinery fault data set.

[0022] Further, the dynamic wide convolution residual network in step 2 comprises a normally connected convolution layer and a dynamic wide convolution layer connected in sequence; and the working process of the dynamic wide convolution layer comprises the following steps:

[0023] Step 2-1, the output f l-1 of the last layer is input into two convolution layers with different kernel sizes, respectively, wherein the first convolution kernel size is 3x1, denoted as F A , and the second convolution kernel size is 3x1 with a dilation rate of 14, denoted as F B ; after the two convolutions, two feature maps U A and U B are obtained, and U A and U B are fused to obtain U:

[0024] U A =F A (f l-1 )

[0025] U B =F B (f l-1 )

[0026] U=Concat(U A ,U B )

[0027] Step 2-2, the U in step 2-1 is compressed by a global average pooling layer to output a feature S C×H×W :

[0028]

[0029] In the formula, GAP is a global average pooling operation, C is a channel number, W is a width, and H is a height, and H and W are both 1;

[0030] Step 2-3, the feature S C×H×W output in step 2-2 is compressed by a fully connected layer FC at a compression rate of r, and the output is:

[0031] V C / r×H×W =FC(S C×H×W )=σ(BN(WS C×H×W ))

[0032] In the formula, σ is an activation function, BN is a batch normalization layer, and W is a fully connected layer weight;

[0033] Step 2-4, the output V C / r×H×W in step 2-3 is processed by two fully connected layers with the same structure and different parameters to obtain two features Z Aand Z B Then, the weights are normalized using the Softmax operation to obtain weights A and B for different convolution kernels and different channels, which are then multiplied by the corresponding U. A and U B The final output f is obtained by weighting. l :

[0034]

[0035] f l =AU A +BU B .

[0036] Furthermore, after acquiring the rotating machinery fault diagnosis data in step 1, the data is divided into a predetermined proportion of training datasets D with non-overlapping segments according to a fixed signal length. train and test dataset D test .

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] The dynamic wide convolutional residual network can dynamically extract rotating machinery fault features at different scales, providing a reliable basis for the calculation of the adaptive symmetric cross-entropy loss function. Furthermore, the dynamic wide convolutional layer can be applied to other convolutional neural networks. Based on this, the proposed adaptive symmetric cross-entropy loss function reduces the L... CCE and L RCE The hyperparameter coefficients are easily calculated and implemented. Experiments show that the proposed method achieves a diagnostic accuracy of over 99% under symmetrical noise tags with different tag noise rates (0%, 10%, 20%, 30%, 40%, and 50%), exhibiting better noise tag robustness compared to other loss functions (CCE, LSCE, GCE, and SCE). [Attached Image Description]

[0039] Figure 1 This is a flowchart of a rotating machinery noise tag fault diagnosis method based on adaptive symmetry loss provided in one embodiment of the present invention;

[0040] Figure 2 This is the dynamic wide convolutional layer structure proposed in this invention;

[0041] Figure 3 This invention proposes a dynamic wide convolutional residual network.

[0042] Figure 4 This is a confusion matrix diagram of the diagnostic results of the method of the present invention under symmetrical noise tags with tag noise rates of 0%, 10%, 20%, 30%, 40%, and 50% in an embodiment of the present invention.

[0043] Figure 5 These are t-SNE clustering visualization diagrams output by each layer of the network in the method of this invention under symmetrical noise labels with a noise label rate of 50% in an embodiment of this invention. (a), (b), (c), (d) and (e) are t-SNE clustering visualization diagrams output by the corresponding layers, respectively.

Detailed Implementation Methods

[0044] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments and accompanying drawings. It should be understood that the illustrative embodiments and descriptions of the present invention are for explaining the present invention only and are not intended to limit the present invention. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined in any way as long as they do not conflict with each other.

[0045] See Figures 1-3 This invention provides a method for fault diagnosis of noise tags in rotating machinery based on adaptive symmetric loss, comprising a dynamic wide convolutional residual network and an adaptive symmetric cross-entropy loss function; the adaptive symmetric cross-entropy loss function is adaptively weighted through network output evaluation, and the parameters are optimized through backpropagation of the network; the method includes the following steps:

[0046] Step 1: Collect rotating machinery fault data and preprocess it into a training dataset D train and test dataset D test , where D train This is a dataset with noise labels;

[0047] Step 2: Construct a dynamic wide convolutional residual network. Input the training dataset into the network in batches (select 128) to obtain the features f extracted by the network.

[0048] Step 3: Based on the features f extracted in Step 2, calculate the weight parameter Var of the adaptive symmetric loss function, and calculate the loss function L for each batch based on this parameter. ASCE Then, backpropagation is performed based on this loss function to optimize the network parameters and obtain the trained diagnostic model. The corresponding calculation process includes the following steps.

[0049] Step 3-1: Perform Softmax calculation on the feature f from Step 2 to obtain the probability distribution of each element in the feature map:

[0050]

[0051] In the formula, f i Let be the value of the i-th element, and n be the number of fault types; this formula will yield a combination of probability distributions, thus forming a probability distribution vector P = (p(f)).1 ), p(f 2 ),..., p(f n ));

[0052] Step 3-2, variance calculation (measure the degree of deviation of the data from the whole) is performed on the probability distribution vector P obtained in step 3-1 to obtain the variance weight Var of the cross-entropy loss function in the adaptive symmetric loss function:

[0053]

[0054] In the formula, is the mean of the probability distribution vector P;

[0055] Step 3-3, according to the variance weight Var in step 3-2, the adaptive symmetric cross-entropy loss function L ASCE is calculated:

[0056]

[0057] In the formula, L CCE represents the cross-entropy loss function, L RCE represents the reverse cross-entropy loss function; p(f i ) represents the predicted label probability value of the i-th classification, y i represents the label value corresponding to the sample, p(f c ) represents the predicted label probability value of the c-th classification; when y i = 0, log(y i ) = A (A takes -7) is defined.

[0058] Step 3-4, according to the adaptive loss function L ASCE in step 3-3, the loss function value of all samples in each Batch is calculated, and the dynamic wide convolutional network constructed is optimized by back propagation to optimize the network parameters until the training is completed (the training round number is 60 rounds);

[0059] Step 4: using the fault diagnosis model trained in step 3 to diagnose the rotating machinery fault data set.

[0060] Specifically, after collecting the rotating machinery fault diagnosis data in step 1, it is divided into mutually non-overlapping training data set D train and test data set D test according to a fixed signal length (2048) and a predetermined proportion (1:1).

[0061] Specifically, the dynamic wide convolutional residual network in step 2 includes sequentially connected ordinary convolutional layers and dynamic wide convolutional layers, and the construction process includes the following steps:

[0062] Step 2-1, the output f l-1 of the last layer is input into two convolutional layer operations with different kernel sizes, respectively, where the first convolutional kernel size is 3x1, denoted as F A , and the second convolutional kernel size is 3x1 with a dilation rate of 14, denoted as F B . (When creating a convolutional operation in the PyTorch environment, set the padding=14 and dilation=14 parameters.) After two convolutional transformations, two feature maps U A and U B are obtained, and U A and U B are fused to obtain U:

[0063] U A = F A (f l-1 )

[0064] U B = F B (f l-1 )

[0065] U = Concat(U A , U B )

[0066] Step 2-2, compress the information of U in step 2-1 through a global average pooling layer to output the feature S C×H×W :

[0067]

[0068] In the formula, GAP is the global average pooling operation, C is the channel number, W is the width, and H is the height, and both H and W are 1.

[0069] Step 2-3, compress the feature S C×H×W output in step 2-2 through a fully connected layer (FC, which is torch.nn.Linear in PyTorch) with a compression rate of r, and the output is:

[0070] V C / r×H×W = FC(S C×H×W ) = σ(BN(WS C×H×W ))

[0071] In the formula, σ is the activation function (ReLU activation function is used to introduce nonlinearity), BN is the batch normalization layer (used to speed up the training and convergence of the network, control the gradient and prevent overfitting), and W is the weight of the fully connected layer.

[0072] Step 2-4, compress the output V C / r×H×WAfter two structure-consistent, parameter-different full connection layers are processed, two features Z of CxHxW shape are obtained A and Z B Then, weight normalization is performed through a Softmax operation, and weights A and B of different convolution kernels and different channels are obtained, which are multiplied by corresponding U A and U B to obtain the final output f l :

[0073]

[0074] f l = AU A + BU B .

[0075] The working process of the application is as follows:

[0076] The working process of the application is as follows:

[0077] I. Data acquisition and processing:

[0078] Taking a rotating machine fault data set as an example, a rotating machine fault data set is collected, and it is divided into training data set D train and test data set D test according to a fixed signal length of 2048 without overlap and in a proportion of 1:1. The specific division setting is shown in Table 1.

[0079] Table 1 Gearbox data set sample processing

[0080]

[0081] II. Model construction:

[0082] Referring to Figure 3 is a dynamic wide convolution residual network structure. The network is based on an 18-layer residual network, and the convolution layers in all residual blocks except the first layer convolution are improved. The structure of each dynamic wide convolution layer is shown in Figure 2 . The dynamic wide convolution residual network is composed of a first layer of convolution, 16 improved convolution layers and a full connection layer, and the specific structure parameters are shown in Table 2.

[0083] Table 2 Network structure parameter table

[0084]

[0085]

[0086] In the data D trainWhen inputting the network, first input it into the first layer of the normal convolution operation in batches, and then input the output f1 into the subsequent network layer. The subsequent network layer is divided into two different residual blocks in a residual connection manner. When the input and output feature channels do not change, residual block 1 is used, and the input feature is directly connected with the original input feature after sequentially passing through a convolution layer, batch normalization, an activation function (ReLU), a convolution layer and batch normalization. When the input and output channels change, residual block 2 is used, and the input feature is convolved after several operations same as residual block 1 to match the changed channel number. Finally, the extracted feature is input into the global average pooling layer for feature information fusion, and then the feature is extracted through the full connection layer to obtain the final classification result.

[0087] III. Model training

[0088] Based on Windows OS and Intel(R) Xeon(R) Silver 4210R CPU@2.40GHz, the training is performed in a PyTorch 1.11.0 environment on a 24GB-NVIDIA's GeForce RTX 3090 GPU. The hyperparameter settings of the training process are shown in Table 3.

[0089] Table 3 Hyperparameter settings

[0090]

[0091] In order to verify the effectiveness of the method, the method is compared with the cross-entropy (CCE), label smoothing cross-entropy (LSCE), generalized cross-entropy (GCE) and symmetric cross-entropy (SCE) loss function methods based on the deep residual network, and the symmetric noise labels with label noise rates of 0%, 10%, 20%, 30%, 40% and 50% are respectively experimented on all data sets. In order to ensure that the experiment is as close to the real working condition as possible, high Gaussian noise with a signal-to-noise ratio (SNR) of 0dB is added to the vibration signal in each data set to verify the robustness of the model to signal noise.

[0092] IV. Model testing and analysis

[0093] Table 4 Gearbox fault diagnosis results

[0094]

[0095] The average diagnostic accuracy and standard deviation of each diagnostic task repeated 10 times are shown in Table 4. Referring to Table 4, when the noise rate is 0%, that is, when the labels of the fault data set are not polluted, the diagnostic accuracy of the method of the application is better than that of other methods. As the label noise rate increases, the methods using CCE, LSCE, GCE, SCE and other loss functions will be affected by the noise labels, resulting in a decrease in diagnostic accuracy. Compared with the case without noise labels, the diagnostic accuracy of the residual network combined with each loss function under 50% label noise rate decreases by 14.58%, 20.23%, 3.55% and 7.14% respectively, and each loss function method is seriously affected by the noise labels. The diagnostic accuracy of the method of the application only decreases by 0.18%, which is relatively stable. This shows that the loss function of the method of the application can maintain good diagnostic accuracy under high noise rate noise labels, and has strong noise resistance. At the same time, as the label noise rate increases, the diagnostic accuracy of the method of the application can be guaranteed to be above 99%, which shows that the method of the application can maintain good noise resistance under different label interference conditions of various noise rates.

[0096] In order to specifically see the influence of the method of the application under different label noise rates, refer to Figure 4 The confusion matrix of the diagnostic results under label noise rates of 0%, 10%, 20%, 30%, 40% and 50% in this embodiment. As can be seen, under the experiments of various label noise rates, the method of the application is more accurate in identifying each category. As the label noise rate increases, the method of the application is less affected, and all classification prediction accuracies are higher than 98.07%, which shows that it has label noise robustness.

[0097] In order to show the feature extraction and clustering ability of the method of the application, refer to Figure 5 The clustering effect diagram of the method of the application in stage 1, stage 2, stage 3, stage 4 and after the full connection layer under 50% noise label rate in this embodiment. As can be seen, as the input sample passes through each stage of the network, the classification boundary of different categories of samples by the method of the application gradually becomes clear and obvious. The classification confusion of stage 2 may be due to the influence of noise labels, making it difficult to identify. After stage 3, the classification boundary gradually becomes clear, but the clustering degree of each sample is not high. After stage 4, it is improved. After full connection, the classification boundary of each sample is relatively clear, which shows that the network can better cluster each classification sample at this time, and the identification boundary is obvious. In summary, the method of the application can better extract the fault diagnosis classification features under noise labels, and output high-precision diagnostic classification results.

[0098] The above examples are illustrative of the present application and are not meant to limit the present application in any way. Any modification of the present application in its spirit or its scope which would be obvious to those skilled in the art is to be included within the scope of the present application.

Claims

1. A method for fault diagnosis of rotating machinery noise tags based on adaptive symmetric loss, characterized in that, The method includes a dynamic wide-range convolutional residual network and an adaptive symmetric cross-entropy loss function; the adaptive symmetric cross-entropy loss function is adaptively weighted through network output evaluation, and the parameters are optimized during network backpropagation; the method includes the following steps: Step 1: Collect rotating machinery fault data and preprocess it into a training dataset D train and test dataset D test , where D train This is a dataset with noise labels; Step 2: Construct a dynamic wide convolutional residual network. Input the training dataset into the network in batches to obtain the features f extracted by the network. Step 3: Based on the features f extracted in Step 2, calculate the weight parameter Var of the adaptive symmetric loss function, and calculate the loss function L for each batch based on this parameter. ASCE Then, backpropagation is performed based on this loss function to optimize the network parameters and obtain the trained diagnostic model. The corresponding calculation process includes the following steps. Step 3-1: Perform Softmax calculation on the feature f from Step 2 to obtain the probability distribution of each element in the feature map: In the formula, f i Let be the value of the i-th element, and n be the number of fault types; this formula will yield a combination of probability distributions, thus forming a probability distribution vector P = (p(f)). 1 ),p(f 2 ),...,p(f n )); Step 3-2: Calculate the variance of the probability distribution vector P obtained in Step 3-1 to obtain the variance weight Var of the cross-entropy loss function in the adaptive symmetric loss function. In the formula, Let P be the mean of the probability distribution vector. Step 3-3: Calculate the adaptive symmetric cross-entropy loss function based on the variance weights Var from Step 3-2. In the formula, L CCE L represents the cross-entropy loss function. RCE p(f) represents the inverse cross-entropy loss function; i y represents the predicted label probability value for the i-th category. i p(f) represents the label value corresponding to the sample. c ) represents the predicted label probability value for class c; when y i When y = 0, define log(y) i ) = A, where A is a constant between -7 and -5; Step 3-4, based on the adaptive loss function L in step 3-3 ASCE Calculate the loss function value for all samples in each batch, and perform backpropagation on the constructed dynamic wide convolutional network to optimize the network parameters until training is complete; Step 4: Use the fault diagnosis model trained in Step 3 to diagnose the rotating machinery fault dataset.

2. The rotating machinery noise tag fault diagnosis method based on adaptive symmetric loss as described in claim 1, characterized in that, The dynamic wide convolutional residual network in step 2 includes sequentially connected ordinary convolutional layers and dynamic wide convolutional layers; the working process of the dynamic wide convolutional layer includes the following steps: Step 2-1, convert the output f of the previous layer... l-1 The inputs are fed into two convolutional layers with different kernel sizes, the first of which has a kernel size of 3×1, denoted as F. A The second convolutional kernel has a size of 3×1 and a dilatation rate of 14, denoted as F. B After two convolutions, two feature maps U are obtained. A and U B , and U A and U B The result of fusion is U: U A =F A (f l-1 ) U B =F B (f l-1 ) U=Concat(U A ,U B ) Step 2-2: Compress the information of U from Step 2-1 through a global average pooling layer to output feature S. C×H×W : In the formula, GAP is the global average pooling operation, C is the number of channels, W is the width, H is the height, and H and W are both 1; Step 2-3: Through a fully connected layer (FC), the features S output in step 2-2 are processed. C×H×W Compression at a compression ratio of r yields the following output: In C / r×H×W =FC(S C×H×W )=σ(BN(WS C×H×W )) In the formula, σ is the activation function, BN is the batch regularization layer, and W is the weight of the fully connected layer; Step 2-4, convert the output V from step 2-3 to... C / r×H×W After processing with two fully connected layers of identical structure but different parameters, two feature Zs with the same shape of C×H×W are obtained. A and Z B Then, the weights are normalized using the Softmax operation to obtain weights A and B for different convolution kernels and different channels, which are then multiplied by the corresponding U. A and U B The final output f is obtained by weighting. l : f l =AU A +BU B .

3. The rotating machinery noise tag fault diagnosis method based on adaptive symmetric loss as described in claim 1, characterized in that, After acquiring the rotating machinery fault diagnosis dataset in step 1, it is divided into a predetermined proportion of training datasets D with fixed signal lengths. train and test dataset D test .

Images