Bearing fault diagnosis method combining multi-scale residual error and multi-task depth separable convolution
By combining the method of separable convolution of multi-scale residual and multi-task depth, sharing and task features are extracted, and attention mechanism and dynamic loss weight algorithm are used to solve the problem of difficulty in completing multi-task bearing fault diagnosis in the existing technology, achieving efficient and accurate multi-task diagnosis effect.
Patent Information
- Application Number
- CN202411732467.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-11
AI Technical Summary
In the prior art, it is difficult to efficiently and accurately complete multi-task fault identification at the same time, such as fault type identification, damage degree identification and operating conditions identification.
Using a method that combines multi-scale residual and multi-task depth separable convolution, shared features are extracted through depth separation convolution, task features are extracted by multi-scale residual module, and features are screened using attention mechanism, dynamic loss weight algorithm balance task training, and online label smoothing optimization model performance.
It realizes efficient and accurate multi-task bearing fault diagnosis, and can simultaneously identify fault types, damage degrees and operating conditions, improving the generalization performance and robustness of the model.
Smart Images
Figure CN119598291A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of mechanical equipment fault diagnosis, and in particular relates to a bearing fault diagnosis method. Background Art
[0002] As the speed, scale and complexity of rotating machinery continue to increase, the requirements for the accuracy and efficiency of fault diagnosis are also getting higher and higher. According to statistics, about 30% of rotating machinery failures in the industrial field are caused by defects in rolling bearings. Due to the harsh working environment of rolling bearings, pitting, wear and fracture are inevitable, which will affect the production efficiency of industrial equipment and even endanger personnel safety. Therefore, it is necessary to monitor the health status of rolling bearings and diagnose them in a timely and accurate manner to ensure the safe and efficient operation of the mechanical system.
[0003] In recent years, deep learning has been widely used in the field of bearing fault diagnosis due to its powerful feature extraction and pattern recognition capabilities. Fault diagnosis methods based on deep learning can avoid the process of manual feature extraction, achieve end-to-end intelligent diagnosis, and improve diagnostic efficiency and accuracy. Common deep learning methods include Convolutional Neural Networks (CNN), Deep Belief Networks (DBN), and Long Short-Term Memory Networks (LSTM). Fault diagnosis methods based on deep learning have achieved great success in single fault diagnosis tasks and have been widely used.
[0004] Gram angular difference field (GADF) is a coding method that converts one-dimensional vibration signals into two-dimensional coding images. It can maintain the original time characteristics while expanding sample features. Multi-task learning (MTL) can learn the shared features of multiple related tasks when training multiple tasks in parallel, and use the useful information between tasks to speed up model training and improve the generalization performance of the model. The depthwise separable network (DSN) effectively reduces the number of model parameters by decomposing the traditional convolution operation into two steps: depth convolution and point-by-point convolution, and maintains the accuracy of the model under the condition of reduced computation. The multi-scale residual module (MRM) can capture features of different scales, improve model performance through rich features, and make the model have good generalization performance. The dynamic loss weight algorithm can effectively balance the training of each subtask in multi-task learning, and keep the learning progress of the subtasks consistent by dynamically adjusting the loss weight of the subtask, thereby accelerating the convergence of the model. Online label smoothing (OLSR) forms soft labels by accepting some erroneous labels when calculating label probabilities, which effectively avoids overfitting in the training process and continuously updates soft labels during model training to improve model performance. Summary of the invention
[0005] The purpose of the present invention is to provide a bearing fault diagnosis method that combines multi-scale residuals with multi-task deep separable convolution to achieve efficient and accurate multi-task bearing fault diagnosis.
[0006] To achieve the above object, the present invention adopts the following technical solution:
[0007] A bearing fault diagnosis method combining multi-scale residuals and multi-task deep separable convolutions comprises the following steps:
[0008] Step a, preprocessing the bearing signal, and then generating a two-dimensional code map through encoding;
[0009] Step b, dividing the two-dimensional code map obtained in step a into a training set, a test set and a validation set;
[0010] Step c: Use depth-wise separable convolution to build a shared network and extract common shared features;
[0011] Step d: Use the multi-scale residual module to build a branch network and extract the task features of the subtask;
[0012] Step e, combining the shared network built in step c with the branch network built in step d to build a multi-task classification model, and inputting the shared features into the branch network after screening them through the attention mechanism;
[0013] Step f, model training: input the training set into the multi-task classification model built in step e, and use the dynamic loss weight algorithm to reasonably allocate computing resources;
[0014] Step g, introduce online label smoothing in the softmax classifier to speed up the model convergence, and output the classification results of each subtask;
[0015] Step h, inputting the validation set into the multi-task classification model after step g to verify the training effect of the model, and fine-tuning the parameters of the model according to the validation results;
[0016] Step i, determine whether the number of network training times m of the multi-task classification model reaches the preset number of iterations N, if yes, proceed to the next step, otherwise repeat step h;
[0017] In step j, the test set is input into the multi-task classification model trained in step i to verify the performance of the multi-task classification model.
[0018] In the step a, the bearing signal is first preprocessed, and a sample signal is obtained by overlapping sampling; then the sample signal is encoded by the Gram angle difference field to generate a two-dimensional encoding diagram; the specific encoding process of the sample signal X by the Gram angle difference field is as follows:
[0019] The sample signal X with n sampling points is n}Normalized to the interval [-1,1], where the first formula is as follows:
[0020]
[0021] After scaling, the sequence Mapped to the polar coordinate system, the value of each element is the corresponding angle cosine, and the corresponding timestamp is the radius. The calculation formula is as follows:
[0022]
[0023] Among them, φ i for The corresponding angle in polar coordinates; t i for The corresponding timestamp; N is a constant coefficient factor, corresponding to The number of elements n; The elements in are evenly distributed on [0,1] according to the timestamps so that they fall on a unit circle;
[0024] The trigonometric difference between each point is used to identify the correlation between elements at different time intervals; the calculation formula of the Gram angle difference field is as follows:
[0025]
[0026] Finally, according to the matrix GADF (i,j) Get the 2D code image and scale it to a 96×96 image.
[0027] In step c, a shared network is constructed using depthwise separable convolution, which is composed of three depthwise separable convolution kernels and four convolution modules. The pooling operation is combined to realize the extraction of common features. The depthwise separable convolution reduces the amount of calculation by decomposing the convolution. The specific calculation formula is as follows:
[0028] N depthwise =D F ×D F ×M
[0029] N pointwise =1×1×M×N
[0030] N S =N depthwise +N pointwise
[0031]
[0032] Among them, N depthwise is the parameter of the feature concatenation layer; D F is the size of the convolution kernel; M is the number of channels of the input feature map; N pointwise is the parameter quantity of the channel fusion layer; N is the number of channels of the output feature map; N S is the total number of parameters of depthwise separable convolution; N C is the parameter amount of ordinary convolution; R is the ratio of the computation amount of depthwise separable convolution to that of ordinary convolution.
[0033] In step d, a branch network is constructed using a multi-scale residual module. Each branch network consists of four multi-scale residual modules. The multi-scale residual module first uses a 1×1 convolution to operate the input feature map; then the obtained feature map x is evenly divided into four subsets, represented as x i ,i∈(1,2,3,4), except x1, all other subsets are operated using 3×3 convolution kernel; the final output feature y i ,i∈(1,2,3,4), merge features through 1×1 convolution, and use residual connection to output features; the specific calculation formula for multi-scale feature extraction is as follows:
[0034]
[0035] Among them, k i () represents the feature map x i The convolution operation performed;
[0036] Set the loss L of each branch network i ,i represents the i-th subtask.
[0037] In step e, the shared features output by the shared network are filtered through the attention mechanism, and the filtered features are Input into the branch network to complete the construction of the multi-task network; the specific calculation formula for the screening feature is as follows:
[0038]
[0039] f=δ(F1([z h ,z w ]))
[0040] g w =σ(F w (f w ))
[0041] g h =σ(F h (f h ))
[0042]
[0043] Where w is the width of the feature; h is the height of the feature; is the pooling result in the vertical direction of the feature map; x c (h,i) is the feature map in the vertical direction; is the pooling result in the horizontal direction of the feature map; x c (j,i) is the feature map in the horizontal direction; f is the feature extracted after the two feature maps are merged; δ is the nonlinear activation function; F1 is a 1×1 convolution that makes the dimension of the output tensor consistent with the dimension of the input X; f w and f h are the two tensors cut out of f; g w f w The corresponding attention weight; g h f h The corresponding attention weight; F w and F h is a 1×1 convolution; X c (i,j) is the input feature; and is the attention weight corresponding to the two directions of the current feature; V c This is the final output of the attention mechanism.
[0044] In step f, the loss L of each branch network is calculated i , the weight of each task is assigned through the dynamic loss weight algorithm; the specific calculation formula is as follows:
[0045]
[0046] Among them, λ i is the loss weight of the current task; s is the sum of loss weights; L all for total mission losses;
[0047] After each group of multi-task learning is completed, the gradient of each subtask is calculated and the parameters are updated according to the gradient direction propagation.
[0048] In step g, the loss is calculated by online label smoothing, and the loss function L i The calculation is improved; the specific calculation formula is as follows:
[0049]
[0050] L i =αL hard +(1-α)L soft
[0051] in, is the target category y i Soft label of p(x i ) is y i The corresponding predicted probability; α is the balance coefficient, the value range is (0,1); L soft is the loss corresponding to the soft label; L hard is the cross entropy loss.
[0052] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0053] (1) The present invention proposes an end-to-end multi-task learning model, which enables different tasks to learn features from each other and can simultaneously complete the fault type identification task, damage degree identification task and operating condition identification task;
[0054] (2) The present invention adopts an attention mechanism in each specific task to screen features that are more effective for the current task, and introduces a multi-scale residual module to extract richer features through convolution kernels of different sizes.
[0055] (3) To address the problem of unbalanced training of tasks in multi-task learning, the present invention introduces a dynamic loss weight algorithm to automatically adjust the weight of the loss function of each task, reasonably allocate computing resources, and optimize the learning progress of different tasks.
[0056] (4) The method of the present invention can automatically learn features and has good intelligence and robustness. The model is simple and interpretable, and is easy for operators to understand and promote. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 It is a schematic diagram of the shared network model structure;
[0058] Figure 2 It is a schematic diagram of the structure of a single branch network model;
[0059] Figure 3 It is a schematic diagram of the multi-task model structure;
[0060] Figure 4 This is the overall algorithm flow chart of the fault diagnosis method that combines multi-scale residuals with multi-task deep separable networks. DETAILED DESCRIPTION
[0061] The present invention will be further explained below in conjunction with the accompanying drawings.
[0062] In view of the fact that the existing fault diagnosis methods can only meet the needs of a single task, the present invention proposes a multi-task learning model. First, the vibration signal is converted into a two-dimensional coding image through the Gram angle difference field to expand the sample features; then, the shared features are extracted through the shared network, and the common features filtered by the attention mechanism are input into the branch network to extract the task features; a dynamic loss weight algorithm is introduced to balance the training progress of the subtasks and accelerate the convergence of the model; finally, the online label smoothing is used in the classification layer to calculate the model loss and improve the model performance. The shared network model structure is as follows: Figure 1 As shown; the branch network model structure is as follows Figure 2 As shown; the multi-task model structure constructed by the shared network and the branch network is as follows Figure 3 As shown; the entire invention flow chart is as follows Figure 4 shown.
[0063] To specifically illustrate the method of the present invention, the present invention analyzes the Case Western Reserve University data set.
[0064] This data set includes the vibration signals of the rolling bearings at the drive end and the fan end measured by using acceleration sensors under different loads. Each load contains four states of the rolling bearing: normal, ball fault, inner ring fault and outer ring fault, and each fault type contains three different fault degrees, namely 0.07 inch damage diameter, 0.14 inch damage diameter and 0.21 inch damage diameter. The required vibration signals can be collected under three operating conditions of 0HP, 1HP and 2HP corresponding to speeds of 1797r / min, 1772r / min and 1750r / min. In order to verify the effectiveness of the multi-task fault diagnosis of the method of the present invention, three classification tasks are set to identify the fault size, fault type and operating condition of the bearing respectively. The specific division method of the data set is shown in Table 1.
[0065] Table 1
[0066]
[0067] A bearing fault diagnosis method combining multi-scale residuals and multi-task deep separable convolution of the present invention comprises the following steps:
[0068] Step a, first preprocess the bearing signal, and obtain the sample signal by overlapping sampling; then encode the sample signal by Gram angle difference field to generate a two-dimensional encoding map; the specific encoding process of the sample signal X by the Gram angle difference field is as follows:
[0069] The sample signal X with n sampling points is n}Normalized to the interval [-1,1], where the first formula is as follows:
[0070]
[0071] After scaling, the sequence Mapped to the polar coordinate system, the value of each element is the corresponding angle cosine, and the corresponding timestamp is the radius. The calculation formula is as follows:
[0072]
[0073] Among them, φ i for The corresponding angle in polar coordinates; t i for The corresponding timestamp; N is a constant coefficient factor, corresponding to The number of elements n; The elements in are evenly distributed on [0,1] according to the timestamps so that they fall on a unit circle;
[0074] The trigonometric difference between each point is used to identify the correlation between elements at different time intervals; the calculation formula of the Gram angle difference field is as follows:
[0075]
[0076] Finally, according to the matrix GADF (i,j) Get the 2D code image and scale it to a 96×96 image.
[0077] Step b, dividing the two-dimensional coding image obtained in step a into a training set, a test set and a validation set.
[0078] Step c: Use depth-wise separable convolution to build a shared network and extract common shared features;
[0079] The shared network is constructed by 3 depth-separable convolution kernels and 4 convolution modules; the pooling operation is combined to realize the extraction of common features; the depth-separable convolution reduces the amount of calculation by decomposing the convolution. The specific calculation formula is as follows:
[0080] N depthwise =D F ×D F ×M
[0081] N pointwise =1×1×M×N
[0082] N S =N depthwise +N pointwise
[0083]
[0084] Among them, N depthwise is the parameter of the feature concatenation layer; D F is the size of the convolution kernel; M is the number of channels of the input feature map; N pointwise is the parameter quantity of the channel fusion layer; N is the number of channels of the output feature map; N S is the total number of parameters of depthwise separable convolution; N C is the parameter amount of ordinary convolution; R is the ratio of the computation amount of depthwise separable convolution to that of ordinary convolution.
[0085] Step d: Use the multi-scale residual module to build a branch network and extract the task features of the subtask;
[0086] Each branch network consists of 4 multi-scale residual modules. The multi-scale residual module first uses 1×1 convolution to operate on the input feature map; then the obtained feature map x is evenly divided into 4 subsets, denoted as x i,i∈(1,2,3,4), except x1, all other subsets are operated using 3×3 convolution kernels; the final output feature y i ,i∈(1,2,3,4), merge features through 1×1 convolution, and use residual connection to output features; the specific calculation formula for multi-scale feature extraction is as follows:
[0087]
[0088] Among them, k i () represents the feature map x i The convolution operation performed;
[0089] Set the loss L of each branch network i ,i represents the i-th subtask.
[0090] Step e, combining the shared network built in step c with the branch network built in step d to build a multi-task classification model, and inputting the shared features into the branch network after screening them through the attention mechanism;
[0091] The shared features output by the shared network are filtered through the attention mechanism, and the filtered features X∈ Input into the branch network to complete the construction of the multi-task network; the specific calculation formula for the screening feature is as follows:
[0092]
[0093] f=δ(F1([z h ,z w ]))
[0094] g w =σ(F w (f w ))
[0095] g h =σ(F h (f h ))
[0096]
[0097] Where w is the width of the feature; h is the height of the feature; is the pooling result in the vertical direction of the feature map; x c (h,i) is the feature map in the vertical direction; is the pooling result in the horizontal direction of the feature map; x c (j,i) is the feature map in the horizontal direction; f is the feature extracted after the two feature maps are merged; δ is the nonlinear activation function; F1 is a 1×1 convolution that makes the dimension of the output tensor consistent with the dimension of the input X; f wand f h are the two tensors cut out of f; g w f w The corresponding attention weight; g h f h The corresponding attention weight; F w and F h is a 1×1 convolution; X c (i,j) is the input feature; and is the attention weight corresponding to the two directions of the current feature; V c This is the final output of the attention mechanism.
[0098] Step f, model training: input the training set into the multi-task classification model built in step e, and use the dynamic loss weight algorithm to reasonably allocate computing resources;
[0099] Calculate the loss L of each branch network i , the weight of each task is assigned through the dynamic loss weight algorithm; the specific calculation formula is as follows:
[0100]
[0101] Among them, λ i is the loss weight of the current task; s is the sum of loss weights; L all for total mission losses;
[0102] After each group of multi-task learning is completed, the gradient of each subtask is calculated and the parameters are updated according to the gradient direction propagation.
[0103] Step g, introduce online label smoothing in the softmax classifier to speed up the model convergence, and output the classification results of each subtask;
[0104] The loss is calculated by online label smoothing, and the loss function L i The calculation is improved; the specific calculation formula is as follows:
[0105]
[0106] L i =αL hard +(1-α)L soft
[0107] in, is the target category y i Soft label of p(x i ) is y i The corresponding predicted probability; α is the balance coefficient, the value range is (0,1); L soft is the loss corresponding to the soft label; Lhard is the cross entropy loss.
[0108] Step h, inputting the validation set into the multi-task classification model after step g to verify the training effect of the model, and fine-tuning the parameters of the model according to the validation results;
[0109] Step i, determine whether the number of network training times m of the multi-task classification model reaches the preset number of iterations N, if yes, proceed to the next step, otherwise repeat step h;
[0110] In step j, the test set is input into the multi-task classification model trained in step i to verify the performance of the multi-task classification model.
[0111] In summary, the present invention proposes a fault diagnosis method combining multi-scale residuals with multi-task deep separable networks. First, a deep separable network is constructed as a parameter sharing unit to extract common features of shared information among multiple tasks; secondly, a task learning module is constructed using a multi-scale residual module to achieve simultaneous training of fault type, damage degree and operating condition tasks; then a dynamic loss weight algorithm is introduced to dynamically adjust the loss weight of each task and control the learning progress of subtasks, so that multiple tasks can complete their goals at the same time; in order to further improve the performance of the network, the screened shared features are input into the branch network through the attention mechanism to improve the correlation between the features and the current task; finally, the one-dimensional vibration signal is converted into a two-dimensional coding image through the Gram angle difference field (GADF) and input into the model for training, so as to improve the model diagnosis efficiency by enriching the features of the fault samples.
[0112] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A bearing fault diagnosis method combining multi-scale residuals and multi-task deep separable convolution, characterized by: The following steps are involved: Step a, preprocessing the bearing signal, and then generating a two-dimensional code map through encoding; Step b, dividing the two-dimensional code map obtained in step a into a training set, a test set and a validation set; Step c: Use depthwise separable convolution to build a shared network and extract common shared features; Step d: Use the multi-scale residual module to build a branch network and extract the task features of the subtask; Step e, combining the shared network built in step c with the branch network built in step d to build a multi-task classification model, and inputting the shared features into the branch network after screening them through the attention mechanism; Step f, model training: input the training set into the multi-task classification model built in step e, and use the dynamic loss weight algorithm to reasonably allocate computing resources; Step g, introduce online label smoothing in the softmax classifier to speed up the model convergence, and output the classification results of each subtask; Step h, inputting the validation set into the multi-task classification model after step g to verify the training effect of the model, and fine-tuning the parameters of the model according to the validation results; Step i, determine whether the number of network training times m of the multi-task classification model reaches the preset number of iterations N, if yes, proceed to the next step, otherwise repeat step h; In step j, the test set is input into the multi-task classification model trained in step i to verify the performance of the multi-task classification model.
2. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In the step a, the bearing signal is first preprocessed, and a sample signal is obtained by overlapping sampling; then the sample signal is encoded by the Gram angle difference field to generate a two-dimensional encoding diagram; the specific encoding process of the sample signal X by the Gram angle difference field is as follows: The sample signal X with n sampling points is n }Normalized to the interval [-1,1], where the first formula is as follows: After scaling, the sequence Mapped to the polar coordinate system, the value of each element is the corresponding angle cosine, and the corresponding timestamp is the radius. The calculation formula is as follows: Among them, φ i for The corresponding angle in polar coordinates; t i for The corresponding timestamp; N is a constant coefficient factor, corresponding to The number of elements n; The elements in are evenly distributed on [0,1] according to the timestamps so that they fall on a unit circle; The trigonometric difference between each point is used to identify the correlation between elements at different time intervals; the calculation formula of the Gram angle difference field is as follows: Finally, according to the matrix GADF (i,j) Get the 2D code image and scale it to a 96×96 image.
3. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In step c, a shared network is constructed using depthwise separable convolution, which is composed of three depthwise separable convolution kernels and four convolution modules. The pooling operation is combined to realize the extraction of common features. The depthwise separable convolution reduces the amount of calculation by decomposing the convolution. The specific calculation formula is as follows: N depthwise =D F ×D F ×M N pointwise =1×1×M×N N S =N depthwise +N pointwise Among them, N depthwise is the parameter quantity of the feature concatenation layer; D F is the size of the convolution kernel; M is the number of channels of the input feature map; N pointwise is the parameter quantity of the channel fusion layer; N is the number of channels of the output feature map; N S is the total number of parameters of depthwise separable convolution; N C is the parameter amount of ordinary convolution; R is the ratio of the computation amount of depthwise separable convolution to that of ordinary convolution.
4. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In step d, a branch network is constructed using a multi-scale residual module. Each branch network consists of four multi-scale residual modules. The multi-scale residual module first uses a 1×1 convolution to operate the input feature map; then the obtained feature map x is evenly divided into four subsets, represented as x i ,i∈(1,2,3,4), except x1, all other subsets are operated using 3×3 convolution kernel; the final output feature y i ,i∈(1,2,3,4), merge features through 1×1 convolution, and use residual connection to output features; the specific calculation formula for multi-scale feature extraction is as follows: Among them, k i () represents the feature map x i The convolution operation performed; Set the loss L of each branch network i ,i represents the i-th subtask.
5. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In step e, the shared features output by the shared network are filtered through the attention mechanism, and the filtered features are Input into the branch network to complete the construction of the multi-task network; the specific calculation formula for the screening feature is as follows: f=δ(F1([z h ,z w ])) g w =σ(F w (f w )) g h =σ(F h (f h )) Where w is the width of the feature; h is the height of the feature; is the pooling result in the vertical direction of the feature map; x c (h,i) is the feature map in the vertical direction; is the pooling result in the horizontal direction of the feature map; x c (j,i) is the feature map in the horizontal direction; f is the feature extracted after the two feature maps are merged; δ is the nonlinear activation function; F1 is a 1×1 convolution that makes the dimension of the output tensor consistent with the dimension of the input X; f w and f h are the two tensors cut out of f; g w f w The corresponding attention weight; g h f h The corresponding attention weight; F w and F h is a 1×1 convolution; X c (i,j) is the input feature; and is the attention weight corresponding to the two directions of the current feature; V c This is the final output of the attention mechanism.
6. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In step f, the loss L of each branch network is calculated i , the weight of each task is assigned through the dynamic loss weight algorithm; the specific calculation formula is as follows: Among them, λ i is the loss weight of the current task; s is the sum of loss weights; L all for total mission losses; After each group of multi-task learning is completed, the gradient of each subtask is calculated and the parameters are updated according to the gradient direction propagation.
7. The bearing fault diagnosis method combining multi-scale residual and multi-task deep separable convolution according to claim 1 is characterized in that: In step g, the loss is calculated by online label smoothing, and the loss function L i The calculation is improved; the specific calculation formula is as follows: L i =αL hard +(1-α)L soft in, is the target category y i Soft label of p(x i ) is y i The corresponding predicted probability; α is the balance coefficient, the value range is (0,1); L soft is the loss corresponding to the soft label; L hard is the cross entropy loss.