A rolling bearing fault diagnosis method based on convolution fusion of global distance matrix

Abstract

The invention provides a rolling bearing fault diagnosis method based on convolution fusion of global distance matrix, comprising the following steps: obtaining vibration data of a rolling bearing to be analyzed; converting the vibration data of the rolling bearing into a two-dimensional image of the rolling bearing to be analyzed based on an image encoding method of the global distance matrix; constructing a fault diagnosis model of an adaptive fusion multi-scale convolutional neural network for analyzing rolling bearing faults; training the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network; inputting the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of the rolling bearing fault. The method can assign learnable weights to data from different sources, and endows the network with the ability to distinguish and reasonably utilize unbalanced diagnostic information in data from different information sources.

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and relates to a rolling bearing fault diagnosis method based on convolution fusion of a global distance matrix. Background Art

[0002] Multi-sensor collaborative fault diagnosis uses different sensors to obtain monitoring data, enhances key information that can reflect the operating status of the equipment, effectively integrates information from different sources through data fusion technology, and combines deep learning models to enrich useful diagnostic information. Some existing multi-sensor collaborative fault diagnosis methods still have the following three problems:

[0003] 1) Most existing methods use time-frequency processing, simple mathematical transformation and reshaping operations to encode one-dimensional signals. There are problems such as reliance on expert experience, complex calculations and insufficient feature mapping. It does not meet the inherent requirements of fault diagnosis and causes the lack of diagnostic information.

[0004] 2) Most existing methods use more traditional convolutional neural networks as the fault diagnosis framework, which has problems such as insufficient feature extraction capabilities and a large number of parameters, resulting in insufficient utilization of diagnostic information and high computational costs.

[0005] 3) Most existing methods use simple addition or concatenation operations to fuse data from different sources, without considering the problem of imbalanced diagnostic information in data from different sensors, which greatly inhibits the improvement of diagnostic performance by multi-sensor collaborative diagnosis. Summary of the invention

[0006] According to the technical problem raised above, a rolling bearing fault diagnosis method based on convolution fusion of global distance matrix is ​​provided, which comprises the following steps: A rolling bearing fault diagnosis method based on convolution fusion of global distance matrix is ​​characterized in that it comprises the following steps:

[0007] Acquiring vibration data of the rolling bearing to be analyzed;

[0008] The image encoding method based on the global distance matrix converts the vibration data of the rolling bearing into a two-dimensional image of the rolling bearing to be analyzed;

[0009] Construct a fault diagnosis model of adaptive fusion multi-scale convolutional neural network for rolling bearing fault analysis;

[0010] The fault diagnosis model of the adaptive fusion multi-scale convolutional neural network is trained to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network;

[0011] The two-dimensional image of the rolling bearing to be analyzed is input into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of rolling bearing faults.

[0012] Further: the process of converting the vibration data of the rolling bearing into the two-dimensional image of the rolling bearing to be analyzed by the image encoding method based on the global distance matrix is ​​as follows;

[0013] First, suppose is a set of rolling bearing time series data with a length of n. Combine X with the timestamp t to get , then, Encoded as a two-dimensional matrix M, each element in M ​​is recorded as ;

[0014] (1)

[0015] (2)

[0016] Where: x i and x j In time series data, i and j The value of the moment, and is the standard deviation of X and t,

[0017] Formula (2) is a process of solving the normalized Euclidean distance, which summarizes the relationship between any two points in the original time series data dimension and the time dimension. Next, M is converted into an antisymmetric matrix ; (3)

[0018] Finally, it is mapped to the range of [0, 255], which is the encoded rolling bearing global distance matrix image GDM. The specific formula is as follows.

[0019] (4)

[0020] Further: the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network includes:

[0021] The first stage multi-scale module: used to extract the multi-scale information of the rolling bearing global distance matrix image;

[0022] Adaptive fusion module: used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, weight the data from different sources in the channel dimension, and suppress interference information while amplifying beneficial information;

[0023] The second stage multi-scale module: used to extract the fusion information of different sensors output by the adaptive fusion module;

[0024] The third stage multi-scale module is used to extract deeper abstract features output by the second stage multi-scale module;

[0025] The fourth stage multi-scale module is used to extract deeper abstract features output by the third stage multi-scale module and output pixel features of a size suitable for processing by the global average pooling layer;

[0026] Global average pooling layer: used to reduce the dimension and number of parameters of the features extracted by the multi-scale module in the fourth stage;

[0027] Fully connected layer: used to classify features output by the global average pooling layer.

[0028] Further: the first-stage multi-scale module includes two multi-convolution sub-modules arranged in parallel;

[0029] The multi-convolution submodule, the second-stage multi-scale module, the third-stage multi-scale module, and the fourth-stage multi-scale module have the same structure;

[0030] The second-stage multi-scale module includes a sequentially connected downsampling layer and a multi-scale convolution block, wherein the multi-scale convolution block is used to extract different levels of abstract features of the global distance matrix image, and the downsampling layer is used to reduce the pixel size;

[0031] The multi-scale convolution block includes different large-scale channel-by-channel convolutions and inverted residual structures, while embedding a global attention mechanism;

[0032] The different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the receptive field of the network and obtain greater spatial information;

[0033] The global attention mechanism utilizes unidirectional average pooling to aggregate position information and related information in the original time series data, and utilizes one-dimensional convolution to enhance the information extracted by the unidirectional average pooling of the position in a low dimension.

[0034] Further: the process of aggregating the position information and the related information in the original time series data by using the unidirectional average pooling, and enhancing the information extracted by the unidirectional average pooling of the position in a low dimension by using the one-dimensional convolution is as follows:

[0035] Assume that the intermediate feature to be processed is X, and the dimension of X is C × H × W , X polymerizes along the H direction, Represents a given channel c Each ofw The value on:

[0036] (5)

[0037] in, The dimension is H × W , the dimension of z is C ×1× W , let z be The aggregate value in the channel dimension represents the relationship between the time series data, which is activated after being mapped to a low dimension and expressed as :

[0038] (6)

[0039] in, is a convolutional layer Conv2d 1×1 of size 1, is the Gaussian error linear unit function;

[0040] Then, a one-dimensional convolution with a size of 3 and a step length of 1 is used to further extract abstract features. The feature extraction of the relationship between time series data is as follows:

[0041] (7)

[0042] in, It is a convolutional layer conv 1 × 1 of size 1, which is used to recover the number of channels of features; is the Gaussian error linear unit function, is the S-type activation function and g is the output of the global attention mechanism.

[0043] Further: the different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the receptive field of the network and obtain a larger spatial information process as follows:

[0044] First, two large-size channel-by-channel convolutions are used to increase the network's receptive field and obtain larger spatial information. Assuming that the intermediate feature to be processed is X, it is evenly divided in the channel dimension. The process is as follows:

[0045] (8)

[0046] Among them, X 1 and X 2 is the averaged feature, X 1 and X 2 The dimension is C / 2× H × W , and then perform multi-scale feature extraction:

[0047] (9)

[0048] (10)

[0049] in, yes The output features of yes The output features of two That is channel-by-channel convolution, S They are 7 and 11 respectively. Next, the two outputs are reconstructed in the channel dimension to obtain the output feature y. The dimension of y is C × H × W , the expression of y is as follows;

[0050] (11)

[0051] Next, we multiply y element-wise with the output g of the global attention mechanism described above;

[0052] (12)

[0053] in, is the weighted output of y, is the Gaussian error linear unit function, and finally, After mapping to the high-dimensional channel, the activation obtains the feature r, and then restores it to the original number of channels to obtain the residual feature R, increases the nonlinear ability, and adds it to the X element to obtain the output of the multi-scale convolution block. :

[0054] (13)

[0055] (14)

[0056] (15)

[0057] in, and are two Conv2d 1×1, is the Gaussian error linear unit function, is a S-type activation function, and the dimension of r is ( C ×2)× H × W , and The dimension is C × H × W .

[0058] Further: the adaptive fusion module is used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, and weight the data from different sources in the channel dimension. The process of amplifying the beneficial information while suppressing the interference information is as follows:

[0059] Assumptions , yes The body and residual of , yes The body and residual of Get the attention feature and use the Sigmoid function to map it to [0, 1] as the weight of the first stage :

[0060] (16)

[0061] Among them, GAP is the global average pooling operation, is Conv2d 1×1, is the S-type activation function;

[0062] Then, use the weight W of the first stage to 1 and X 2 Perform the first weighting and repeat the above process to obtain the weight of the second stage using the attention feature and the S-type activation function ;

[0063] (17)

[0064] Among them, GAP is the global average pooling operation, is Conv2d 1×1, is the S-type activation function; finally, using the weights of the second stage The input features are weighted twice to obtain the output F of the adaptive feature fusion module.

[0065] (18)

[0066] A readable storage medium stores a program module, wherein the program module can implement any of the methods described above when executed in a processor.

[0067] A rolling bearing fault diagnosis device based on convolution fusion of global distance matrix, comprising:

[0068] Acquisition module: used to acquire vibration data of the rolling bearing to be analyzed;

[0069] Image conversion module: used for converting the vibration data of rolling bearing into a two-dimensional image of rolling bearing to be analyzed based on the image encoding method of global distance matrix;

[0070] Building module: used to build a fault diagnosis model of adaptive fusion multi-scale convolutional neural network for rolling bearing fault analysis;

[0071] Self-training module: used to train the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network;

[0072] Fault diagnosis module: used to input the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of rolling bearing faults.

[0073] The present invention provides a rolling bearing fault diagnosis method based on convolution fusion of global distance matrix, which is designed to further improve the recognition accuracy of some existing multi-sensor collaborative fault diagnosis methods. It has the following advantages:

[0074] 1) The one-dimensional signal is encoded into a two-dimensional image using the global relationship matrix method. The proposed method is implemented using the standard Euclidean distance and does not involve complex mathematical transformations or reliance on expert experience knowledge; the global relationship matrix is ​​a position matrix that adds a time dimension, which highlights the relationship between each data in the time series data while retaining the time correlation.

[0075] 2) The backbone network of the adaptive fusion multi-scale convolutional neural network is designed using advanced convolution concepts. The convolution block of the adaptive fusion multi-scale convolutional neural network uses two different large-scale channel-by-channel convolutions, which can extract more spatial information at different scales and increase the receptive field of the network; the convolution block uses the inverted residual structure, which is a deep separable convolution as a whole. This structure balances the number of network parameters while ensuring powerful abstract feature extraction capabilities; the new global attention mechanism embedded in the convolution block enables the network to pay more attention to the location information in the image and the global correlation in the original time series data.

[0076] 3) The adaptive fusion module is used to realize the data fusion process of adaptive fusion multi-scale convolutional neural network. The adaptive fusion module is essentially an attention feature fusion method that can assign learnable weights to data from different sources; the adaptive fusion module contains two attention weighting processes, which gives the network the ability to distinguish and reasonably utilize the unbalanced diagnostic information in data from different information sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0078] Figure 1 The global distance matrix images encoded by the vibration signals of the bearing in five different health states; (a) rolling element fault; (b) inner race fault; (c) normal; (d) outer race fault; (e) mixed fault;

[0079] Figure 2 It is an adaptive fusion multi-scale convolutional neural network in the network architecture diagram;

[0080] Figure 3 It is the multi-scale convolution block in the network architecture diagram;

[0081] Figure 4 is the global attention embedded in the convolutional block in the network architecture diagram;

[0082] Figure 5 It is the adaptive fusion module in the network architecture diagram;

[0083] Figure 6 This is the module diagram of the rolling bearing fault simulation test bench;

[0084] Figure 7 It is the outer ring fault diagram in the fault category diagram;

[0085] Figure 8 It is a rolling fault diagram in the fault category diagram;

[0086] Fig. 9 It is the inner circle fault diagram in the fault category diagram;

[0087] Fig.10 It is a diagram of data processing results during the diagnosis process. DETAILED DESCRIPTION

[0088] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0089] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0090] A rolling bearing fault diagnosis method based on convolution fusion of global distance matrix includes the following steps:

[0091] S1: Acquire vibration data of the rolling bearing to be analyzed;

[0092] S2: Based on the image encoding method of the global distance matrix, the vibration data of the rolling bearing is converted into a two-dimensional image of the rolling bearing to be analyzed;

[0093] S3: constructing a fault diagnosis model of an adaptive fusion multi-scale convolutional neural network for analyzing rolling bearing faults;

[0094] S4: training the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network;

[0095] S5: Input the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of the rolling bearing fault.

[0096] Rolling bearing failures include: rolling element failure, inner ring failure, outer ring failure, and mixed failure of outer ring and rolling element;

[0097] The steps S1 / S2 / S3 / S4 / S5 are performed sequentially;

[0098] The global distance matrix is ​​a position matrix that preserves the temporal correlation while highlighting the correlation between each data point.

[0099] Figure 1 The global distance matrix images encoded by the vibration signals of the bearing in five different health states; (a) rolling element fault; (b) inner race fault; (c) normal; (d) outer race fault; (e) mixed fault;

[0100] The process of converting the vibration data of the rolling bearing into the two-dimensional image of the rolling bearing to be analyzed by the image encoding method based on the global distance matrix is ​​as follows;

[0101] In the image encoding process of the global distance matrix, first, let is a set of rolling bearing time series data with a length of n. Combine X with the timestamp t to get , which is a new two-dimensional array, then Encoded as a two-dimensional matrix M, each element in M ​​is recorded as ;

[0102] (1)

[0103] (2)

[0104] Where: x i and x j In time series data, i and j The value of the moment, and are the standard deviations of X and t, respectively,

[0105] Formula (2) is a process of solving the normalized Euclidean distance, which summarizes the relationship between any two points in the original time series data dimension and the time dimension. The purpose of dividing by the standard deviation is to eliminate the influence of different scales. Next, M is converted into an antisymmetric matrix ;

[0106] (3)

[0107] Finally, it is mapped to the range of [0, 255], which is the encoded rolling bearing global distance matrix image Global Distance Matrix (GDM). The specific formula is as follows.

[0108] (4)

[0109] Figure 2 It is an adaptive fusion multi-scale convolutional neural network in the network architecture diagram;

[0110] Figure 3 It is the multi-scale convolution block in the network architecture diagram;

[0111] Figure 4 is the global attention embedded in the convolutional block in the network architecture diagram;

[0112] Figure 5 It is the adaptive fusion module in the network architecture diagram;

[0113] Further: the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network includes:

[0114] The first stage multi-scale module: used to extract the multi-scale information of the rolling bearing global distance matrix image;

[0115] Adaptive fusion module: used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, weight the data from different sources in the channel dimension, and suppress interference information while amplifying beneficial information;

[0116] The second stage multi-scale module: used to extract the fusion information of different sensors output by the adaptive fusion module;

[0117] The third stage multi-scale module is used to extract deeper abstract features output by the second stage multi-scale module;

[0118] The fourth stage multi-scale module is used to extract deeper abstract features output by the third stage multi-scale module and output pixel features of a size suitable for processing by the global average pooling layer;

[0119] Global average pooling layer: used to reduce the dimension and number of parameters of the features extracted by the multi-scale module in the fourth stage, saving computing resources and improving computing efficiency;

[0120] Fully connected layer: used to classify features output by the global average pooling layer.

[0121] The first-stage multi-scale module includes two multi-scale sub-modules arranged in parallel;

[0122] The multi-scale submodule, the second-stage multi-scale module, the third-stage multi-scale module, and the fourth-stage multi-scale module have the same structure;

[0123] The second-stage multi-scale module includes a sequentially connected downsampling layer and a multi-scale convolution block; wherein the multi-scale convolution block is used to extract different levels of abstract features of the global distance matrix image, and the downsampling layer is used to reduce the pixel size, save computing resources and enhance the robustness of feature extraction;

[0124] The multi-scale convolution block includes different large-scale channel-by-channel convolutions and inverted residual structures, while embedding a global attention mechanism;

[0125] The different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the receptive field of the network and obtain greater spatial information;

[0126] The global attention mechanism utilizes unidirectional average pooling to aggregate position information and related information in the original time series data, and utilizes one-dimensional convolution to enhance the information extracted by the unidirectional average pooling of the position in a low dimension.

[0127] The global attention mechanism enables the network to pay more attention to the location information in the global distance matrix image and the correlation in the original time series data;

[0128] Furthermore, the process of aggregating the position information and the related information in the original time series data by using unidirectional average pooling, and enhancing the position information and the related information in a low dimension by using one-dimensional convolution is as follows:

[0129] Assume that the intermediate feature to be processed is X, and the dimension of X is C × H × W ,along H Direction aggregation, Represents a given channel c Each of w The value on:

[0130] (5)

[0131] in, The dimension is H × W , the dimension of z is C ×1× W , let z be The aggregate value in the channel dimension represents the relationship between the time series data, which is activated after being mapped to a low dimension and expressed as :

[0132] (6)

[0133] in, is a convolutional layer Conv2d 1×1 of size 1, It is the GeLU (Gaussian Error Linear Units) Gaussian error linear unit function;

[0134] Then, a one-dimensional convolution Conv1d with a size of 3 and a step length of 1 is used to further extract abstract features. This step is to extract features from the mutual relationship of time series data so that the network pays more attention to these favorable features. The details are as follows:

[0135] (7)

[0136] in, It is a convolutional layer conv 1 × 1 of size 1, which is used to recover the number of channels of features; is the GeLU function, is the Sigmoid function (S-type activation function), g is the output of the global attention mechanism;

[0137] Further: the different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the receptive field of the network and obtain a larger spatial information process as follows:

[0138] First, two large-size channel-by-channel convolutions are used to increase the network's receptive field and obtain larger spatial information. Assuming that the intermediate feature to be processed is X, it is evenly divided in the channel dimension. The process is as follows:

[0139] (8)

[0140] Among them, X 1 and X 2 is the averaged feature, X 1 and X 2 The dimension is C / 2× H × W , and then perform multi-scale feature extraction:

[0141] (9)

[0142] (10)

[0143] in, yes The output features of yes The output features of two That is channel-by-channel convolution, S are 7 and 11 respectively. Next, the two outputs are reconstructed in the channel dimension to obtain the output y, the dimension of y is C × H × W , the expression of y is as follows;

[0144] (11)

[0145] Next, we multiply y element-wise with the output g of the global attention mechanism described above;

[0146] (12)

[0147] in, is the weighted output of y, is the GeLU function. Finally, After mapping to the high-dimensional channel, the activation is restored to the original number of channels to increase the nonlinear ability, and the output of the multi-scale convolution block is obtained by adding the X element. :

[0148] (13)

[0149] (14)

[0150] (15)

[0151] in, and are two Conv2d 1×1, is the GeLU function, is the Sigmoid function, and the dimension of r is ( C ×2)× H × W , and The dimension is C × H × W .

[0152] Further: the adaptive fusion module is used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, and weight the data from different sources in the channel dimension. The process of amplifying the beneficial information while suppressing the interference information is as follows:

[0153] The adaptive fusion module processes the residual type convolution block.

[0154] Assumptions , yes The body and residual of , yes The body and residual of Get the attention feature and use the Sigmoid function to map it to [0, 1] as the weight W of the first stage:

[0155] (16)

[0156] Among them, GAP is the global average pooling operation, is Conv2d 1×1, is the Sigmoid function;

[0157] Then, use the weight W of the first stage to 1 and X 2Perform the first weighting and repeat the above process to obtain the weight of the second stage using the attention feature and Sigmoid function ;

[0158] (17)

[0159] Among them, GAP is the global average pooling operation, is Conv2d 1×1, is the Sigmoid function; finally, using the weights of the second stage The input features are weighted twice to obtain the output F of the adaptive feature fusion module.

[0160] (18)

[0161] The process of training the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain the trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network is as follows:

[0162] The two-dimensional images of rolling bearings to be analyzed are divided into training samples and test samples, with training samples accounting for 80% and test samples accounting for 20%. The fault diagnosis model of adaptive fusion multi-scale convolutional neural network is trained by gradient back propagation algorithm. When training the model, the batch size is 16, the number of iterations is set to 50, the stochastic gradient descent optimizer is used for weight update, the initial learning rate is set to 0.001, and the loss function is cross entropy loss, which is used to measure the accuracy of the model in the classification task;

[0163] The fault diagnosis performance of the fault diagnosis model is tested using test samples. When the accuracy of the test samples of the fault diagnosis model no longer improves within a number of iterations, the training stops automatically. Finally, the trained adaptive fusion multi-scale convolutional neural network can effectively identify different types of rolling bearing faults.

[0164] The adaptive fusion multi-scale convolutional neural network fault diagnosis model is trained by using a gradient back propagation algorithm, which belongs to the prior art;

[0165] A readable storage medium stores a program module, wherein the program module is executed in a processor to implement any of the methods described above.

[0166] A rolling bearing fault diagnosis device based on convolution fusion of global distance matrix, comprising:

[0167] Acquisition module: used to acquire vibration data of the rolling bearing to be analyzed;

[0168] Image conversion module: used for converting the vibration data of rolling bearing into a two-dimensional image of rolling bearing to be analyzed based on the image encoding method of global distance matrix;

[0169] Building module: used to build a fault diagnosis model of adaptive fusion multi-scale convolutional neural network for rolling bearing fault analysis;

[0170] Self-training module: used to train the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network;

[0171] Fault diagnosis module: used to input the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of rolling bearing faults.

[0172] Embodiment 1:

[0173] Experimental data introduction

[0174] By Figure 6 The rolling bearing fault simulation test bench shown in the figure collects data of the bearing in various health states to verify the effectiveness of the proposed method. The experimental platform includes a motor, a flywheel, a bearing seat, a test bearing and two vibration sensors.

[0175] Figure 7 It is the outer ring fault diagram in the fault category diagram;

[0176] Figure 8 It is a rolling fault diagram in the fault category diagram;

[0177] Fig. 9 It is the inner circle fault diagram in the fault category diagram;

[0178] Fig.10 It is a graph showing the actual results of data processing during the diagnosis process;

[0179] During the data acquisition process, two sensors collected dual-channel acceleration signals in the horizontal and vertical directions respectively, and the motor speed was 1800r / min. The experiment simulated four bearing fault modes, including three single faults and one compound fault, with a sampling frequency of 20kHz. Each signal was divided into 1000 samples, of which 80% of the random samples were used for training and 20% of the samples were used for testing, with a sample length of 1000 data points. No repeated sampling was used in the sample division process, which made the experimental results more objective. The detailed division of category labels and samples is shown in Table 1.

[0180] Table 1 Detailed classification information of category labels and samples

[0181]

[0182] 1. Comparison method

[0183] The purpose of this section is to verify the effectiveness and superiority of the proposed global relationship matrix image and adaptive fusion multi-scale convolutional neural network.

[0184] The hardware environment of the experimental verification process is Intel Core i7-12700F CPU processor, single NVIDIAGeForce RTX 3090 graphics processing unit, and Win10 operating system. During the training process of the proposed method and the comparative method, the model update times were set to 50 times and the batch processing times were set to 32. The Adam optimization algorithm was used to update the feature extractor, and the initial learning rate was set to 0.001, and then decreased by 0.5 times with 5 steps. In order to reduce the randomness of the experimental results, each group of comparative experiments was repeated 5 times during the algorithm verification.

[0185] Comparison method 1 (global relationship matrix image + ResNet18): ResNet18 is a classic method for image classification problems. We set the output of its fully connected layer to 5, the learning rate to 0.001, and decrease it with the step size.

[0186] Comparison method 2 (global relationship matrix image + ResNet50): ResNet50 is also a classic method. Similarly, the output of its fully connected layer is set to 5, the learning rate is set to 0.001, and it decreases with the step size.

[0187] Comparison method 3 (global relationship matrix image + MobileNetV3-large): MobileNetV3-large is a lightweight convolutional neural network with an SE module inserted. After modifying the number of output layers of the fully connected layer, the model parameters are about 4M.

[0188] Comparison method 4 (global relationship matrix image + MobileNetV3-small): MobileNetV3-small is also a lightweight convolutional neural network with an SE module inserted. After modifying the number of output layers of the fully connected layer, the number of model parameters is about 2M, which is similar to the number of parameters of the adaptive fusion multi-scale convolutional neural network.

[0189] Comparison method 5 (global relationship matrix image + ConvNext-tiny): ConvNext-tiny is a latest classic convolutional neural network that adopts the structure of deep separable convolution and inverted residual.

[0190] Comparative method 6 (Gram angular field image + multi-scale convolutional neural network): The signal is converted into a Gram angular field image and recognized using a multi-scale convolutional neural network.

[0191] Comparative method 7 (global relationship matrix image + multi-scale convolutional neural network-V1): Different from the multi-scale convolutional neural network, the data fusion of the two channels of the multi-scale convolutional neural network-V1 is achieved through the traditional element-by-element addition operation. The multi-scale convolutional neural network-V1 is an ablation technique of the multi-scale convolutional neural network.

[0192] Comparative method 8 (global relationship matrix image + multi-scale convolutional neural network-V2): Different from the multi-scale convolutional neural network, the data fusion of the two channels of the multi-scale convolutional neural network-V2 is achieved through traditional splicing operations. The multi-scale convolutional neural network-V2 is also an ablation technology of the multi-scale convolutional neural network.

[0193] Results Analysis

[0194] Table 2 shows the statistical results of the diagnostic accuracy of the proposed method and other comparative methods. We conducted 5 tests and used the average and maximum diagnostic accuracy to represent the classification results. It should be noted that the self-built bearing dataset has a certain degree of noise during the data collection process, which is more in line with the real industrial environment and increases the rationality of the proposed method.

[0195] 1) The global relationship matrix image can achieve good performance in different models, and the diagnostic accuracy is quite stable, which proves the effectiveness of the global relationship matrix image.

[0196] 2) Among different types of models, the performance of the multi-channel fusion model is better than that of the single-channel model. Even the multi-scale convolutional neural network-V1 and multi-scale convolutional neural network-V2 can achieve an accuracy of 98.6% and 98.2%, which fully demonstrates the necessity of multi-channel fusion in intelligent diagnosis.

[0197] 3) The diagnostic accuracy of comparison method 6 can reach 98.1%, which is significantly lower than the 99.6% achieved by the proposed method, proving that the performance of the proposed global relationship matrix image in this work is better than the existing Gram angle field image.

[0198] 4) The proposed multi-scale convolutional neural network achieved an optimal classification accuracy of 99.6%, which is significantly higher than the traditional splicing and element-wise addition fusion methods, proving the superiority of the multi-scale convolutional neural network.

[0199] Table 2 Statistics of the accuracy of different diagnostic methods

[0200]

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on convolution fusion of global distance matrix, characterized by: The following steps are involved: Acquiring vibration data of the rolling bearing to be analyzed; The image encoding method based on the global distance matrix converts the vibration data of the rolling bearing into a two-dimensional image of the rolling bearing to be analyzed; Construct a fault diagnosis model of adaptive fusion multi-scale convolutional neural network for rolling bearing fault analysis; The fault diagnosis model of the adaptive fusion multi-scale convolutional neural network is trained to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network; Input the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of the rolling bearing fault; The process of converting the vibration data of the rolling bearing into the two-dimensional image of the rolling bearing to be analyzed by the image encoding method based on the global distance matrix is ​​as follows; First, let X = [x1, x2, ..., x n ] is a set of rolling bearing time series data with a length of n. Combining X with the timestamp t, we get X′=[(x1,1),(x2,2),...,(x n ,n)], then, encode X′ into a two-dimensional matrix M, each element in M ​​is recorded as Where: x i and x j are the values ​​at time i and j in the time series data, σ x and σ t is the standard deviation of X and t, Formula (2) is a process of solving the normalized Euclidean distance, which summarizes the relationship between any two points in the original time series data dimension and the time dimension. Next, M is converted into an antisymmetric matrix M′; Finally, it is mapped to the range of [0, 255], which is the encoded rolling bearing global distance matrix image GDM. The specific formula is as follows:

2. The rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to claim 1 is characterized in that: The fault diagnosis model of the adaptive fusion multi-scale convolutional neural network includes: The first stage multi-scale module: used to extract the multi-scale information of the rolling bearing global distance matrix image; Adaptive fusion module: used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, weight the data from different sources in the channel dimension, and suppress interference information while amplifying beneficial information; The second stage multi-scale module: used to extract the fusion information of different sensors output by the adaptive fusion module; The third stage multi-scale module is used to extract deeper abstract features output by the second stage multi-scale module; The fourth stage multi-scale module is used to extract deeper abstract features output by the third stage multi-scale module and output pixel features of a size suitable for processing by the global average pooling layer; Global average pooling layer: used to reduce the dimension and number of parameters of the features extracted by the multi-scale module in the fourth stage; Fully connected layer: used to classify features output by the global average pooling layer.

3. The rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to claim 2 is characterized in that: The first-stage multi-scale module includes two multi-scale convolution sub-modules arranged in parallel; The multi-scale convolution submodule, the second-stage multi-scale module, the third-stage multi-scale module, and the fourth-stage multi-scale module have the same structure; The second-stage multi-scale module includes a sequentially connected downsampling layer and a multi-scale convolution block, wherein the multi-scale convolution block is used to extract different levels of abstract features of the global distance matrix image, and the downsampling layer is used to reduce the pixel size; The multi-scale convolution block includes different large-scale channel-by-channel convolutions and inverted residual structures, while embedding a global attention mechanism; The different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the receptive field of the network and obtain greater spatial information; The global attention mechanism utilizes unidirectional average pooling to aggregate position information and related information in the original time series data, and utilizes one-dimensional convolution to enhance the information extracted by the unidirectional average pooling of the position in a low dimension.

4. The rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to claim 3 is characterized in that: The process of using unidirectional average pooling to aggregate position information and related information in the original time series data, and using one-dimensional convolution to enhance the information extracted by the unidirectional average pooling in a low dimension is as follows: Assume that the intermediate feature to be processed is X, the dimension of X is C×H×W, X is aggregated along the H direction, and Z c w Represents the value at each w at a given channel c: Among them, x c The dimension of is H×W, the dimension of z is C×1×W, and z is denoted by Z c w The aggregate value in the channel dimension represents the relationship between the time series data, which is activated after being mapped to a low dimension and represented as f: f=σ1(F w1 (z)) (6) Among them, F w1 is a convolutional layer Conv2d 1×1 of size 1, σ1 is the Gaussian error linear unit function; Then, a one-dimensional convolution with a size of 3 and a step length of 1 is used to further extract abstract features. The feature extraction of the relationship between time series data is as follows: g=σ3(F w2 (σ2(Conv1d(f)))) (7) Among them, F w2 is a convolutional layer conv 1×1 of size 1, used to recover the number of channels of the feature; σ2 is the Gaussian error linear unit function, σ3 is the S-type activation function, and g is the output of the global attention mechanism.

5. The rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to claim 3 is characterized by: The different large-sized channel-by-channel convolutions use two large-sized channel-by-channel convolutions to increase the network's receptive field and obtain greater spatial information in the following process: First, two large-size channel-by-channel convolutions are used to increase the network's receptive field and obtain larger spatial information. Assuming that the intermediate feature to be processed is X, it is evenly divided in the channel dimension. The process is as follows: X1,X2=Split(X,dim=1) (8) Among them, X1 and X2 are the features after equal division, and the dimensions of X1 and X2 are C / 2×H×W. Then multi-scale feature extraction is performed: y1=DWconv s×s (X1) (9) y2=DWconv s×s (X2) (10) Among them, y1 is the output feature of X1, y2 is the output feature of X2, and the two DWconv s×s That is, channel-by-channel convolution, S is 7 and 11 respectively. Next, the two outputs are reconcatenated in the channel dimension to obtain the output feature y. The dimension of y is C×H×W. The expression of y is as follows; y=Concat(y1,y2,dim=C) (11) Next, we multiply y element-wise with the output g of the global attention mechanism described above; y'=0;(y☉g)(12) Among them, y′ is the weighted output of y, σ3 is the Gaussian error linear unit function, and finally, y′ is mapped to the high-dimensional channel and activated to obtain the feature r, and then restored to the original number of channels to obtain the residual feature R, increase the nonlinear ability, and add it to the X element to obtain the output X′ of the multi-scale convolution block: r=σ4(F1(y′)) (13) R=σ5(F2(r)) (14) X′=R+X (15) Among them, F1 and F2 are two Conv2d 1×1, σ4 is the Gaussian error linear unit function, σ5 is the S-type activation function, the dimension of r is (C×2)×H×W, and the dimension of R and X′ is C×H×W.

6. The rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to claim 2 is characterized by: The adaptive fusion module is used to distribute the multi-scale information output by the multi-scale module in the first stage, assign learnable weights to the data of different sensors during the fusion process, and weight the data from different sources in the channel dimension. The process of amplifying the beneficial information and suppressing the interference information is as follows: Assuming X1′ and R1′ are the body and residual of X1, and X2′ and R2′ are the body and residual of X2, first use the calculated R1′ and R2′ to obtain the attention features, and use the Sigmoid function to map them to [0,1] as the weight W of the first stage: W=σ6(F3(GAP(R1′+R2′))) (16) Among them, GAP is the global average pooling operation, F3 is Conv2d 1×1, and σ6 is the S-type activation function; Then, the weight W of the first stage is used to weight X1 and X2 for the first time, and the above process is repeated to obtain the weight W′ of the second stage using the attention feature and the S-type activation function; W′=0,(F(GAP(W☉X1+(1-W)☉X,)))(17) Among them, GAP is the global average pooling operation, F4 is Conv2d 1×1, and σ7 is the S-type activation function; finally, the input features are weighted twice using the weight W′ of the second stage to obtain the output F of the adaptive feature fusion module: FW☉+(1-W')☉X(18).

7. A readable storage medium storing a program module, characterized in that: The program module is executed in a processor to implement the method according to any one of claims 1 to 6.

8. A rolling bearing fault diagnosis device based on convolution fusion of global distance matrix. A rolling bearing fault diagnosis method based on convolution fusion of global distance matrix according to any one of claims 1 to 6, characterized in that: include: Acquisition module: used to acquire vibration data of the rolling bearing to be analyzed; Image conversion module: used for converting the vibration data of rolling bearing into a two-dimensional image of rolling bearing to be analyzed based on the image encoding method of global distance matrix; Building module: used to build a fault diagnosis model of adaptive fusion multi-scale convolutional neural network for rolling bearing fault analysis; Self-training module: used to train the fault diagnosis model of the adaptive fusion multi-scale convolutional neural network to obtain a trained fault diagnosis model of the adaptive fusion multi-scale convolutional neural network; Fault diagnosis module: used to input the two-dimensional image of the rolling bearing to be analyzed into the fault diagnosis model of the trained adaptive fusion multi-scale convolutional neural network to realize the diagnosis of rolling bearing faults.

Images