Early weak fault diagnosis system and diagnosis method for ball bearing

Abstract

The application discloses a ball bearing early weak fault diagnosis system and method based on a double-channel dense connection convolutional network, which comprises the following steps: acquiring vibration signal values of a ball bearing early weak fault sampled in time sequence and dividing the vibration signal values into training samples and test samples; constructing a double-channel dense connection convolutional network model; adopting a classification module to classify total feature vectors output by the model; training the double-channel dense connection convolutional network model by using the training samples; and inputting the test samples into the trained double-channel dense connection convolutional network model. The application uses deep learning to realize shallow and deep fault feature learning of small sample data, adopts a new dense connection network model to improve network performance, and can be used for early weak fault detection of ball bearings and has strong self-adaptive detection capability and high detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment diagnosis and health management, specifically to an early weak fault diagnosis system and method for ball bearings based on a dual-channel densely connected convolutional network. Background Technology

[0002] Ball bearings are critical components widely used in mechanical equipment. Due to early-stage failures, overloads, fatigue, wear, and corrosion, ball bearings are prone to damage during machine operation. In fact, early-stage failures of ball bearings can lead to severe equipment shaking, machine shutdown, production stoppage, and even personal injury. Generally, early-stage ball bearing failures are complex and difficult to detect. Therefore, monitoring and analyzing the condition of ball bearings is crucial, as it can detect early-stage failures (including outer ring failure, inner ring failure, and ball failure) and prevent losses caused by these failures. Recently, the fault detection and diagnosis of ball bearings has received considerable attention. Among all methods for diagnosing early-stage failures of various types of ball bearings, vibration signal analysis is one of the most important and useful tools.

[0003] Intelligent fault diagnosis is of great significance for the early weak fault diagnosis of ball bearings, mainly in improving diagnostic accuracy, enabling real-time monitoring, and reducing maintenance costs. It is crucial in modern complex rotating machinery systems. Therefore, to ensure the production efficiency and operational safety of various complex high-speed rotating machinery systems, research on early weak fault detection and diagnosis methods for ball bearings is receiving increasing attention. Common fault diagnosis methods include artificial neural networks, support vector machines, and extreme learning machines. While these methods have achieved certain results in fault diagnosis, they are all shallow learning methods, requiring manual feature extraction and exhibiting weak robustness. To avoid the shortcomings of traditional diagnostic methods that require manual feature extraction and to improve the robustness of the model under non-stationary operating conditions, a deep learning-based early weak fault diagnosis system and method for ball bearings based on a dual-channel densely connected convolutional network is needed. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an early weak fault diagnosis system and method for ball bearings based on a dual-channel densely connected convolutional network using deep learning, which addresses the shortcomings of the prior art. This diagnostic system and method uses deep learning to achieve shallow and deep fault feature learning with small sample data, eliminating the need for manual feature extraction. A new densely connected network model is adopted, which improves network performance. This invention can be applied to the detection of early weak faults in ball bearings (fault types include: outer ring faults, inner ring faults, and ball faults), with strong adaptive detection capability and high detection accuracy.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0006] An early weak fault diagnosis method for ball bearings based on a dual-channel densely connected convolutional network includes:

[0007] Step 1: Obtain vibration signal values ​​of ball bearings in a time-series continuous sampling, and randomly divide them into training samples and test samples, and proportionally divide the training samples into unordered verification samples.

[0008] Step 2: Construct a dual-channel dense connection module;

[0009] Step 3: Construct a dual-channel densely connected convolutional network model, including a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, a fully connected layer, and a classification and recognition module; the deep feature extraction module includes three dual-channel densely connected modules; the input features first pass through the shallow feature extraction module, then through the three dual-channel densely connected modules for deep feature learning, and then through the parameter reduction module before being imported into the fully connected layer to obtain the total feature vector;

[0010] A classification and recognition module is used to classify early weak faults in ball bearings based on the total feature vector output by the fully connected layer;

[0011] Step 4: Set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, and validate it using proportionally divided validation samples. Save the training weights of the best parameters for easy testing, and obtain the trained dual-channel densely connected convolutional network model.

[0012] Step 5: Input the test samples into the trained dual-channel densely connected convolutional network model; set up the interface to achieve automatic diagnosis of early weak faults in ball bearings and integrate it into an expert system.

[0013] As a further improvement to the present invention, step 1 specifically comprises:

[0014] 1.1 Obtain the vibration signal values ​​of all bearing faults by continuous sampling in time series, filter out the time series and continuous sampling vibration signals of early weak faults of ball bearings, including the original sample data, check whether the labels and data correspond, and ensure that the labels and data correspond accurately.

[0015] 1.2 Each sample includes 6000 vibration signal sampling points obtained by sequential sampling in chronological order, with sampling point numbers ranging from 1 to 6000. The 6000 vibration signal sampling points of each sample are stored in a CSV file row by row, with 0, 1, 2, and 3 representing no fault, outer ring fault, inner ring fault, and ball bearing fault, respectively. A portion of the vibration signal data and labels are stored in corresponding CSV files as the training dataset. The other portion of the vibration signal data is stored unlabeled in a CSV file as the test set. The ratio of training set to test set is 8:2. The numerical data and labels of each sample are kept completely paired to ensure that there is no data loss or mismatch.

[0016] 1.3 Define the CSVDataset class, a CSV data loader for the early weak fault network of ball bearings, to read sample vibration signal numerical data, extract relevant parameters from the labels, and arrange them in a uniform format: , The value is assigned to the category to which the label belongs. To record the true, time-sequential, continuously sampled vibration signals, the first 6000 vibration signal sampling points and the last label column are converted into tensors and set as floating-point numbers.

[0017] As a further improvement to the present invention, step 2 specifically comprises:

[0018] 2.1 Construct a dual-channel dense connection module, which includes two dual-channel convolutional layers and one parameter reduction layer, with the two dual-channel convolutional layers and the parameter reduction layer connected in series.

[0019] 2.2 A single dual-channel convolutional layer includes a left convolutional channel module and a right convolutional channel module. The left convolutional channel module includes one 3×3 convolutional layer with a stride of 1 and an inner margin of 1, one 1×1 convolutional layer with a stride of 1, one batch layer, and two activation layers. The right convolutional channel module includes one 5×5 convolutional layer with a stride of 1 and an inner margin of 2, one 1×1 convolutional layer with a stride of 1, one batch layer, and two activation layers.

[0020] After processing from the data loader, let the input feature of a single dual-channel convolutional layer be A. After A passes through the left convolutional channel module of the first dual-channel convolutional layer in the dual-channel densely connected module, the output feature is A1. After A passes through the right convolutional channel module of the first dual-channel convolutional layer, the output feature is A2. The output feature maps of the left and right convolutional channel modules maintain the same size but change the number of channels, extracting simple feature information while preserving spatial location information. Finally, A... 1、 The result of concatenating A2 and input feature A is A+ A1+ A2;

[0021] 2.3. Densely connect the two dual-channel convolutional layers, i.e., the output feature in step 2.2 is A + A1 + A2 = B. Then, after being superimposed with the input feature A, the input feature of the second dual-channel convolutional layer is A + B = C. C passes through the left convolutional channel module of the second dual-channel convolutional layer and outputs the feature C1. C passes through the right convolutional channel module of the second dual-channel convolutional layer and outputs the feature C2. Concatenate C1 and C2 to get C1 + C2 = C3. Superimpose C3 with the input feature C of the second dual-channel convolutional layer and the input feature A to get the output feature A + C + C3 = D. Change the number of channels to extract deep feature information.

[0022] 2.4 The parameter reduction layer consists of one convolutional layer, one batch processing layer, one activation function layer, one max pooling layer, and one dropout layer, configured to perform in-depth processing on the final output latent feature D in 2.3; the activation function layer uses a faulty modified linear unit, and the dropout rate P of the dropout layer is set to 0.2.

[0023] As a further improvement to the present invention, step 3 specifically comprises:

[0024] 3.1 Construct a dual-channel densely connected convolutional network model, including a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, and a fully connected layer;

[0025] 3.2 The shallow feature extraction module includes a 3×3 convolutional layer with a stride of 1 and an inner margin of 1, a batch processing layer, and an activation function layer; the feature map size remains unchanged, simple feature information is extracted, and spatial location information is preserved. After the input features are processed by the shallow feature extraction module, the output feature is X.

[0026] 3.3 The deep feature extraction module includes three dual-channel densely connected modules; the three dual-channel densely connected modules are densely connected sequentially; the output feature X processed by the shallow feature extraction module in step 3.2 is used as the input feature of the first dual-channel densely connected module, and the output feature after processing by the first dual-channel densely connected module is X1; since the three dual-channel densely connected modules are densely connected, the input feature of the second dual-channel densely connected module is the superposition of the output feature X1 of the first dual-channel densely connected module and the output feature X of the shallow feature extraction module, which is X + X1 = X2, thus obtaining feature X2; feature X2 is processed by the... After processing by two dual-channel dense connection modules, the output feature is Z. The superposition feature of Z with the input feature X2 of the second dual-channel dense connection module and the output feature X of the shallow feature extraction module is X+X2+Z=Z1, which gives feature Z1. Z1 is used as the input feature of the third dual-channel dense connection module. The output feature after passing through the third dual-channel dense connection module is Z2. Z2 is superimposed with the input feature Z1 of the third dual-channel dense connection module, the output feature X of the shallow feature extraction module, and the input feature X2 of the second dual-channel dense connection module to give the final output feature Z2+Z1+X+X2=Y, which gives feature Y.

[0027] 3.4 After feature Y passes through a convolutional layer, a batch processing layer, an activation function layer, and an average pooling layer in the parameter reduction module, the pooled feature Y1 is obtained.

[0028] 3.5 After the feature Y1 is flattened, it is processed by a fully connected layer, and the output feature tensor is obtained based on the weight matrix W and the bias vector b of the connected layer. The formula for a fully connected component is: , , i=1,2,…,m; where, It is the weight in the i-th row and j-th column of the weight matrix W. , The output feature tensor of the fully connected layer is used to complete the construction of a dual-channel densely connected convolutional network model.

[0029] 3.6 The classification and recognition module includes a Softmax function, which converts the output feature tensor of the fully connected layer in step 3.5 into a variable. The data is processed sequentially using the Softmax function, and the classification results are output.

[0030] The formula for the Softmax function is:

[0031] ;

[0032] In the formula, This represents the probability that the i-th type of fault sample is correctly identified. Let represent the output value of the i-th network node, and C represent the number of fault sample categories. Let O represent the output values ​​of all network nodes for the fault category of the c-th sample; the probability distribution tensor O for all fault category labels can be obtained:

[0033] ;

[0034] in, The function converts the output matrix into a probability distribution.

[0035] As a further improvement to the present invention, step 4 specifically comprises:

[0036] 4.1 Initialize the training and validation dataset split ratio η, learning rate η, and batch size S; define and initialize the accuracy and loss changes during training and validation, initialize the model weight parameters, and define the storage path for the model parameters best_model_wts when the best accuracy is achieved.

[0037] 4.2. Divide the dataset according to step 1.3. Load the continuous time series data and extract the label information. In steps 2 and 3, build a dual-channel densely connected convolutional network model and initialize the parameter pool to store the information of each round of traversal.

[0038] 4.3 Determine the number of training rounds and batch size, select the Adam optimizer, and define the cross-entropy loss function; in each training round, after the training step size reaches the specified batch size, the training ends and the next round begins;

[0039] 4.4 In each batch of training, set up a CSVDataset data loader, select all data samples in a batch, convert the 6000 columns of feature data of each sample into tensors, and then perform standardization, normalization, randomization, and structural reconstruction; convert the corresponding label values ​​into tensors as well, and import the feature values ​​and label values ​​into the GPU for processing at the same time; perform forward propagation calculation of the model.

[0040] 4.5 Input the loaded data feature tensor into the shallow feature extraction module in step 3.2, and output the feature tensor M after convolution, normalization and activation function;

[0041] 4.6 The feature tensor M in step 4.5 is used as the input to the deep feature extraction module in step 3.3. After passing through the deep feature extraction module in step 3.3, it is then input to the parameter reduction module in step 3.4 for parameter reduction processing, and the output feature is N.

[0042] 4.7 The output feature tensor N in step 4.6 is output as a feature tensor after passing through the fully connected layer in step 3.5;

[0043] 4.8. The feature tensor output by the fully connected layer in step 4.7 is used as the output of the forward propagation calculation of the model. After being processed by the argmax function, the predicted label value is obtained. The loss value between the predicted label and the true label is calculated using the loss function, and then the gradient is calculated using the backpropagation algorithm. The model parameter best_model_wts is updated. At this point, the training of one batch of data is completed. Repeat the training process of steps 4.5, 4.6, 4.7, and 4.8 to start the next round of training.

[0044] 4.9 During training, the average prediction accuracy (mAP) of the network model when traversing the training and validation sets is visualized. The recall rate is used as the evaluation metric for object detection. F1 score evaluation index;

[0045] ;

[0046] ;

[0047] ;

[0048] in, The total number of categories, Indicates the first The average precision of the class It is a category index; This indicates the number of targets that were correctly detected. That is, the number of undetected targets;

[0049] 4.10. Based on the target detection evaluation metrics, continuously update the weight parameters when the model training results are at their best, and save the data.

[0050] Training terminates when the number of iterations reaches epochs.

[0051] As a further improvement to the present invention, step 5 specifically comprises:

[0052] 5.1 After all vibration signal samples in the training set have been traversed, import the optimal weight parameters into the model, traverse the vibration signal samples in the test set, and execute steps 4.5 to 4.7. After the traversal is complete, calculate the prediction average accuracy (mAP) and recall rate according to step 4.9. ), F1 value;

[0053] 5.2 Finally, the probability distribution of early weak faults in ball bearings is obtained through the Softmax function in the classification and recognition module. By obtaining the index of the maximum value of the probability distribution, the label corresponding to the feature can be obtained, and thus the type of early weak faults in ball bearings can be obtained, thereby achieving classification and recognition.

[0054] 5.3 Use the pyside2 module in Qt Designer to edit the interface. Its main functions include: displaying results, selecting target detection files, starting prediction, closing the page, displaying prediction accuracy, and displaying prediction time.

[0055] A button to select a target detection file is connected to the resource manager; clicking this button allows users to select a file for prediction. A "Start Prediction" button is connected to the data loader, the dual-channel densely connected convolutional network model, and the classification module. Clicking this button imports the selected file into the data loader, processes it, and then sequentially imports it into the dual-channel densely connected convolutional network model with optimal weight parameters and the classification module. A text box displaying the results is connected to the results from the classification module, providing the prediction results, prediction time, average prediction accuracy, and F1 score. A "Close Page" button is connected to a "close" command; clicking this closes the interface. This encapsulation results in the expert system application interface.

[0056] To achieve the above-mentioned technical objectives, another technical solution adopted by the present invention is as follows:

[0057] An early weak fault diagnosis system for ball bearings based on a dual-channel densely connected convolutional network includes:

[0058] The dataset processing and loading module is used to read and load CSV files containing vibration signal values ​​sampled continuously over time for early weak faults, for subsequent training and testing.

[0059] Model building module: Constructs a dual-channel densely connected convolutional network model, including a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, and a fully connected layer; the deep feature extraction module includes three dual-channel densely connected modules; the input features first pass through the shallow feature extraction module, then through the three dual-channel densely connected modules for deep feature learning, and then through the parameter reduction module before being imported into the fully connected layer to obtain the total feature vector;

[0060] The classification and recognition module is used to classify early weak faults in ball bearings based on the total feature vector output by the dual-channel densely connected convolutional network model.

[0061] The model training module is used to set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, and validate it using proportionally divided validation samples. The optimal training weights are saved for easy testing, resulting in a well-trained dual-channel densely connected convolutional network model.

[0062] The expert system module is used to set up the interface to enable automatic diagnosis of early minor faults in ball bearings and to integrate it into the expert system.

[0063] The beneficial effects of this invention are as follows:

[0064] (1) This invention optimizes the structure of DenseNet and dual-channel networks in the deep learning framework, and uses the dual-channel densely connected convolutional network model for continuous time-series acquisition of early weak bearing faults. Among them, the shallow feature extraction module is used to extract simple feature information while retaining spatial location information; the deep feature extraction module is used to extract deep semantic information, improving the training stability and convergence speed of the network; the classification module is used to intelligently distinguish the types of early weak bearing faults, achieving the functions of automatic identification and intelligent diagnosis.

[0065] (2) This invention proposes a dual-channel dense connection module, which increases the width and depth of feature transmission, allowing two channels to process different types of information simultaneously, fuse the information from the two channels, make timely responses, and improve the speed and efficiency of information processing; by reusing features, each layer can focus more on learning specific features without having to repeat learning features already learned in previous layers, thereby reducing the complexity of the model, improving the robustness of the model to noise and missing data, improving generalization ability, and enabling better adaptation and accurate prediction.

[0066] (3) This invention proposes a network structure that densely connects dual-channel dense connection modules. Dense connection can accelerate the convergence speed of the network and reduce training time; increase the fluidity of gradients and alleviate the gradient vanishing problem; and avoid the loss and decay of features during the transmission process, promote the flow and fusion of information, and improve the training efficiency and performance of the network. Attached Figure Description

[0067] Figure 1 This is a flowchart of an early weak fault diagnosis method for ball bearings based on a dual-channel densely connected convolutional network.

[0068] Figure 2 A visualization of the vibration signal dataset for early weak faults in ball bearings.

[0069] Figure 3 This is a structural diagram of a dual-channel dense connection module.

[0070] Figure 4 This is a structural diagram of a dual-channel densely connected convolutional network model.

[0071] Figure 5 This is a flowchart illustrating the training process of the dual-channel densely connected convolutional network model of this invention.

[0072] Figure 6 The diagram shows the early weak fault diagnosis results of the ball bearing after importing the dual-channel densely connected convolutional network model into the test set.

[0073] Figure 7 This is a comparison chart of the accuracy of CNN and the dual-channel densely connected convolutional network of this invention. Detailed Implementation

[0074] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0075] An early weak fault diagnosis method for ball bearings based on a dual-channel densely connected convolutional network, such as... Figure 1 As shown, it includes:

[0076] Step 1: Obtain vibration signal values ​​of the ball bearing in a time-series continuous sampling manner for early minor faults. The samples are randomly divided into training samples and test samples, and unordered validation samples are proportionally divided from the training samples. t is the unit time of the total time series.

[0077] Step 2: Construct a dual-channel dense connection module.

[0078] Step 3: Construct a dual-channel densely connected convolutional network model, including a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, and a fully connected layer. The deep feature extraction module includes three dual-channel densely connected modules. Input features are first extracted through the shallow feature extraction module, then the input feature vector is learned through the three dual-channel densely connected modules, and then imported into the fully connected layer for classification and detection after passing through the parameter reduction module, obtaining the total feature vector and forming a dual-channel densely connected convolutional network model. The classification module is used to classify the early weak faults of the ball bearing using the total feature vector output by the fully connected layer.

[0079] Step 4: Set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, realize shallow and deep fault feature learning with small sample data, and use proportionally divided validation samples for validation, save the training weights of the best parameters for easy testing, and obtain the trained dual-channel densely connected convolutional network model.

[0080] Step 5: Input the test samples into the trained dual-channel densely connected convolutional network model; set up the interface to achieve automatic identification and intelligent diagnosis of early weak faults in ball bearings, and integrate it into an expert system.

[0081] Step 1 specifically includes:

[0082] 1.1 Obtain vibration signal values ​​of all bearing faults, continuously sampled in time series. The time series of early weak faults belonging to ball bearings and continuously sampled vibration signals were screened out, including the original numerical data of the samples. The labels and data were checked to ensure that the labels and data corresponded accurately.

[0083] 1.2, such as Figure 2 As shown, each sample includes 6000 vibration signal sampling points obtained by sequential sampling in chronological order, with sampling point numbers ranging from 1 to 6000. The 6000 vibration signal sampling points of each sample are stored in a CSV file row by row, with 0, 1, 2, and 3 representing no fault, outer ring fault, inner ring fault, and ball bearing fault, respectively. A portion of the vibration signal data and labels are stored in corresponding CSV files as the training dataset. The other portion of the vibration signal data is stored unlabeled in a CSV file as the test set. The ratio of the training set to the test set is 8:2, ensuring that the numerical data and labels of each sample are completely paired and that there is no data loss or mismatch.

[0084] 1.3 Define a CSVDataset class for the early weak fault network of ball bearings. This class reads sample vibration signal numerical data, extracts relevant parameters from the labels, and arranges the data in a uniform format. For example: , The value is assigned to the category to which the label belongs. To record the true, time-sequential, continuously sampled vibration signals, the first 6000 vibration signal sampling points and the last label column are converted into tensors and set as floating-point numbers. When setting up the data loader, parameters such as batch size, instance, and whether to shuffle the data are set.

[0085] Step 2 specifically includes:

[0086] 2.1 Construct a dual-channel dense connection module, such as Figure 3 As shown, the dual-channel dense connection module includes two dual-channel convolutional layers (i.e., Figure 3 The first and second dual-channel convolutional layers are connected in series with one parameter-reducing layer and two dual-channel convolutional layers.

[0087] 2.2 A single dual-channel convolutional layer includes a left convolutional channel module and a right convolutional channel module. The left convolutional channel module includes one 3×3 convolutional layer with a stride of 1 and an inner margin of 1, one 1×1 convolutional layer with a stride of 1, one batching layer (BN), and two activation layers (ReLU). The right convolutional channel module includes one 5×5 convolutional layer with a stride of 1 and an inner margin of 2, one 1×1 convolutional layer with a stride of 1, one batching layer, and two activation layers.

[0088] After processing from the data loader, let the input feature of a single dual-channel convolutional layer be A. After A passes through the left convolutional channel module of the first dual-channel convolutional layer in the dual-channel densely connected module, the output feature is A1. After A passes through the right convolutional channel module of the first dual-channel convolutional layer, the output feature is A2. The output feature maps of the left and right convolutional channel modules maintain the same size but change the number of channels, extracting simple feature information while preserving spatial location information. Finally, A... 1、 The result of concatenating A2 and input feature A is A+ A1+ A2;

[0089] 2.3. Densely connect the two dual-channel convolutional layers, i.e., the output feature in step 2.2 is A + A1 + A2 = B. Then, after being superimposed with the input feature A, the input feature of the second dual-channel convolutional layer is A + B = C. C passes through the left convolutional channel module of the second dual-channel convolutional layer and outputs the feature C1. C passes through the right convolutional channel module of the second dual-channel convolutional layer and outputs the feature C2. Concatenate C1 and C2 to get C1 + C2 = C3. Superimpose C3 with the input feature C of the second dual-channel convolutional layer and the input feature A to get the output feature A + C + C3 = D. Change the number of channels to extract deep feature information.

[0090] 2.4 The parameter reduction layer consists of one convolutional layer, one batch normalization (BN) layer, one ReLU activation function layer, one max pooling layer, and one dropout layer. It is configured to process the latent features D of the final output in 2.3 in greater depth. It can also reduce the number of parameters that need to be processed, improve the model speed, and obtain more representative fault information. The ReLU activation function layer uses a faulty modified linear unit, and the dropout rate P of the dropout layer is set to 0.2.

[0091] Step 3 specifically includes:

[0092] 3.1 Construct a dual-channel densely connected convolutional network model, such as... Figure 4 As shown, it includes a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, and a fully connected layer;

[0093] 3.2 The shallow feature extraction module includes a 3×3 convolutional layer with a stride of 1 and an inner margin of 1, a batch processing layer (batch normalization), and a ReLU activation function layer; the feature map size remains unchanged, simple feature information is extracted, and spatial location information is preserved. After the input features are processed by the shallow feature extraction module, the output feature is X.

[0094] 3.3 The deep feature extraction module includes three dual-channel densely connected modules; and these three dual-channel densely connected modules are densely convolved; the output feature X processed by the shallow feature extraction module in step 3.2 is used as the input feature of the first dual-channel densely connected module, and the output feature after processing by the first dual-channel densely connected module is X1; since the three dual-channel densely connected modules are densely connected, the input feature of the second dual-channel densely connected module is the superposition of the output feature X1 of the first dual-channel densely connected module and the output feature X of the shallow feature extraction module, which is X + X1 = X2, thus obtaining feature X2; feature X2 After processing by the second dual-channel dense connection module, the output feature is Z. The superposition feature of Z with the input feature X2 of the second dual-channel dense connection module and the output feature X of the shallow feature extraction module is X + X2 + Z = Z1, which is the feature Z1. Z1 is used as the input feature of the third dual-channel dense connection module. The output feature after the third dual-channel dense connection module is Z2. Z2 is superimposed with the input feature Z1 of the third dual-channel dense connection module, the output feature X of the shallow feature extraction module, and the input feature X2 of the second dual-channel dense connection module to form the final output feature Z2 + Z1 + X + X2 = Y, which is the feature Y.

[0095] 3.4 After the feature Y passes through a convolutional layer, a batch normalization (BN) layer, a ReLU activation function layer, and an average pooling layer in the parameter reduction module, dimensionality reduction and downsampling are performed to preserve the features of the data, making it easier for the model to learn the essential features of the data, resulting in the pooled feature Y1.

[0096] 3.5 After the feature Y1 is flattened, it is processed by a fully connected layer, and the output feature tensor is obtained based on the weight matrix W and the bias vector b of the connected layer. The formula for a fully connected component is: , , i=1,2,…,m; where, It is the weight in the i-th row and j-th column of the weight matrix W. , The output feature tensor of the fully connected layer; such as Figure 4 As shown, the dual-channel densely connected convolutional network model is now complete.

[0097] 3.6 The classification and recognition module includes a Softmax function. After the input features pass through the dual-channel dense connection module (dense connection, pooling, and fully connected), the output feature tensor of the fully connected layer in step 3.5 is... The data is processed sequentially using the Softmax function, and the classification results are output.

[0098] The formula for the Softmax function is:

[0099] ;

[0100] In the formula, This represents the probability that the i-th type of fault sample is correctly identified. Let represent the output value of the i-th network node, and C represent the number of fault sample categories. Let O represent the output values ​​of all network nodes for the fault category of the c-th sample; the probability distribution tensor O for all fault category labels can be obtained:

[0101] ;

[0102] in, The function converts the output matrix into a probability distribution, where y is the tensor result after fully connected processing.

[0103] Step 4 specifically includes:

[0104] 4.1 Initialize the training and validation dataset split ratio η, learning rate η, and batch size S; define and initialize the accuracy and loss changes during training and validation, initialize the model weight parameters, and define the storage path for the model parameters best_model_wts when the best accuracy is achieved.

[0105] 4.2. Divide the dataset according to step 1.3. Load the continuous time series data and extract the label information. Steps 2 and 3. Build the complete deep learning algorithm and the dual-channel densely connected convolutional network model and initialize the parameter pool (used to store the information of each round of traversal).

[0106] 4.3 Determine the number of training epochs and batch size, select the Adam optimizer, and define the CrossEntropy Loss function; in each training epoch, after the training step size reaches the specified batch size, the next training epoch begins.

[0107] 4.4 In each batch of training, set up a CSVDataset data loader, select all data samples in a batch, convert the 6000 columns of feature data of each sample into tensors, and then perform standardization, normalization, randomization, and structural reconstruction; convert the corresponding label values ​​into tensors as well, and import the feature values ​​and label values ​​into the GPU for processing at the same time; perform forward propagation calculation of the model.

[0108] 4.5 Input the loaded data feature tensor into the shallow feature extraction module in step 3.2, and output the feature tensor M after convolution, normalization and activation function;

[0109] 4.6 The feature tensor M in step 4.5 is used as the input to the deep feature extraction module in step 3.3. After passing through the deep feature extraction module in step 3.3, it is then input to the parameter reduction module in step 3.4 for parameter reduction processing, and the output feature is N.

[0110] 4.7 The output feature tensor N in step 4.6 is output as a feature tensor after passing through the fully connected layer in step 3.5;

[0111] 4.8. The feature tensor output by the fully connected layer in step 4.7 is used as the output of the forward propagation calculation of the model. After being processed by the argmax function, the predicted label value Pre is obtained. The loss value between the predicted label and the true label is calculated using the loss function, and then the gradient is calculated using the backpropagation algorithm.

[0112] ;

[0113] ;

[0114] In the formula, and These are first-order momentum and second-order momentum, respectively. and Let represent the first-order momentum and second-order momentum after bias correction, respectively; ε is the hyperparameter of the network model; and η is the initial learning rate. It is the true probability distribution. It is the predicted probability distribution. The probability of occurrence, where n is the number of all possible events. The cross-entropy of the weighted summation result, For iterative updates of parameters that change over time;

[0115] Update the model's parameters best_model_wts; this completes the training of one batch of data. Repeat steps 4.5, 4.6, 4.7, and 4.8 to begin the next round of training.

[0116] 4.9 During training, the average prediction accuracy (mAP) of the network model when traversing the training and validation sets is visualized. The recall rate is used as the evaluation metric for object detection. F1 score evaluation index;

[0117] ;

[0118] ;

[0119] ;

[0120] in, The total number of categories, Indicates the first The average precision of the class It is a category index; This indicates the number of targets that were correctly detected. That is, the number of undetected targets;

[0121] 4.10. Based on the target detection evaluation metrics, continuously update the weight parameters when the model training results are at their best, and save the data.

[0122] When the iteration reaches the number of epochs, training is terminated. The overall training process is as follows: Figure 5 As shown.

[0123] Step 5 specifically includes:

[0124] 5.1 After all vibration signal samples in the training set have been traversed, import the optimal weight parameters into the model, traverse the vibration signal samples in the test set, and execute steps 4.5 to 4.7. After the traversal is complete, calculate the prediction average accuracy (mAP) and recall rate according to step 4.9. The F1 score and the average prediction accuracy of a portion of the sample data are shown below. Figure 6 As shown;

[0125] 5.2 Finally, the probability distribution of early weak faults in ball bearings is obtained through the Softmax function in the classification and recognition module. By obtaining the index of the maximum value of the probability distribution, the label corresponding to the feature can be obtained, thereby obtaining the type of early weak faults in ball bearings and realizing classification and recognition.

[0126] 5.3 Use the pyside2 module in Qt Designer to edit the interface. Its main functions include: displaying results, selecting target detection files, starting prediction, closing the page, displaying prediction accuracy, and displaying prediction time.

[0127] A button to select a target detection file is connected to the resource manager; clicking this button allows users to select a file for prediction. A "Start Prediction" button is connected to the data loader (dataset loading module), the dual-channel densely connected convolutional network model, and the classification module. Clicking this button imports the selected file into the data loader, processes it, and then sequentially imports it into the dual-channel densely connected convolutional network model with optimal weight parameters and the classification module. A text box displaying the results is connected to the results from the classification module, providing the prediction results, prediction time, average prediction accuracy, and F1 score. A "Close Page" button is connected to the "close" command; clicking this closes the interface. This encapsulation results in the expert system application interface.

[0128] Figure 7The image shows a comparison of the accuracy of existing CNNs and the dual-channel densely connected convolutional network DDBDN. The prediction accuracy of the dual-channel densely connected convolutional network DDBDN of this invention is much greater than that of CNNs.

[0129] This embodiment also provides an early weak fault diagnosis system for ball bearings based on a dual-channel densely connected convolutional network, including:

[0130] The dataset processing and loading module is used to read and load CSV files containing vibration signal values ​​sampled continuously over time for early weak faults, for subsequent training and testing.

[0131] Model building module: Constructs a dual-channel densely connected convolutional network model, including a shallow feature extraction module (extracting simple feature information, reducing data parameters, and retaining spatial location information), a deep feature extraction module (for extracting deep and complex features, improving environmental information processing and exploration capabilities), a parameter reduction module (reducing the number of parameters to be processed, improving model speed, and obtaining more representative fault information), and a fully connected layer; the deep feature extraction module includes three dual-channel densely connected modules; the input features first pass through the shallow feature extraction module, then through the three dual-channel densely connected modules for deep feature learning, and finally through the parameter reduction module before being imported into the fully connected layer to obtain the total feature vector;

[0132] The classification and recognition module is used to classify early weak faults of ball bearings based on the total feature vector output by the dual-channel densely connected convolutional network model. When the dual-channel densely connected convolutional network outputs features, it determines the type of early weak fault that the bearing is experiencing based on the probability distribution of the fault label and automatically identifies and intelligently diagnoses it.

[0133] The model training module is used to set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, and validate it using proportionally divided validation samples. The optimal training weights are saved for easy testing, resulting in a well-trained dual-channel densely connected convolutional network model.

[0134] The expert system module is used to set up the interface to achieve automatic diagnosis of early minor faults in ball bearings and integrate it into the expert system. The entire fault diagnosis module and process are packaged and integrated into the interface, which facilitates data input, data processing, and result observation, simplifies the fault diagnosis process, and makes it easier to observe the results.

[0135] The scope of protection of this invention includes, but is not limited to, the above embodiments. The scope of protection of this invention is defined by the claims. Any substitutions, modifications, or improvements to this technology that are easily conceived by those skilled in the art fall within the scope of protection of this invention.

Claims

1. A method for early weak fault diagnosis of ball bearings based on a dual-channel densely connected convolutional network, characterized in that, include: Step 1: Obtain vibration signal values ​​of ball bearings in a time-series continuous sampling, and randomly divide them into training samples and test samples, and proportionally divide the training samples into unordered verification samples. Step 2: Construct a dual-channel dense connection module. The dual-channel dense connection module includes two dual-channel convolutional layers and one parameter reduction layer connected in sequence. Each dual-channel convolutional layer includes a left convolutional channel module and a right convolutional channel module. The left convolutional channel module is used to extract the first scale feature, and the right convolutional channel module is used to extract the second scale feature. The input features of the dual-channel convolutional layer are concatenated with the first scale feature and the second scale feature along the channel dimension to obtain the output features of the dual-channel convolutional layer. Step 3: Construct a dual-channel densely connected convolutional network model, which includes a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, a fully connected layer, and a classification and recognition module; The deep feature extraction module includes three dual-channel densely connected modules connected in sequence; The input features are first processed by a shallow feature extraction module, and then by three sequentially connected dual-channel dense connection modules for deep feature learning. The output features of the deep feature extraction module are then processed by a parameter reduction module and imported into a fully connected layer to obtain the total feature vector. A classification and recognition module is used to classify early weak faults in ball bearings based on the total feature vector output by the fully connected layer; Step 4: Set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, and validate it using proportionally divided validation samples. Save the training weights of the best parameters for easy testing, and obtain the trained dual-channel densely connected convolutional network model. Step 5: Input the test samples into the trained dual-channel densely connected convolutional network model; set up the interface to achieve automatic diagnosis of early weak faults in ball bearings and integrate it into an expert system.

2. The method for early weak fault diagnosis of ball bearings based on dual-channel densely connected convolutional networks according to claim 1, characterized in that, Step 1 specifically includes: 1.1 Obtain the vibration signal values ​​of all bearing faults by continuous sampling in time series, filter out the time series and continuous sampling vibration signals of early weak faults of ball bearings, including the original sample data, check whether the labels and data correspond, and ensure that the labels and data correspond accurately. 1.2 Each sample includes 6000 vibration signal sampling points obtained by sequential sampling in chronological order, with sampling point numbers ranging from 1 to 6000. The 6000 vibration signal sampling points of each sample are stored in a CSV file row by row, with 0, 1, 2, and 3 representing no fault, outer ring fault, inner ring fault, and ball bearing fault, respectively. A portion of the vibration signal data and labels are stored in corresponding CSV files as the training dataset. The other portion of the vibration signal data is stored unlabeled in a CSV file as the test set. The ratio of training set to test set is 8:

2. The numerical data and labels of each sample are kept completely paired to ensure that there is no data loss or mismatch. 1.3 Define the CSVDataset class, a CSV data loader for the early weak fault network of ball bearings, to read sample vibration signal numerical data, extract relevant parameters from the labels, and arrange them in a uniform format: , The value is assigned to the category to which the label belongs. To record the true, time-sequential, continuously sampled vibration signals, the first 6000 vibration signal sampling points and the last label column are converted into tensors and set as floating-point numbers.

3. The method for early weak fault diagnosis of ball bearings based on dual-channel densely connected convolutional networks according to claim 1, characterized in that, Step 2 specifically includes: 2.1 Construct a dual-channel dense connection module, which includes two dual-channel convolutional layers and one parameter reduction layer, with the two dual-channel convolutional layers and the parameter reduction layer connected in series. 2.2 A single dual-channel convolutional layer includes a left convolutional channel module and a right convolutional channel module. The left convolutional channel module includes one 3×3 convolutional layer with a stride of 1 and an inner margin of 1, one 1×1 convolutional layer with a stride of 1, one batch layer, and two activation layers. The right convolutional channel module includes one 5×5 convolutional layer with a stride of 1 and an inner margin of 2, one 1×1 convolutional layer with a stride of 1, one batch layer, and two activation layers. After processing from the data loader, let the input feature of a single dual-channel convolutional layer be A. After A passes through the left convolutional channel module of the first dual-channel convolutional layer in the dual-channel densely connected module, the output feature is A1. After A passes through the right convolutional channel module of the first dual-channel convolutional layer, the output feature is A2. The output feature maps of the left and right convolutional channel modules maintain the same size but change the number of channels, extracting simple feature information while preserving spatial location information. Finally, A... 1、 The result of concatenating A2 and input feature A is B; 2.

3. Densely connect the two dual-channel convolutional layers: The output feature in step 2.2 is B. Then, B is superimposed with the input feature A to obtain the input feature C of the second dual-channel convolutional layer. C passes through the left convolutional channel module of the second dual-channel convolutional layer to output feature C1. C passes through the right convolutional channel module of the second dual-channel convolutional layer to output feature C2. Concatenate C1 and C2 to obtain C3. Superimpose C3 with the input feature C of the second dual-channel convolutional layer and the input feature A to obtain the output feature D. By changing the number of channels, deep feature information is extracted. 2.4 The parameter reduction layer consists of one convolutional layer, one batch processing layer, one activation function layer, one max pooling layer, and one dropout layer, configured to perform in-depth processing on the final output latent feature D in 2.3; the activation function layer uses Leaky ReLU with leakage, and the dropout rate P of the dropout layer is set to 0.

2.

4. The method for early weak fault diagnosis of ball bearings based on dual-channel densely connected convolutional networks according to claim 3, characterized in that, Step 3 specifically includes: 3.1 Construct a dual-channel densely connected convolutional network model, including a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, and a fully connected layer; 3.2 The shallow feature extraction module includes a 3×3 convolutional layer with a stride of 1 and an inner margin of 1, a batch processing layer, and an activation function layer; the feature map size remains unchanged, simple feature information is extracted, and spatial location information is preserved. After the input features are processed by the shallow feature extraction module, the output feature is X. 3.3 The deep feature extraction module includes three dual-channel densely connected modules; the three dual-channel densely connected modules are densely connected sequentially; the output feature X processed by the shallow feature extraction module in step 3.2 is used as the input feature of the first dual-channel densely connected module, and the output feature after processing by the first dual-channel densely connected module is X1; since the three dual-channel densely connected modules are densely connected, the input feature of the second dual-channel densely connected module is the superposition feature of the output feature X1 of the first dual-channel densely connected module and the output feature X of the shallow feature extraction module, which is X2, thus obtaining feature X2; After feature X2 is processed by the second dual-channel dense connection module, the output feature is Z. The superposition of Z with the input feature X2 of the second dual-channel dense connection module and the output feature X of the shallow feature extraction module is Z1, which is the feature Z1. Z1 is used as the input feature of the third dual-channel dense connection module. After passing through the third dual-channel dense connection module, the output feature is Z2. Z2 is superimposed with the input feature Z1 of the third dual-channel dense connection module, the output feature X of the shallow feature extraction module, and the input feature X2 of the second dual-channel dense connection module to obtain the final output feature Y, which is the feature Y. 3.4 After feature Y passes through a convolutional layer, a batch processing layer, an activation function layer, and an average pooling layer in the parameter reduction module, the pooled feature Y1 is obtained. 3.5 After the feature Y1 is flattened, it is processed by a fully connected layer, and the output feature tensor is obtained based on the weight matrix W and the bias vector b of the connected layer. The formula for a fully connected component is: , , i=1,2,…,m; where, Let j be the j-th element of the flattened eigenvector. It is the weight in the i-th row and j-th column of the weight matrix W. , The output feature tensor of the fully connected layer is used to complete the construction of a dual-channel densely connected convolutional network model. 3.6 The classification and recognition module includes a Softmax function, which converts the output feature tensor of the fully connected layer in step 3.5 into a Softmax function. The classification result is output after passing through the Softmax function; The formula for the Softmax function is: ； In the formula, This represents the probability that the i-th type of fault sample is correctly identified. Let represent the output value of the i-th network node, and C represent the number of fault sample categories. This represents the output values ​​of all network nodes for the fault category of the c-th sample; the probability distribution tensor O for all fault category labels is obtained: ； in, The function converts the output matrix into a probability distribution.

5. The method for early weak fault diagnosis of ball bearings based on dual-channel densely connected convolutional networks according to claim 4, characterized in that, Step 4 specifically includes: 4.1 Initialize the training and validation dataset split ratio η, learning rate η, and batch size S; define and initialize the accuracy and loss changes during training and validation, initialize the model weight parameters, and define the storage path for the model parameters best_model_wts when the best accuracy is achieved. 4.

2. Divide the dataset according to step 1.

3. Load the continuous time series data and extract the label information. In steps 2 and 3, build a dual-channel densely connected convolutional network model and initialize the parameter pool to store the information of each round of traversal. 4.3 Determine the number of training rounds and batch size, select the Adam optimizer, and define the cross-entropy loss function; in each training round, after the training step size reaches the specified batch size, the training ends and the next round begins; 4.4 In each batch of training, set up the CSVDataset data loader, select all data samples in a batch, convert the 6000 columns of feature data of each sample into tensors, and then perform standardization, normalization, randomization, and structure recombination; convert the corresponding label values ​​into tensors as well, and import the feature values ​​and label values ​​into the GPU for processing at the same time. Forward propagation calculation of the model; 4.5 Input the loaded data feature tensor into the shallow feature extraction module in step 3.2, and output the feature tensor M after convolution, normalization and activation function; 4.6 The feature tensor M in step 4.5 is used as the input to the deep feature extraction module in step 3.

3. After passing through the deep feature extraction module in step 3.3, it is then input to the parameter reduction module in step 3.4 for parameter reduction processing, and the output feature is N. 4.7 The output feature tensor N in step 4.6 is output as a feature tensor after passing through the fully connected layer in step 3.5; 4.

8. The feature tensor output from the fully connected layer in step 4.7 is used as the output of the model's forward propagation calculation. After processing by the Softmax and argmax functions, the predicted label value is obtained. The loss function is used to calculate the loss value between the predicted label and the true label, and then the gradient is calculated using the backpropagation algorithm. The model parameters best_model_wts are updated. At this point, the training of one batch of data is completed. Repeat the training process of steps 4.5, 4.6, 4.7, and 4.8 to start the next round of training. 4.9 During training, visualize the average accuracy (mAP) of the dual-channel densely connected convolutional network model constructed in step 3.1 after traversing the training and validation sets. The recall rate is used as the evaluation metric for object detection. F1 score evaluation index; ； ； ； in, The total number of categories, Indicates the first The average precision of the class It is a category index; This indicates the number of targets that were correctly detected. That is, the number of undetected targets; 4.

10. Based on the target detection evaluation metrics, continuously update the weight parameters when the model training results are at their best, and save the data. Training terminates when the number of iterations reaches epochs.

6. The method for early weak fault diagnosis of ball bearings based on dual-channel densely connected convolutional networks according to claim 5, characterized in that, Step 5 specifically includes: 5.1 After all vibration signal samples in the training set have been traversed, import the optimal weight parameters into the model, traverse the vibration signal samples in the test set, and execute steps 4.5 to 4.

7. After the traversal is complete, calculate the mean accuracy (mAP) and recall according to step 4.

9. F1 value; 5.2 Finally, the probability distribution of early weak faults in ball bearings is obtained through the Softmax function in the classification and recognition module. By obtaining the index of the maximum value of the probability distribution, the label corresponding to the feature is obtained, and thus the type of early weak faults in ball bearings is obtained, achieving classification and recognition. 5.

3. Use Qt Designer to edit the interface and develop functions based on the pyside2 module. The functions include: displaying results, selecting target detection files, starting prediction, closing the page, displaying prediction accuracy, and displaying prediction time. A button to select a target detection file is connected to the resource manager. Clicking the button selects a file for prediction. A button to start prediction is connected to the data loader, the dual-channel densely connected convolutional network model, and the classification module. Clicking the start prediction button imports the selected file into the data loader, processes it, and then imports it sequentially into the dual-channel densely connected convolutional network model with optimal weight parameters and the classification module. A text box displaying the results is connected to the results from the classification module, providing the prediction results, prediction time, average accuracy, and F1 score. A button to close the page is connected to the close command; clicking it closes the interface. After encapsulation, the expert system application interface is obtained.

7. An early weak fault diagnosis system for ball bearings based on a dual-channel densely connected convolutional network, characterized in that, include: The dataset processing and loading module is used to read and load CSV files containing vibration signal values ​​sampled continuously over time for early weak faults, for subsequent training and testing. The model building module is used to construct a dual-channel densely connected module, which includes two sequentially connected dual-channel convolutional layers and one parameter reduction layer. Each dual-channel convolutional layer includes a left convolutional channel module and a right convolutional channel module. The left convolutional channel module is used to extract features at the first scale, and the right convolutional channel module is used to extract features at the second scale. The input features of the dual-channel convolutional layer are concatenated with the first and second scale features along the channel dimension to obtain the output features of the dual-channel convolutional layer. The module also constructs a dual-channel densely connected convolutional network model, which includes a shallow feature extraction module, a deep feature extraction module, a parameter reduction module, a fully connected layer, and a classification and recognition module. The deep feature extraction module includes three sequentially connected dual-channel densely connected modules. The input features first pass through the shallow feature extraction module, and then through the three sequentially connected dual-channel densely connected modules for deep feature learning. The output features of the deep feature extraction module are then fed into the fully connected layer after passing through the parameter reduction module to obtain the total feature vector. The classification and recognition module is used to classify the early weak faults of the ball bearing based on the total feature vector output by the fully connected layer. The model training module is used to set the initial parameters of the dual-channel densely connected convolutional network model, train the dual-channel densely connected convolutional network model using training samples, and validate it using proportionally divided validation samples. The optimal training weights are saved for easy testing, resulting in a well-trained dual-channel densely connected convolutional network model. The expert system module is used to input test samples into a trained dual-channel densely connected convolutional network model; it is also used to set up an interface to achieve automatic diagnosis of early minor faults in ball bearings and to integrate it into the expert system.

Images