Carrying equipment bearing fault diagnosis method and device based on time-frequency dual-domain prediction

Abstract

The invention discloses a carrying equipment bearing fault diagnosis method and device based on time-frequency dual-domain prediction, and the method comprises the following steps: collecting a bearing acceleration signal, carrying out the normalization preprocessing of the bearing signal, dividing the data into a training set, a fine tuning set and a test set, carrying out the enhancement processing of the data, and carrying out the fault diagnosis of the bearing. And inputting the time domain representation and the frequency domain representation into an encoder for feature extraction, performing model pre-training according to a set predefined proxy task, constructing a target diagnosis model based on a pre-training model, performing fine tuning on the target diagnosis model by using a fine tuning set, and finally inputting a test set into the fine-tuned target diagnosis model for diagnosis. According to the method, diversified data can be generated, potential feature information in unlabeled data can be efficiently extracted, diagnosis is rapid and accurate, and bearing fault diagnosis of intelligent carrying equipment can be realized.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology for transport equipment, and in particular to a method and apparatus for bearing fault diagnosis of transport equipment based on time-frequency dual-domain prediction. Background Technology

[0002] In transport equipment, bearings, as critical components for load-bearing and support, operate under conditions of constant disturbance. They must simultaneously withstand periodic impacts caused by vehicle vibration, torque fluctuations during steering, and time-varying stresses such as transient axial / radial combined loads caused by uneven road surfaces. This environment easily induces various failure types. In addition to bearing wear, electrolytic corrosion, and overheating, the associated rotating system may also experience rotor imbalance, motor winding insulation aging, and inter-turn short circuits. Once a failure occurs, it will not only weaken the efficiency of the drive train and the stability of the system, but also significantly increase the safety risks to personnel and cargo.

[0003] Currently, deep learning algorithms are being applied to bearing fault diagnosis. However, existing deep learning methods typically require a large amount of labeled data for training, but the actual collected data is mostly unlabeled. Manual labeling is time-consuming and labor-intensive, limiting the application of most intelligent diagnostic models. Furthermore, the complexity, diversity, and nonlinear characteristics of bearing fault signals under actual operating conditions pose significant challenges to the practical application of existing intelligent fault diagnosis technologies. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for diagnosing bearing faults in transport equipment based on time-frequency dual-domain prediction, in order to solve the problem that existing intelligent diagnostic methods cannot achieve effective fault diagnosis of bearings in transport equipment in scenarios where labeled data is scarce.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: According to one aspect of the present invention, a method for fault diagnosis of bearings in transport equipment based on time-frequency dual-domain prediction is provided, comprising the following steps: S01. Data Acquisition: Acquire acceleration data of bearings under different fault conditions during operation of the transport equipment; S02, Data Processing: Normalize the acceleration data under different fault states, divide the acceleration data under each fault state into multiple sample data using a sliding window, divide the sample data into training set, fine-tuning set and test set, perform data augmentation on the data in the training set to obtain enhanced time domain representation and perform fast Fourier transform to obtain frequency domain representation. S03, Model pre-training: Input the time domain representation and frequency domain representation into the time domain encoder and frequency domain encoder respectively, extract the time domain and frequency domain features, set a predefined proxy task, and build a prediction head based on the predefined proxy task. Input the time domain and frequency domain features into the prediction head, and use the frequency domain features to predict the corresponding time domain features. S04, Model Fine-tuning: Retain the time encoder pre-trained in step S3, construct the target diagnostic model TFDDP based on the pre-trained model, and fine-tune the learned feature representation using a small amount of labeled data; S05. Model Diagnosis: Input the signals from the test set into the fine-tuned target diagnostic model TFDDP, connect the classifier head to the tail of the temporal encoder, extract the features from the test signals, and classify different types of features.

[0006] Preferably, in step S2, the acceleration data for different fault states are normalized and calculated using the following formula: in, Representing the The first of the columns 1 eigenvalue, Represents the normalized eigenvalues. Indicates the first All feature values ​​of the column.

[0007] Preferably, in step S02, the ratio of the training set, the fine-tuning set, and the test set is 177:3:20.

[0008] Preferably, in step S03, both the temporal encoder and the frequency encoder consist of a three-layer convolutional application network and a four-layer residual block network. The residual block network consists of two depthwise separable convolutional layers, two batchnorm layers, two rule function layers, and two attention modules. The attention modules include a double-pulse self-attention mechanism and a convolutional block attention module. The double-pulse self-attention mechanism includes a local attention mechanism and a global attention mechanism.

[0009] Preferably, step S03 further includes a time projection head and a frequency projection head, which are used to project the high-dimensional features extracted by the encoder to a low-dimensional subspace. The time projection head and the frequency projection head are respectively: in, Indicates a time-domain projection head; Indicates a frequency domain projection head; This indicates the output of the time-domain projection head; This indicates the output of the frequency domain projection head.

[0010] Preferably, step S03 further includes a nonlinear prediction head, which consists of two fully connected layers. The first fully connected layer of the nonlinear prediction head is used to expand the input low-dimensional features to high-dimensional features, and the second fully connected layer of the nonlinear prediction head is used to restore the high-dimensional features to the input low-dimensional features.

[0011] Preferably, in step S03, the pre-training process of the predefined agent task is supervised by a cross-correlation smoothing matrix loss function, which is: The first term represents the squared difference loss of the diagonal elements after label smoothing; the second term represents the squared difference loss of the off-diagonal elements; and the third term represents the L1 regularization loss after label smoothing. and Represents the regularization parameter; This represents the elements in the cross-correlation matrix; The dimensions represent the time-domain and frequency-domain features; i and j are the indices of the dimensions. Let C be the identity matrix assuming smoothness, and C be the cross-correlation matrix of the frequency domain projector and the time domain projector in the batch dimension.

[0012] Preferably, the pre-trained model is fine-tuned, and the specific steps are as follows: S401. Use a small amount of labeled data from the fine-tuning set to fine-tune the target diagnostic model TFDDP, and correct the mapping relationship between the features learned by the target diagnostic model TFDDP and the corresponding fault categories. S402. Retain the temporal encoder structure in the pre-training process, and transfer the potential fault feature representation and fault mapping relationship learned in the pre-trained model to the downstream diagnostic task.

[0013] Preferably, in step S5, the fault diagnosis accuracy threshold is set to 95%.

[0014] Preferably, a bearing fault diagnosis device for transport equipment based on time-frequency dual-domain prediction includes a bearing acceleration signal acquisition module, a signal preprocessing module, a data augmentation module, a model pre-training module, a model fine-tuning module, and a fault diagnosis module. The bearing acceleration signal acquisition module is used to acquire the bearing acceleration signal; the signal preprocessing module is used to normalize the acquired bearing acceleration signal and randomly divide the normalized preprocessed signal into a training set, a fine-tuning set, and a test set according to a certain proportion; the data augmentation module is used to augment the time domain data; the model pre-training module is used for model pre-training; the model fine-tuning module uses a small amount of labeled data from the fine-tuning set to fine-tune the target diagnostic model TFDDP based on the errors existing in the potential fault feature representation obtained after training in the pre-trained model; the fault diagnosis module inputs the test set data into the fine-tuned target diagnostic model TFDDP for diagnosis.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention utilizes predefined agent tasks to learn latent feature representations from a large amount of unlabeled data, which can be used for intelligent fault diagnosis of bearings in various equipment in scenarios where labeled data is scarce, greatly improving the accuracy of bearing fault diagnosis and the real-time performance of practical applications; 2. This invention uses time-domain and frequency-domain encoders to extract the time-domain and frequency-domain features of the signal respectively, and uses frequency-domain representation to predict the proxy task of time-domain representation, so that the time-domain and frequency-domain features under the same sample are closer, thereby learning the potential fault feature representation in unlabeled data in the process. 3. This invention constructs a novel dual-pulse self-attention mechanism in the residual block of the encoder and introduces an improved convolutional block attention module to better extract relevant features. At the same time, it sets up a depth-separable convolutional network, thereby significantly reducing the amount of computation.

[0016] 4. This invention incorporates a cross-correlation smoothing matrix loss function, which fully considers the noise and uncertainty that may exist in the labels, maximizes the correlation between time-domain features and frequency-domain features, and ensures the independence between different features, thereby effectively improving the generalization ability of the model.

[0017] 5. The bearing intelligent fault diagnosis device constructed in this invention has significant advantages such as high diagnostic accuracy, fast diagnostic speed and strong stability. It can meet the actual diagnostic application needs of various bearings and can also be generalized to other fields, providing a reference paradigm for diagnostic methods in related fields. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method steps of the present invention.

[0019] Figure 2 This is a schematic diagram of the workflow of the ResDSCNet encoder based on the present invention.

[0020] Figure 3This is a schematic diagram of the working process of the dual-pulse self-attention mechanism of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the working process of the improved convolutional block attention module of the present invention.

[0022] Figure 5 This is a schematic diagram of the fault diagnosis process of the present invention.

[0023] Figure 6 This is a structural block diagram of the bearing fault diagnosis device for transport equipment according to the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0025] Please see Figure 1 This invention provides a method for fault diagnosis of bearings in transport equipment based on time-frequency dual-domain prediction, comprising the following steps: S01. Use an accelerometer to simultaneously collect data on the operation of bearings with different faults in the transport equipment, where the acceleration data includes time-domain data of different bearing fault states.

[0026] Specifically, acceleration signals of seven bearing health states at different speeds were collected using a test bench for motor bearings of transport equipment. The acceleration signals included time-domain data of different bearing fault states.

[0027] S02: Normalize the acceleration data under different states, and divide the data into multiple sample sets using a sliding window. These sample sets are then divided into training, fine-tuning, and test sets. The training set consists of unlabeled data; the fine-tuning set consists of labeled data used for model fine-tuning; and the test set consists of unlabeled data used to validate the model. The unlabeled data in the training set is then subjected to Fast Fourier Transform and data augmentation to obtain the frequency domain representation of the original samples and the augmented time domain representation.

[0028] Specifically, it includes the following steps: S201. The maximum-minimum normalization method is used to normalize the acceleration signals of various bearing faults, so that the amplitude of the time-domain signal is distributed between 0 and 1. The maximum-minimum normalization method satisfies: in, Representing the The first of the columns 1 eigenvalue, Represents the normalized eigenvalues. Indicates the first All feature values ​​of the column.

[0029] S202. Number the corresponding bearing fault types. For example, the seven bearing fault types are inner ring fault, outer ring fault, rolling element fault, inner ring and outer ring fault, inner ring and rolling element fault, outer ring and rolling element fault, and inner ring, outer ring and rolling element fault. The numbers are 0, 1, 2, 3, 4, 5, and 6, respectively. Use a sliding window to segment the time series data using the normalized acceleration signal. The width of the sliding window is 1024, and the overlap rate of the sliding window is 50%. This yields samples of the bearing under different fault states.

[0030] S203. After obtaining samples for each type of fault, 2000 samples from each fault type are selected to form the total dataset for the experiment. The fault samples are then randomly divided into a training set, a fine-tuning set, and a test set in a ratio of 177:3:20. The training set contains 1770 unlabeled samples for each fault type, the fine-tuning set contains 30 labeled samples for each fault type, and the test set contains 200 unlabeled samples for each fault type.

[0031] S204. Construct a data augmentation module. The data augmentation module includes Gaussian signal-to-noise ratio noise reduction, time offset, masking, fade-in, time-domain inversion, random pruning, and gain adjustment. The acceleration signal is input into the data augmentation module to obtain the enhanced time-domain data signal. Data augmentation modules can diversify the original signal and improve the model's adaptability to data diversity and invariance. Assume the input time-domain sample is of length... signal The enhancement methods are shown in the following formulas: in, This indicates that the mean is 0 and the standard deviation is 0. Gaussian noise, SNR Indicates in The signal-to-noise ratio is randomly selected.

[0032] in, Indicates from The displacement of a randomly selected signal; This indicates a cyclic shift of the signal, meaning that when the index goes out of range, it will wrap back to the starting position of the signal.

[0033] Where s is the starting position of the mask, in Within the range; L is the length of the mask, in Randomly generated internally. ,and Indicates rounding down.

[0034] in, Indicates the fade-in factor; This indicates a single factor.

[0035] in, The length of the distorted signal is denoted as . Round the result to the nearest integer. .

[0036] in, This indicates that a random integer will be generated within this range to represent the starting position of the cut. For the clipping factor, , To round down, use... This operation is used to splice the signal, ensuring that the length of the trimmed signal is [missing information]. .

[0037] in, Indicates the gain factor. .

[0038] S205. Perform a Fast Fourier Transform on the acceleration signal to obtain its frequency domain representation. .

[0039] S03. The two representations of the two sets of data are input into the time-domain encoder and the frequency-domain encoder, respectively. Through a predefined surrogate task (predicting the time-domain representation from the frequency-domain representation), the model learns robust representations while solving the task. A novel cross-correlation loss function is designed for the model to minimize the difference between the frequency prediction and the time representation. Finally, the pre-training of the model is completed.

[0040] Specifically, the two representations of the acceleration signal are input into the encoder of TFDDP for feature extraction. The entire execution process includes the following steps: S301. Input the two representations of the two sets of data into the time-domain and frequency-domain encoders respectively for feature extraction. Use the latent representation in the frequency domain to predict the augmented time-domain representation. The prediction leverages the correlation between frequency-domain features and their corresponding augmented time-domain representations, which are independent of the time-domain representations of other samples. The encoder's main structure consists of the ResDSCNet framework, such as... Figure 2 As shown, the main structure of ResDSCNet follows the traditional convolutional neural network structure for the first three layers. After passing through these three layers, the data is input into a module consisting of four identical residual blocks. The residual block structure comprises two depthwise separable convolutional layers, two batchnorm layers, two rule functions, and two attention modules. The temporal encoder and frequency domain encoder have identical model architectures; they operate independently, differing only in their inputs and outputs. The attention modules within the residual blocks are a novel Dual Spike Self-Attention (DSSA) mechanism and an Improved Convolutional Block Attention Module (ICBAM), which decompose standard convolution into depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational cost, thus improving model efficiency.

[0041] S302, such as Figure 3 As shown, DSSA includes local attention and global attention mechanisms. By combining local and global features, it helps capture key information at different scales and suppress irrelevant features, thereby improving the model's ability to identify complex signals. The specific process is as follows: First, input features M Local features are generated using the query convolutional network Query_Conv1 and the key convolutional network Key_Conv1 within the local attention module. Q 1 and K 1. The local attention weight matrices of both are calculated through transpose and normalization operations. B 1. And normalize using the Softmax function to obtain D 1. This attention weight matrix B 1. Features generated by the weighted numerical convolutional network Value_Conv1 V 1. This allows us to obtain the output features after local attention weighting. O 1.

[0042] Meanwhile, input features M Global features are generated using the query convolutional network Query_Conv2 and the key convolutional network Key_Conv2 within the global attention module. Q 2 and K2. Using the global attention weight matrix D 2 pairs of global features V 2. Perform weighted summaries to obtain global attention output features. O 2. In DSSA, the combination of local and global features helps to capture key information at different scales and suppress irrelevant features, thereby improving the model's ability to identify complex signals.

[0043] Finally, the output features will be... O 1 and O 2. By scaling parameters respectively γ 1 and γ 2. Scale the input features and compare them with the original input features. M Add them together to obtain the final output features. M 1.

[0044] ICBAM utilizes global average pooling and global max pooling to extract global information about features. Its workflow is as follows: Figure 4 As shown, features X After passing through channel attention and spatial attention mechanisms respectively, the final output features are obtained. X 9.

[0045] This design significantly enhances the model's ability to focus on useful features and suppress irrelevant ones. The channel attention mechanism in ICBAM adaptively adjusts the weights of each channel based on global information, enabling the model to better focus on important feature channels. Specifically, global average pooling and global max pooling are used to extract global information about features, and weights for each channel are generated through fully connected layers and activation functions. Subsequently, a spatial attention mechanism further refines feature selection by calculating attention weights at each location, highlighting key spatial positions. This dual attention mechanism not only improves the quality of feature representations but also enhances the model's expressive power and robustness, resulting in better performance across various tasks.

[0046] Finally, the time-domain and frequency-domain representations of the encoder output are as follows: in, Indicates a time-domain encoder. Indicates a frequency domain encoder. This represents the output of the time-domain signal after passing through the encoder. This represents the output of the frequency domain signal after passing through the encoder.

[0047] S303. After the signal in the time and frequency domains is processed by the encoder, the extracted features are high-dimensional features, which are difficult to learn from their latent information. Therefore, a time and frequency projection head is set to project the high-dimensional features extracted by the encoder into a low-dimensional subspace, thereby obtaining low-dimensional feature vectors in the time and frequency domains, as shown in the following formula: in, Indicates the time-domain projection head. This refers to a frequency domain projection head. This indicates the output of the time-domain projection head. This indicates the output of the frequency domain projection head.

[0048] S304. After obtaining the low-dimensional feature representations in the time and frequency domains, a prediction head is constructed based on the designed predefined proxy task. Its function is to utilize the low-dimensional features of frequency. Predict the temporal characteristics of the corresponding augmented data. The nonlinear prediction head It consists of two fully connected layers. The first layer expands the input low-dimensional features into high-dimensional features. After batch normalization and ReLU activation, the second layer restores the dimension to the input dimension and finally outputs the frequency features. Through this process, it learns the latent fault feature representation in unlabeled data, which can then be transferred to downstream fault diagnosis tasks, thereby achieving effective fault identification under conditions of scarce labeled data.

[0049] S305. In the above pre-training process, the pre-training process of the predefined proxy task is supervised by the designed cross-correlation smoothing matrix loss function to measure the frequency prediction. and temporal features The correlation between them is discussed. By introducing label smoothing, the target matrix is ​​adjusted to a smooth identity matrix, preventing the model from becoming overly dependent on the training data during training, thereby reducing the risk of overfitting and improving the model's generalization ability. Furthermore, an L1 regularization term is added to the loss calculation, achieved by summing the absolute differences between the prediction matrix and the label matrix. This regularization method encourages the sparsity of model parameters, making it more adaptable to new data, while also improving the model's interpretability and aiding in feature selection. The loss function is defined as follows: In the above formula, the first term represents the squared difference loss of the diagonal elements after label smoothing, and the second term represents the squared difference loss of the off-diagonal elements, multiplied by the regularization parameter. The third term represents the L1 regularization loss after label smoothing, multiplied by the regularization parameter. . Represents the elements in the cross-correlation matrix, where, It is a constant that balances the contributions of decorrelation terms. N yes and Dimensions i and j It is a dimension index. Assuming a smooth identity matrix, C In terms of batch size and Cross-correlation matrix between them: in yes l 2-norm, ensuring and z t Within a comparable range, and C converges to N × N The identity matrix. Through the supervision of this loss function, the corresponding time-domain and frequency-domain features can be better aligned during pre-training, thereby learning the potential fault feature representations in unlabeled data to meet the needs of subsequent efficient fault diagnosis.

[0050] S04. Retain the time encoder completed in the previous pre-training stage, and build the target diagnostic model TFDDP based on the pre-trained model. In order to eliminate the deviation between the potential fault feature representation obtained in the training process and the actual fault, use a small amount of labeled data to fine-tune the learned feature representation. Specifically, in the fine-tuning phase, the pre-trained model is fine-tuned using a small amount of labeled data from the fine-tuning set, including the following steps: S401. Since there is a certain deviation between the potential fault feature representation and the actual fault during the pre-training process, a small amount of labeled data in the fine-tuning set is used to fine-tune the target diagnostic model TFDDP to correct the mapping relationship between the features learned by the target diagnostic model TFDDP and the corresponding fault categories.

[0051] S402. Subsequently, the temporal encoder structure in the pre-training process is retained, and the potential fault feature representation and fault mapping relationship learned in the pre-training model are transferred to the downstream diagnostic task, providing a foundation for effective fault identification in the subsequent testing process.

[0052] S05. Input the test set signals into the fine-tuned target diagnostic model TFDDP, connect the classifier head to the tail of the time-domain encoder, extract the features in the test signals, and classify different types of features to achieve bearing fault diagnosis.

[0053] Specifically, when training and test set samples are input into the TFDDP model for training and optimization, if the time-frequency dual-domain prediction-based fault diagnosis model for transport equipment bearings exhibits good fault diagnosis accuracy on both the training and test sets, it is used to diagnose transport equipment bearing faults. The fault diagnosis accuracy threshold is set to 95%. If the fault diagnosis accuracy of the TFDDP model is greater than 95%, the model is considered to have good fault diagnosis accuracy. To verify that the TFDDP diagnostic model can effectively identify the fault categories of transport equipment bearings, signals from the test set are input into the trained and fine-tuned TFDDP model. The model performs a series of processing steps on the input test signals, classifies different types of features, and outputs the diagnostic accuracy. If the TFDDP diagnostic accuracy on the test set is greater than 95%, it is used for fault diagnosis of transport equipment bearings.

[0054] like Figure 5 The diagram shows the flowchart of self-supervised fault diagnosis for bearings in transport equipment. First, fault acceleration data of various bearings under different operating conditions are collected on the experimental platform. Then, the time-series data is segmented using a sliding window method to obtain samples for each type of fault. A very small portion of the signals is labeled as a fine-tuning set, while the remaining signals remain unlabeled for initial signal processing. Next, during self-supervised pre-training, FFT is performed on a portion of the pre-processed time-series data to convert the unlabeled time-domain data to the frequency domain. Subsequently, a series of data augmentations are performed on the time-domain signals. At this point, frequency-domain data and augmented time-domain data are obtained. These two sets of data are input into the time-domain encoder and frequency-domain encoder, respectively. Through a predefined interface task, the model learns robust representations while solving the task and transfers the feature representations to other downstream tasks. A novel cross-correlation loss function is designed for the model to minimize the difference between frequency prediction and time representation. Finally, the time encoder pre-trained in the previous stage is retained, and the learned feature representations are finely adjusted using a small amount of labeled data. The classifier head is connected to the tail of the time-domain encoder, thereby achieving bearing fault category classification.

[0055] Table 1 shows the detailed model structure of the experiment. The original time-domain data of 1024 data points was input into the time encoder. Similarly, the time-domain data input FFT dimension was reduced to 512, thus the spectrum containing only 512 positive frequencies was input into the frequency encoder. The projection head projects the high-dimensional features of 512 dimensions into a low-dimensional subspace of 128 dimensions. The prediction head consists of two fully connected layers. The first layer expands the input low-dimensional features from 128 to 256, performs batch normalization and ReLU activation, and then the second layer restores the dimension to 128, finally outputting the frequency features.

[0056] Table 1 shows the dynamic changes in the size of the feature representation across the components of TFDDP. Furthermore, such as Figure 6 As shown, this invention also proposes a fault diagnosis device for bearings of transport equipment based on time-frequency dual-domain prediction, see [link to relevant documentation]. Figure 6 It includes a bearing acceleration signal acquisition module 601, a signal preprocessing module 602, a data augmentation module 603, a model pretraining module 604, a model fine-tuning module 605, and a fault diagnosis module 606. The bearing acceleration signal acquisition module, the signal preprocessing module, the data augmentation module, the model pretraining module, the model fine-tuning module, and the fault diagnosis module are electrically connected. The bearing acceleration signal acquisition module 601 has acceleration sensors installed on the bearing housing in the radial, vertical and axial directions to acquire the acceleration signal of the bearing. The signal preprocessing module 602 is responsible for normalizing the collected bearing acceleration signal and randomly dividing the normalized signal into training set, fine-tuning set and test set according to the proportion. The number of fine-tuning set is very small and the fine-tuning set samples are labeled data, while the training set and test set are unlabeled data. The data augmentation module 603 is composed of seven data augmentation methods and is responsible for enhancing the time domain data in multiple ways, thereby diversifying the time domain data and making the model more robust.

[0057] The model pre-training module 604 is responsible for inputting the enhanced time-domain and frequency-domain signals from the bearing training set into the feature extractor, using the frequency domain to predict the time domain according to the predefined proxy task, maximizing the correlation between time-domain features and frequency-domain features, and learning the potential fault feature representation in the signal in the process. The model fine-tuning module 605, based on the error in the potential fault feature representation obtained after training in the pre-trained model, uses a small amount of labeled data in the fine-tuning set to fine-tune the target diagnostic model TFDDP, thereby obtaining a more accurate mapping between features and faults. The fault diagnosis module 606 inputs the test set data into the finely tuned target diagnosis model TFDDP for diagnosis, thereby achieving efficient diagnosis of bearing faults in scenarios where labeled data is scarce.

[0058] Meanwhile, to verify the effectiveness of the proposed time-frequency dual-domain prediction-based bearing fault diagnosis method for transport equipment, this method is compared with other self-supervised learning methods and semi-supervised deep learning methods, including TFAI, TFPred, DDTLN, SSMN, and ISSML. A set of examples containing fault data under seven different operating conditions is used for verification; Table 2 shows the results of each method under different performance indicators.

[0059] Table 2 shows the performance of each method under different performance indicators. According to the data in Table 2, the TFDDP method outperforms other comparative methods in terms of average accuracy at different speeds. The TFDDP method achieves the highest average accuracy at 2100 rpm, reaching 99.47%, which is 2.77% higher than TFAI, 1.08% higher than TFPred, 1.93% higher than DDTLN, 1.38% higher than SSMN, and 5.00% higher than ISSMN. Examples demonstrate that the proposed time-frequency dual-domain prediction-based bearing fault diagnosis method and device for transport equipment can effectively perform bearing condition monitoring and fault diagnosis.

[0060] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of a bearing of a load carrying equipment based on time-frequency dual domain prediction, characterized in that, The method comprises the following steps: S01, data collection: collecting acceleration data of the bearing of the carrying equipment in different fault states during operation; S02, data processing: normalizing the acceleration data in different fault states, dividing the acceleration data in each fault state using a sliding window to obtain a plurality of sample data, dividing the sample data into a training set, a fine-tuning set and a test set, and performing data enhancement on the data in the training set to obtain enhanced time domain representation and performing fast Fourier transform to obtain frequency domain representation; S03, model pre-training: inputting the time domain representation and the frequency domain representation into a time domain encoder and a frequency domain encoder respectively, extracting features in the time domain and the frequency domain, setting a predefined proxy task, and constructing a prediction head according to the predefined proxy task, inputting the features in the time domain and the frequency domain into the prediction head, and predicting the corresponding time domain features using the frequency domain features; S04, model fine-tuning: retaining the time encoder pre-trained in step S3, constructing a target diagnosis model TFDDP based on the pre-trained model, and fine-tuning the learned feature representation using a small amount of labeled data; S05, model diagnosis: inputting the signals in the test set into the fine-tuned target diagnosis model TFDDP, connecting the classifier head to the tail of the time domain encoder, extracting features in the test signals, and classifying different types of features.

2. The bearing fault diagnosis method for load handling equipment based on time-frequency dual-domain prediction according to claim 1, characterized in that: In step S02, the normalization processing of the acceleration data in different fault states is calculated by the following formula: wherein, represents the first column of the first feature value, denotes the normalized feature value, denotes the first column of the feature values.

3. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S02, the division ratio of the training set, the fine-tuning set and the test set is 177:3:

20.

4. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S03, the time domain encoder and the frequency domain encoder are both composed of three layers of convolutional neural network and four layers of residual block network, the residual block network is composed of two layers of deep separable convolution layer, two layers of Batchnorm layer, two layers of Rule function layer and two layers of attention module, the attention module includes a double pulse self-attention mechanism and a convolution block attention module, the double pulse self-attention mechanism includes a local attention mechanism and a global attention mechanism.

5. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S03, it also includes a time projection head and a frequency projection head, the time projection head and the frequency projection head are used to project the high-dimensional features extracted by the encoder to a low-dimensional subspace, and the time projection head and the frequency projection head are respectively: wherein denotes a time domain projection head; denotes a frequency domain projection head; denotes an output of the time domain projection head; denotes an output of the frequency domain projection head.

6. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S03, it also includes a nonlinear prediction head, the nonlinear prediction head is composed of two layers of full connection layer, the first layer full connection layer of the nonlinear prediction head is used to expand the input low-dimensional features to high-dimensional features, and the second layer full connection layer of the nonlinear prediction head is used to restore the high-dimensional features to the input low-dimensional features.

7. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S03, the pre-training process of the predefined proxy task is supervised by a cross-correlation smoothing matrix loss function, and the cross-correlation smoothing matrix loss function is: wherein the first term represents the square difference loss of the diagonal elements after label smoothing; the second term represents the square difference loss of the non-diagonal elements; and the third term represents the L1 regularization loss after label smoothing; and denotes a regularization parameter; denotes an element in the cross-correlation matrix; denotes the dimensions of the time-domain feature and the frequency-domain feature; i and j are the indices of the dimensions; is a unit matrix assuming smoothing, and C is the cross-correlation matrix of the frequency-domain projection head and the time-domain projection head in the batch dimension.

8. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: The pre-trained model is fine-tuned, and the specific steps are as follows: S401, fine-tune the target diagnosis model TFDDP using a small amount of labeled data in the fine-tuning set, and correct the mapping relationship between the features learned by the target diagnosis model TFDDP and the corresponding fault categories; S402, the time domain encoder structure in the pre-training process is reserved, and the learned potential fault feature representation and fault mapping relationship in the pre-training model are migrated to the downstream diagnosis task.

9. The time-frequency dual-domain prediction based fault diagnosis method of load carrying equipment bearings according to claim 1, characterized in that: In step S05, the fault diagnosis accuracy threshold is set to 95%.

10. A device for bearing fault diagnosis of load-carrying equipment based on time-frequency dual-domain prediction, applied to the bearing fault diagnosis method of load-carrying equipment based on time-frequency dual-domain prediction according to any one of claims 1-9, characterized in that, The bearing acceleration signal acquisition module, the signal preprocessing module, the data enhancement module, the model pre-training module, the model fine-tuning module and the fault diagnosis module are included. The bearing acceleration signal acquisition module is used to collect the acceleration signal of the bearing; the signal preprocessing module is used to normalize and preprocess the collected bearing acceleration signal, and the normalized and preprocessed signal is randomly divided into a training set, a fine-tuning set and a test set according to a proportion; the data enhancement module is used to enhance the time domain data; the model pre-training module is used for model pre-training; the model fine-tuning module is used to fine-tune the target diagnosis model TFDDP based on the error of the potential fault feature representation obtained after the training of the pre-training model, using a small amount of labeled data in the fine-tuning set; The fault diagnosis module inputs the test set data into the fine-tuned target diagnosis model TFDDP for diagnosis.

Images