A Bearing Fault Depth Regression Method and System Based on Cross-Channel Gated and Residual Shrinkage Network

By employing cross-channel gating and residual shrinkage networks, multi-channel adaptive weighted fusion and noise-resistant feature extraction were achieved, solving the fault diagnosis problem of multi-channel heterogeneous sensor signals under strong noise conditions and improving the accuracy and robustness of bearing fault classification and damage depth estimation.

CN123069321APending Publication Date: 2026-11-10GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611267487.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing bearing fault diagnosis methods have shortcomings in adaptive fusion of multi-channel heterogeneous sensors, extraction of weak fault impact features under strong noise background, and fault classification and damage depth estimation, making it difficult to achieve effective end-to-end diagnosis.

Method used

A cross-channel gating and residual shrinkage network is adopted. The cross-channel gating module generates dynamic weights for multi-channel adaptive weighted fusion. Combined with a deep residual shrinkage network, noise is suppressed and fault features are extracted. Finally, a multi-task diagnosis module is used to achieve collaborative diagnosis of fault classification and damage degree.

Benefits of technology

It improves the feature representation capability and diagnostic stability of multi-channel signals, enhances the identification capability of weak fault features and the accuracy of damage depth estimation, and realizes end-to-end inference from the original monitoring signal to the damage degree estimate. It has good engineering practicality and generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069321A_ABST
    Figure CN123069321A_ABST
Patent Text Reader

Abstract

This invention discloses a bearing fault depth regression method and system based on cross-channel gating and residual shrinkage networks. The method includes the following steps: S1, acquiring monitoring signals synchronously collected by multi-channel sensors to obtain multi-channel time-series samples; S2, inputting the multi-channel time-series samples into a cross-channel gating module to obtain weighted multi-channel signals; S3, inputting the weighted multi-channel signals into a deep residual shrinkage network to output feature representations; S4, inputting the feature representations into a multi-task diagnostic module, which includes a classification head and multiple regression heads. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type, used to estimate the damage degree of the corresponding type. This invention achieves multi-channel adaptive weighted fusion through cross-channel gating, extracts noise-resistant fault features through a deep residual shrinkage network, and achieves collaborative diagnosis of fault classification and damage degree estimation through a multi-task diagnostic module.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis and intelligent operation and maintenance technology, specifically to a bearing fault depth regression method and system based on cross-channel gating and residual shrinkage network. Background Technology

[0002] Rotating machinery such as internal combustion engines, gas turbines, centrifugal pumps, and compressors are widely used in industries such as energy, petrochemicals, and equipment manufacturing, and are an important component of the national economy's critical infrastructure. The long-term stable operation of this equipment directly affects the continuity of production lines and the economic benefits of enterprises; sudden failures of critical components often lead to production shutdowns for maintenance, capacity losses, and even safety accidents. Core components such as bearings, gears, and rotors are prone to accelerated wear, fatigue crack propagation, and surface spalling under long-term high-speed operation, load fluctuations, and complex environments. To address these needs, developing intelligent diagnostic technologies that can sense equipment status online, detect fault signs early, and assist in decision-making has become an important research direction for the intelligent operation and maintenance of rotating machinery, and has significant engineering implications for improving the inherent safety level and operational reliability of major equipment.

[0003] Currently, fault diagnosis of rotating machinery mainly employs traditional methods that combine signal processing with expert experience. These methods rely on signal processing techniques such as time-frequency analysis, envelope spectrum analysis, and empirical mode decomposition to extract manually designed feature vectors from the raw monitoring signals. Then, they use pattern recognition models such as support vector machines and neural networks to complete fault classification. However, this process is highly dependent on prior knowledge and manual parameter tuning. The feature extraction and classification stages are independent of each other, making end-to-end joint optimization difficult. Furthermore, in situations with variable operating conditions, significant noise interference, or limited samples, the generalization performance is often limited, failing to meet the practical diagnostic needs of complex engineering scenarios.

[0004] In recent years, intelligent diagnostic methods, represented by deep neural networks, have developed rapidly. Convolutional neural networks (CNNs), in particular, can automatically learn hierarchical representations from raw signals, reducing the tedious manual feature design process and demonstrating excellent end-to-end modeling capabilities on one-dimensional time-series data such as mechanical vibration. Fault identification methods based on one-dimensional convolutional networks have gradually become one of the mainstream research approaches, providing a new technical path for directly mining fault information from raw monitoring data. However, in actual engineering sites, multiple heterogeneous sensors are typically used for synchronous acquisition. In addition to single-channel vibration signals, it is often necessary to fuse information from multiple physical quantities such as gearbox vibration, torque, speed, current, voltage, displacement, and acoustics. Most existing diagnostic frameworks based on one-dimensional convolutional neural networks use direct splicing or fixed-weight fusion methods for multi-channel signals, making it difficult to adaptively weight each sample based on the signal quality and fault sensitivity of each channel within the current time window. When some channels have high noise or sensor malfunctions, effective fault features are easily diluted by low-quality channel signals, thus affecting diagnostic accuracy.

[0005] Furthermore, existing convolutional neural networks lack targeted suppression mechanisms for background noise components during feature extraction, resulting in insufficient fidelity in preserving weak impact features under strong noise conditions. If only global average pooling is used for feature aggregation after encoding, it is difficult to fully characterize the temporal evolution of long-window signals. In fault diagnosis task design, existing methods often simplify fault classification and damage depth estimation to a single regression head or a multi-task single-head structure. The different damage formation mechanisms of inner and outer circle faults, and the lack of damage depth in normal samples, are not adequately modeled, leading to limited accuracy in fine-grained regression of damage depth.

[0006] In summary, existing bearing fault diagnosis methods still have significant shortcomings in areas such as adaptive fusion of multi-channel heterogeneous sensors, extraction of weak fault impact features under strong noise backgrounds, and joint estimation of fault classification and damage depth. There is an urgent need for a bearing fault diagnosis method that can achieve multi-channel adaptive weighted fusion, noise-resistant depth feature extraction, and joint estimation of fault identification and damage depth. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a bearing fault depth regression method based on cross-channel gating and residual shrinkage network. This method achieves multi-channel adaptive weighted fusion through cross-channel gating, extracts noise-resistant fault features through deep residual shrinkage network, and achieves collaborative diagnosis of fault classification and damage degree estimation through multi-task diagnosis module.

[0008] To achieve one of the objectives of this invention, the following solution is adopted: The bearing fault depth regression method based on cross-channel gating and residual shrinkage network includes the following steps: S1. Acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and preprocess the monitoring signals to obtain multi-channel time-series samples; S2. Input the multi-channel time series samples into the cross-channel gating module; the cross-channel gating module calculates the time-domain energy description value of each channel, constructs a time-domain energy description vector from the time-domain energy description values ​​of each channel, generates dynamic weights based on the time-domain energy description vectors, and uses the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal; S3. Input the weighted multi-channel signal into a deep residual shrinkage network. The deep residual shrinkage network includes an adaptive soft threshold shrinkage layer to suppress background noise and extract fault features, and output feature representation. S4. Input the feature representation into the multi-task diagnostic module. The multi-task diagnostic module includes a classification head and multiple regression heads. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type to estimate the damage degree of the corresponding type. Based on the identification result of the classification head, select the corresponding regression head from each regression head to output the damage degree estimate.

[0009] Further, in step S1, the preprocessing includes: time-aligning the monitoring signal, dividing it into fixed-length multi-channel time series samples using the sliding window method, and standardizing the samples to eliminate dimensional differences between different physical quantity channels.

[0010] Furthermore, in step S2, the cross-channel gating module generates dynamic weights in the following way: Calculate the time-domain energy description value for each channel separately. ,in Indicates the first The first channel in the The signal value at each time point, Indicates the length of time; The description vector is composed of the time-domain energy description values ​​of each channel. The description vector is input into a two-layer fully connected network to obtain channel scores, and then processed... Normalization yields the dynamic weights of each channel; The dynamic weights of each channel are applied to the corresponding channel at each time point in the current time series sample to obtain a weighted multi-channel signal, thereby achieving adaptive fusion of multi-channel signals.

[0011] Further, in step S3, the deep residual shrinking network includes an initial convolutional stem and multiple residual shrinking blocks, each of which embeds the adaptive soft threshold shrinking layer; The adaptive soft threshold shrinkage layer adaptively determines the threshold based on the global statistics of the current feature map and performs a soft threshold shrinkage operation on the feature map to retain the fault impact features while suppressing noise components.

[0012] Furthermore, the process between step S3 and step S4 also includes: The feature representation output from step S3 is input into a bidirectional long short-term memory network for temporal enhancement, and the output after temporal enhancement is averaged along the time dimension to obtain a window-level feature vector. The window-level feature vector is then input into the multi-task diagnostic module.

[0013] Furthermore, the multi-task diagnostic module used in step S4 is obtained through the following training process: The multi-task diagnostic module is trained using a joint loss function, which includes classification loss and regression loss. The classification loss is calculated based on the predicted probability of the state category output by the classification head and the true state category label; During the training phase, the true state category label is used as the routing basis for the regression head. When the true state category label is a fault category, the regression loss is calculated only based on the regression head output value and damage degree label corresponding to the fault category. Other regression heads do not participate in the calculation of the regression loss. When the true state category label is normal state, the normal sample does not participate in the regression loss calculation, and its corresponding regression loss term is 0.

[0014] Furthermore, in step S4, the classification loss adopts cross-entropy loss, and the regression loss adopts smoothing L1 loss; The smoothing parameter for the smoothed L1 loss is set to 0.2; The regression loss weights for different regression heads are adaptively adjusted using learnable parameters.

[0015] Furthermore, in step S4, the bearing condition category includes at least a normal condition, an inner ring fault condition, and an outer ring fault condition; Each of the described regression heads includes at least an inner-circle dedicated regression head and an outer-circle dedicated regression head; When the classification head identifies the damage as normal, the damage level output is 0; When an inner ring fault is identified, the damage level is output by calling the inner ring dedicated regression head. When an outer ring fault is identified, the damage level is output by calling the dedicated outer ring regression head.

[0016] Furthermore, in step S4, the degree of damage is the etching damage depth, which is measured in millimeters; the method employs an end-to-end joint training approach, utilizing an optimizer to train the network parameters.

[0017] The second objective of this invention is to provide a bearing fault depth regression system based on cross-channel gating and residual shrinkage network. It achieves multi-channel adaptive weighted fusion through cross-channel gating, extracts noise-resistant fault features through deep residual shrinkage network, and achieves collaborative diagnosis of fault classification and damage degree estimation through multi-task diagnosis module.

[0018] To achieve the second objective of this invention, the following solution is adopted: A bearing fault depth regression system based on cross-channel gating and residual shrinkage networks includes: The data acquisition and preprocessing module is used to acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and to preprocess the monitoring signals to obtain multi-channel time-series samples. A cross-channel gating module is used to receive the multi-channel time-series samples, calculate the time-domain energy description value of each channel, construct a time-domain energy description vector from the time-domain energy description values ​​of each channel, generate dynamic weights based on the time-domain energy description vectors, and use the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal. The deep residual shrinking network encoder module includes a deep residual shrinking network, which contains an adaptive soft threshold shrinking layer for receiving the weighted multi-channel signal, suppressing background noise and extracting fault features, and outputting feature representations. The multi-task diagnostic module includes a classification head and multiple regression heads for receiving the feature representation. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type for estimating the damage degree of the corresponding type. Based on the identification result of the classification head, the module selects the corresponding regression head from the regression heads to output the damage degree estimate.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention generates dynamic weights through a cross-channel gating module and performs weighted fusion of signals from each channel, achieving adaptive fusion of information from multiple heterogeneous sensors and improving the efficiency of fault information utilization. Traditional methods use direct splicing or fixed-weight fusion of multi-channel signals, making it difficult to adaptively weight each sample based on the signal quality and fault sensitivity of each channel; when some channels have high noise or sensor anomalies, effective fault features are easily diluted by low-quality channel signals. To address these issues, this invention generates dynamic weights in the cross-channel gating module based on the time-domain energy description vector composed of the time-domain energy description values ​​of each channel. These dynamic weights are then used to weight and fuse signals from each channel, enabling adaptive adjustment of the contribution of each channel based on the current time-series sample, thus improving feature representation and diagnostic stability under multi-physical quantity monitoring conditions. Experimental results show that after removing the cross-channel gating module, the root mean square error (RMSE) of the fault sample damage depth regression increased from 0.01214 mm to 0.03171 mm, indicating that this invention can significantly improve the accuracy of damage degree estimation.

[0020] 2. This invention suppresses background noise and extracts fault features through an adaptive soft-thresholding contraction layer in a deep residual contraction network, achieving noise-resistant feature extraction and improving the identification capability of weak fault features. Conventional convolutional residual networks lack targeted suppression of background noise components during feature extraction, resulting in insufficient fidelity preservation of weak impact features under strong noise conditions. In contrast, this invention uses a deep residual contraction network as the backbone of feature extraction. This network adaptively determines the threshold based on feature map statistics, performing a soft-thresholding contraction operation on the feature map, suppressing noise components while retaining effective feature information such as fault impact. Experimental results show that after removing the deep residual contraction module, the RMSE of fault samples increases to 0.13165 mm, indicating that this invention effectively improves diagnostic robustness under strong noise conditions.

[0021] 3. This invention utilizes a multi-task diagnostic module comprised of a classification head and multiple dedicated regression heads. Based on the classification results, it selects the corresponding regression head to output a damage degree estimate, achieving collaborative diagnosis of fault classification and damage degree estimation, thus improving the accuracy of damage depth estimation. Existing methods often simplify fault classification and damage depth estimation to a single regression head or a multi-task single-head structure, failing to adequately model the different damage formation mechanisms of inner and outer ring faults, and the lack of damage depth in normal samples. To address these issues, this invention constructs a cascaded multi-task diagnostic framework that first classifies and then performs branched regression. The classification head identifies the bearing state category, and each regression head estimates the damage degree corresponding to a different fault type. Inner and outer ring faults are modeled separately by dedicated regression heads, while normal samples do not participate in regression. Experimental results show that when using a single-head regression structure, the RMSE of fault samples increases to 0.01709 mm, indicating that this invention effectively reduces the error caused by task coupling and significantly improves the accuracy of damage depth prediction.

[0022] 4. This invention constitutes a complete technical chain from data acquisition, multi-channel adaptive weighted fusion, noise-resistant feature extraction to split-path regression, realizing direct processing and diagnostic result output of multi-channel raw sliding window signals without the need for cumbersome manual feature engineering. This invention does not rely on traditional signal processing methods such as time-frequency analysis, envelope spectrum, and empirical mode decomposition for manual feature design, reducing dependence on manual feature engineering and expert experience. It achieves end-to-end inference from raw monitoring signals to damage level estimates, demonstrating good engineering practicality and generalization ability. Attached Figure Description

[0023] Figure 1 This is a flowchart of the bearing fault depth regression method based on cross-channel gating and residual shrinkage network in an embodiment of the present invention; Figure 2 This is a schematic diagram of the cross-channel reliability gating module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Deep Residual Shrinking Network (DRSN) in an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall structure of the bearing fault diagnosis framework based on cross-channel gating and deep residual shrinkage network in an embodiment of the present invention. Figure 5 This is a block diagram of the bearing fault depth regression system based on cross-channel gating and residual shrinkage network in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0025] Example 1 like Figure 1-4 As shown, this embodiment of the invention provides a bearing fault depth regression method based on cross-channel gating and residual shrinkage network. It can simultaneously classify bearings as normal / inner ring fault / outer ring fault, and quantitatively estimate the etching damage depth of the inner and outer rings under the condition of synchronous acquisition by multiple heterogeneous sensors.

[0026] The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to embodiments of the present invention includes the following steps: S1. Acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and preprocess the monitoring signals to obtain multi-channel time-series samples.

[0027] S2. Input the multi-channel time-series samples into the cross-channel gating module; the cross-channel gating module calculates the time-domain energy description value of each channel, constructs a time-domain energy description vector from the time-domain energy description values ​​of each channel, generates dynamic weights based on the time-domain energy description vectors, and uses the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal.

[0028] S3. Input the weighted multi-channel signal into a deep residual shrinkage network. The deep residual shrinkage network includes an adaptive soft threshold shrinkage layer to suppress background noise and extract fault features, and output feature representation.

[0029] S4. Input the feature representation into the multi-task diagnostic module. The multi-task diagnostic module includes a classification head and multiple regression heads. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type to estimate the damage degree of the corresponding type. Based on the identification result of the classification head, select the corresponding regression head from each regression head to output the damage degree estimate.

[0030] The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to the present invention will be further described in detail below.

[0031] To address the challenges of adaptive fusion of signals from multi-channel heterogeneous sensors, insufficient feature extraction under noise interference, and the difficulty of accurately estimating the damage depth of the inner and outer races using single-head multi-task regression, this invention designs a cross-channel reliability gating module. This module dynamically weights multi-channel signals sample-by-sample based on the temporal energy description of each channel. To achieve robust extraction of weak fault impacts, this invention employs a deep residual shrinking network, embedding an adaptive soft-threshold shrinking layer within the residual block to suppress background noise components while preserving fault features. To enhance the dynamic modeling capability of long-window time-series signals, this invention introduces a bidirectional long short-term memory network (BiLSTM) after the encoder for temporal enhancement. To achieve joint diagnosis consistent with the physical mechanism of bearing faults, this invention constructs a cascaded multi-task head that first classifies (inner race fault, outer race fault, normal) and then performs branched regression (inner race fault, outer race fault). Normal bearing samples do not participate in deep regression training.

[0032] In this embodiment, the experimental dataset is divided into two types: datasets from self-developed experimental devices and public datasets.

[0033] Specifically, the experimental data acquisition device (Private Dataset, PRI) collects vibration signals from rotating machinery using seven sensors: two accelerometers, one torque / speed sensor, a current sensor, a voltage sensor, a displacement sensor, and an acoustic sensor. These sensors capture vibration signals from eight different types of rotating machinery (torque and speed are measured using the same sensor). The experimental conditions include three system states: Normal, Inner etching fault, and Outer etching fault. The etching damage depth label for fault samples is a physical dimension in millimeters (mm), with ten discrete increments: 0.2mm, 0.4mm, 0.6mm, 0.8mm, 1.0mm, 1.2mm, 1.4mm, 1.6mm, 1.8mm, and 2.0mm. The damage depth for normal samples is 0. After data acquisition, the signals from each channel are time-aligned, and a sliding window method is used to divide them into multi-channel time-series samples of fixed length. In this embodiment, the window length W = 4096 sampling points, and the sliding step size S = 2048 sampling points. Data was partitioned according to the experimental file number: experimental data with filenames ending in "_1" were used for training, and experimental data with filenames ending in "_2" were used for testing and verification, in order to avoid data leakage from the same experiment.

[0034] Specifically, there are three publicly available datasets: the Shandong University of Science and Technology (SDUST) rotating machinery fault dataset, the Mendeley three-phase motor bearing dataset, and the Case Western Reserve University (CWRU) bearing dataset. The SDUST rotating machinery dataset includes three discrete levels: normal operation and faults with inner and outer ring diameters of 0.2 mm, 0.4 mm, and 0.6 mm. The Mendeley three-phase motor bearing dataset includes six discrete levels: normal operation and faults with inner and outer ring diameters of 0.7 mm, 0.9 mm, 1.1 mm, 1.3 mm, 1.5 mm, and 1.7 mm. The Case Western Reserve University bearing dataset includes three discrete levels: normal operation and faults with inner and outer ring diameters of 0.007 inches, 0.014 inches, and 0.021 inches. Since the damage unit in this experiment is mm, these values ​​are converted to three discrete levels: 0.1778 mm, 0.3556 mm, and 0.5334 mm. The aforementioned public datasets differ from the self-built PRI in terms of sensor type, number of channels, and fault category definition. Therefore, in this embodiment of the invention, model training and performance evaluation are completed using its eight-channel synchronously acquired data. On the above public datasets, pre-trained models are performed to verify the cross-dataset generalization ability.

[0035] In the experiments of this invention, the number of test set samples varied due to differences in datasets. The main experiment used a self-built PRI full set sliding window sample for end-to-end training, rather than a small sample setting. The training set contained 15,486 samples, and the validation and test sets each contained 7,728 samples. The Shandong University of Science and Technology dataset had 6,454 samples in the training set, 2,149 samples in the validation set, and 2,156 samples in the test set, totaling 10,759 samples; the Mendeley three-phase motor bearing dataset had 844 samples in the training set, 422 samples in the validation set, and 422 samples in the test set, totaling 1,684 samples; and the Case Western Reserve University bearing dataset had 1,517 samples in the training set, 232 samples in the validation set, and 468 samples in the test set, totaling 2,217 samples.

[0036] In this embodiment, after time alignment of the original monitoring signals of each channel, the sliding window method is used to divide them into fixed-length multi-channel time series samples; the mean and standard deviation of each channel are estimated on the training set, and the training set, validation set and test set are Z-score standardized to eliminate the difference in the dimensions of different physical quantity channels.

[0037] In this embodiment, no separate filtering and noise reduction module is set at the input end; the suppression of background noise is achieved in the deep residual shrinking network encoder through the adaptive soft threshold shrinking layer in the residual block.

[0038] Optionally, a small amount of Gaussian noise can be applied to the input samples during the training phase to enhance the model's noise-resistant generalization ability.

[0039] The damage depth label uses the physical quantity (mm) of the etching size; the damage degree index can be normalized to 0 to 1 according to the maximum fault size for auxiliary evaluation.

[0040] In this embodiment, the specific contents of the cross-channel reliability gating module are as follows: The cross-channel reliability gating module proposed in this invention, such as Figure 2 As shown, the calculation process is as follows.

[0041] definition To input multi-channel signals, For batch size, For the number of channels, For time length; defined This is the gated weighted signal. The time-domain energy description of each channel is as follows: Descriptor vector The channel score is obtained by inputting a two-layer fully connected network. : The channel weights are then obtained by applying the normalized exponential function (Softmax): Weighting of signals from each channel: Its weighted multi-channel signal The data is then fed into a subsequent deep residual shrinking network encoder. The channel weights change dynamically with each time window sample, achieving adaptive fusion of multiple physical quantity channels.

[0042] In this embodiment, the specific details of the deep residual shrinkage network encoder are as follows: Soft-threshold shrinkage layers in Deep Residual Shrinkage Networks (DRSNs) Figure 3 As shown, it includes the initial convolution stem and multiple residual shrinkage blocks.

[0043] Stem The following steps are performed sequentially: one-dimensional convolution (Conv1d, kernel size 15) → batch normalization → corrected linear units ( ) Activate → Max Pooling (MaxPool1d, pooling kernel size 4).

[0044] Each residual shrinking block includes: Conv1d→BN→ → Soft threshold shrinkage layer (Shrinkage1d) → Conv1d → BN → Added with shortcut branch → .

[0045] For feature maps Calculate adaptive thresholds by channel And perform soft threshold shrinkage: Where GAP stands for Global Average Pooling, FC stands for Fully Connected Subnetwork, and σ stands for Sigmoid function.

[0046] After element-wise addition of the main branch and shortcut branch of each residual shrinking block, ... activation: The calculation method for one-dimensional convolution is as follows: in Indicates the first Layer The first channel 1 node For activation function, For the first Number of channels per layer For input channels, The location of the convolution kernel. The kernel size is [size]. Step size, The fill size is used. Pooling is calculated as follows: in For the pooling process 1 node For pooling step size, This is the pooled window size.

[0047] In this embodiment, the specific content of the bidirectional long short-term memory network temporal enhancement module is as follows: Will After transposition, the input is a bidirectional long short-term memory network ( Perform timing enhancement: in The hidden layer dimension in each direction is 128, and the output dimension after bidirectional concatenation is 256. After applying dropout (dropout rate of 0.2) to the output, the arithmetic mean is taken along the time dimension to obtain the batch window-level feature matrix: In this embodiment, the specific contents of the multi-task diagnostic head are as follows: The classification and regression calculations for cascaded multi-task heads are as follows: The classification probability is: The bearing condition category is predicted during the inference phase as follows: The final damage depth estimate is synthesized by the routing function: Take the real labels during the training phase As a basis for routing (mandated by the teacher); taken during the reasoning stage As a basis for routing.

[0048] In this embodiment, the training method is as follows: end-to-end joint training is performed on a self-built eight-channel training set, using the Adam optimizer; test samples are evaluated in batches.

[0049] The joint loss function is: The classification loss is the cross-entropy loss function: The regression loss is a smooth L1 loss (β=0.2), and it is calculated only for the corresponding regression heads on the inner / outer circle samples: when or When the set is empty, the corresponding regression loss is 0. Normal samples are not included in the regression loss. The weights of the inner and outer regression loss can be learned adaptively. Learning rate: 0.001.

[0050] This invention designs a multi-physical quantity adaptive fusion method for cross-channel reliability gating. Based on the temporal energy description of each channel, channel weights are obtained through mapping via a two-layer fully connected network and Softmax normalization. The eight-channel sliding window signals are then dynamically weighted and fused sample by sample. The weights change with the samples, suppressing low-reliability channels and highlighting fault-sensitive channels when sensor quality is uneven or a single channel is abnormal. This invention employs an adaptive soft-threshold noise-resistant feature extraction method using a deep residual shrinking network. A soft-threshold shrinking layer is embedded in a one-dimensional residual shrinking block, and the threshold is adaptively determined by the global statistics of the absolute value of the feature map and a small sub-network. This method suppresses background noise components while retaining weak fault impact features, improving diagnostic robustness under strong noise conditions. This invention constructs a multi-task damage depth diagnosis framework that first classifies and then performs multi-path regression. It constructs a classification head and inner and outer dedicated regression heads: during the training phase, the teacher forces routing with the real class labels, and the inner / outer circle samples calculate the loss for their respective dedicated regression heads. Normal samples do not participate in damage depth regression. During the inference phase, the routing is based on the classification results. The damage depth output for normal samples is 0, and the inner / outer circle is estimated by the inner circle head and the outer circle head, respectively.

[0051] The bearing fault depth regression method based on cross-channel gating and residual shrinkage network of the present invention will be described below through specific applications.

[0052] In this embodiment, the eight-channel dataset (PRI) generated by the self-developed rotating machinery fault diagnosis test bench is mainly used as the core validation data to verify the performance of the present invention in jointly estimating bearing condition category identification and etching damage depth (mm) under the condition of simultaneous acquisition of multiple physical quantities. Independent tests were also conducted on the Shandong University of Science and Technology (SDUST) rotating machinery fault dataset, the Mendeley three-phase motor bearing fault dataset (triaxial vibration), and the Case Western Reserve University (CWRU) bearing fault dataset to verify the generalization ability of the fault diagnosis framework of the present invention under different numbers of acquisition channels, different operating condition divisions, and different damage annotation forms. All embodiments uniformly use damage depth regression performance as the main evaluation index, including: root mean square error of damage depth of fault samples (RMSE, mm), mean absolute error of fault samples (MAE, mm), and MAE (mm) of fault location for inner and outer rings.

[0053] Table 1 Experimental Results Continued from the table above The following verification of the invention is based on the dataset (PRI) from the self-developed eight-channel bearing fault experimental platform: This embodiment uses the self-developed PRI dataset as the primary verification object. The test bench simultaneously acquires eight channels of signals: bearing vibration, gearbox vibration, torque, speed, current, voltage, displacement, and audio. Operating conditions include normal, inner ring and outer ring etching faults. Damage depth labels are etched dimensions (mm), with fault levels ranging from 0.2 to 2.0 mm and increments of 0.2 mm. Data is divided according to experimental batches (…). train, (Test), sliding window length 4096, step size 2048.

[0054] End-to-end joint training was performed using the complete method of this invention. The experimental results are shown in Table 1, No. 1: The RMSE of the fault sample was 0.01214 mm, the MAE was 0.00644 mm, and the MSE was 0.000147 mm. 2 The inner ring MAE is 0.00452mm, and the outer ring MAE is 0.00836mm.

[0055] The method of this invention was compared with traditional two-stage methods, namely 1D-CNN single-head, CNN-LSTM single-head, and SVM / SVR combined with handcrafted features. The results are shown in Table 1, numbers 2-4. It can be seen that the method of this invention has the lowest RMSE and MAE for fault samples, indicating that on the eight-channel PRI dataset, this invention outperforms deep network comparisons and traditional machine learning methods in terms of damage depth estimation accuracy.

[0056] To further verify the contributions of each key module, ablation experiments were conducted, and the results are shown in Table 1, numbers 17-20. Removing the cross-channel gating improved the fault RMSE to 0.03171 mm; removing the DRSN soft threshold shrinkage improved it to 0.13165 mm; removing the cascaded dual regression head (replacing it with a single-head multi-tasking module) improved it to 0.01709 mm; and removing the BiLSTM timing enhancement improved it to 0.02169 mm. These results indicate that cross-channel gating, the deep residual shrinkage network, the cascaded branch regression, and the timing enhancement module all substantially contribute to the accuracy of the final damage depth estimation.

[0057] The following verification of the invention is based on the Shandong University of Science and Technology Rotating Machinery Fault Dataset (SDUST):

[0058] This embodiment verifies the applicability of the present invention in a single-channel vibration scenario based on the SDUST rotating machinery fault public dataset. The input is a single-channel vibration signal, with a sliding window length of 4096 and a step size of 4096; the remaining modules of the network and the training strategy are the same as in Embodiment 1, only the number of input channels is adjusted according to the dataset.

[0059] The experimental results are shown in Table 1, numbers 5-8. For the complete method of this invention (number 5), the fault sample RMSE is 0.00610 mm, MAE is 0.00422 mm, and MSE is 0.000037 mm.2 The MAE of the inner circle is 0.00523 mm, and the MAE of the outer circle is 0.00321 mm. Compared with 1D-CNN, CNN-LSTM, and feature + SVM / SVR in items 6-8, the method of this invention has the lowest error in all aspects, indicating that the present invention has good damage depth estimation performance and stability on the SDUST dataset, and has a certain generalization ability on single-channel public data.

[0060] The following verification of the invention is based on the Mendeley three-phase motor bearing fault dataset: This embodiment is validated based on the Mendeley three-phase motor bearing fault dataset. This dataset provides X / Y / Z triaxial vibration signals with three input channels; a sliding window length of 2048 and a step size of 1024; and a fault damage depth of approximately 0.7–1.7 mm. Data partitioning employs a batch file strategy. For training, Used for verification and testing.

[0061] The experimental results are shown in Table 1, numbers 9-12. For the complete method of this invention (number 9), the fault sample RMSE is 0.11273 mm, MAE is 0.06392 mm, and MSE is 0.012708 mm. 2 The MAE of the inner ring was 0.04640 mm, and the MAE of the outer ring was 0.08234 mm. Compared with the controls in series 10-12, the method of this invention had the lowest RMSE and MAE. This dataset has a small sample size and relatively complex operating conditions, resulting in an overall higher regression error than PRI and CWRU. However, the method of this invention is still significantly better than the three controls, indicating that this invention can still effectively perform quantitative estimation of damage depth in multi-axis vibration scenarios of three-phase motors.

[0062] The invention is validated below using a bearing failure dataset from Case Western Reserve University (CWRU): This embodiment is validated based on the CWRU classic bearing fault dataset. A 12kHz single-channel vibration acceleration at the drive end is used, with a sliding window length of 4096 and a step size of 2048; the fault diameters of the inner and outer rings are 0.007, 0.014, and 0.021 inches (approximately 0.1778, 0.3556, and 0.5334 mm), respectively.

[0063] The experimental results are shown in Table 1, numbers 13-16. For the complete method of this invention (number 13), the fault sample RMSE is 0.00626 mm, MAE is 0.00409 mm, and MSE is 0.000039 mm. 2The inner circle MAE was 0.00168 mm, and the outer circle MAE was 0.00650 mm, the lowest among the four methods. The RMSE of feature number 16 + SVM / SVR control was 0.00948 mm, also achieving a good result, but still higher than the method of this invention. The results show that this invention achieves excellent damage depth estimation accuracy on the widely used benchmark dataset CWRU, further validating the universality and stability of the method.

[0064] Advantages of the bearing fault depth regression method based on cross-channel gating and residual shrinkage network in this invention: 1. This invention achieves adaptive fusion of information from multiple heterogeneous sensors, improving the efficiency of fault information utilization. This invention proposes a cross-channel reliability gating module, which calculates the time-domain energy description value of each sensor channel separately, constructs a time-domain energy description vector from these values, and dynamically generates channel weights using a gating network, achieving sample-by-sample adaptive weighted fusion of multi-channel signals. Compared with traditional fixed-weight fusion or direct splicing methods, this method can automatically enhance the contribution of fault-sensitive channels based on the current sample state, suppress interference from low-quality and abnormal channels, and improve feature representation ability and diagnostic stability under multi-physical quantity monitoring conditions. Experimental results show that after removing the cross-channel gating module, the root mean square error (RMSE) of the fault sample damage depth regression increases from 0.01214 mm to 0.03171 mm, indicating that this module can significantly improve the accuracy of damage degree estimation.

[0065] 2. This invention achieves noise-resistant feature extraction, improving the ability to identify weak fault features. This invention employs a deep residual shrinking network as a feature encoder, introducing an adaptive soft threshold shrinking mechanism into the residual structure. This mechanism automatically determines the threshold size based on feature map statistics, suppressing noise components while retaining effective feature information such as fault impact. Compared to traditional convolutional networks that rely solely on convolutional kernels to learn features, this invention exhibits better robustness and generalization ability under strong noise conditions. Experimental results show that after removing the deep residual shrinking module, the RMSE of fault samples increases to 0.13165 mm, indicating that this module plays a crucial role in improving regression accuracy.

[0066] 3. This invention achieves collaborative diagnosis of classification and regression that conforms to the fault mechanism, improving the accuracy of damage depth estimation. This invention constructs a cascaded multi-task diagnostic framework that first classifies and then performs branched regression. First, the bearing condition is classified as normal, inner ring fault, or outer ring fault. Then, based on the classification results, corresponding dedicated regression heads are called to estimate the damage depth. Since inner ring faults and outer ring faults have different formation mechanisms and signal characteristics, using a dedicated regression model can effectively reduce the error caused by task coupling. Simultaneously, normal samples do not participate in the regression loss calculation, avoiding interference from invalid samples in the regression model training. Experimental results show that when using a single-head regression structure, the RMSE of fault samples increases to 0.01709 mm, indicating that the cascaded branched regression mechanism proposed in this invention can significantly improve the accuracy of damage depth prediction.

[0067] 4. This invention establishes a multi-module collaborative diagnostic mechanism to achieve integrated diagnosis of fault identification and damage depth estimation. The cross-channel gating module first dynamically weights and fuses multi-source heterogeneous information based on the time-domain energy description vector; the depth residual shrinkage module further suppresses random noise in the gated signal; the temporal enhancement module learns long-term dependencies using the gating and denoised features; and the classification routing mechanism calls the corresponding regression head to complete damage size estimation based on the fault location. These modules form a collaborative diagnostic link from information filtering, noise suppression, temporal modeling to branched regression. The modules cooperate and work together to achieve unified modeling for bearing condition category identification and quantitative damage depth assessment.

[0068] 5. This invention can perform end-to-end joint training and inference on normalized multi-channel original sliding window signals without the need for cumbersome manual feature engineering.

[0069] Example 2 like Figure 5 As shown, embodiments of the present invention also provide a bearing fault depth regression system based on cross-channel gating and residual shrinkage networks, including: The data acquisition and preprocessing module is used to acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and to preprocess the monitoring signals to obtain multi-channel time-series samples.

[0070] The cross-channel gating module is used to receive the multi-channel time-series samples, calculate the time-domain energy description value of each channel, construct a time-domain energy description vector from the time-domain energy description values ​​of each channel, generate dynamic weights based on the time-domain energy description vectors, and use the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal.

[0071] The deep residual shrinking network encoder module includes a deep residual shrinking network, which contains an adaptive soft threshold shrinking layer for receiving the weighted multi-channel signal, suppressing background noise, extracting fault features, and outputting feature representations.

[0072] The multi-task diagnostic module includes a classification head and multiple regression heads for receiving the feature representation. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type for estimating the damage degree of the corresponding type. Based on the identification result of the classification head, the module selects the corresponding regression head from the regression heads to output the damage degree estimate.

[0073] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A bearing fault depth regression method based on cross-channel gating and residual shrinkage network, characterized in that, Includes the following steps: S1. Acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and preprocess the monitoring signals to obtain multi-channel time-series samples; S2. Input the multi-channel time series samples into the cross-channel gating module. The cross-channel gating module generates dynamic weights based on the time-domain energy description vector composed of the time-domain energy description values ​​of each channel, and uses the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal. S3. Input the weighted multi-channel signal into a deep residual shrinkage network. The deep residual shrinkage network includes an adaptive soft threshold shrinkage layer to suppress background noise and extract fault features, and output feature representation. S4. Input the feature representation into the multi-task diagnostic module. The multi-task diagnostic module includes a classification head and multiple regression heads. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type to estimate the damage degree of the corresponding type. Based on the identification results of the classification head, the corresponding regression head is selected from each regression head to output the damage level estimate.

2. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, In step S1, the preprocessing includes: aligning the monitoring signal in time, dividing it into fixed-length multi-channel time series samples using a sliding window method, and standardizing the samples to eliminate dimensional differences between different physical quantity channels.

3. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, In step S2, the cross-channel gating module generates dynamic weights in the following way: Calculate the time-domain energy description value for each channel separately. ,in Indicates the first The first channel in the The signal value at each time point, Indicates the length of time; The description vector is composed of the time-domain energy description values ​​of each channel. The description vector is input into a two-layer fully connected network to obtain channel scores, and then processed... Normalization yields the dynamic weights of each channel; The dynamic weights of each channel are applied to the corresponding channel at each time point in the current time series sample to obtain a weighted multi-channel signal, thereby achieving adaptive fusion of multi-channel signals.

4. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, In step S3, the deep residual shrinking network includes an initial convolutional stem and multiple residual shrinking blocks, each of which embeds the adaptive soft threshold shrinking layer; The adaptive soft threshold shrinkage layer adaptively determines the threshold based on the global statistics of the current feature map and performs a soft threshold shrinkage operation on the feature map to retain the fault impact features while suppressing noise components.

5. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, Between step S3 and step S4, the following is also included: The feature representation output from step S3 is input into a bidirectional long short-term memory network for temporal enhancement, and the output after temporal enhancement is averaged along the time dimension to obtain a window-level feature vector. The window-level feature vector is then input into the multi-task diagnostic module.

6. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, The multi-task diagnostic module used in step S4 is obtained through the following training process: The multi-task diagnostic module is trained using a joint loss function, which includes classification loss and regression loss. The classification loss is calculated based on the predicted probability of the state category output by the classification head and the true state category label; During the training phase, the true state category label is used as the routing basis for the regression head. When the true state category label is a fault category, the regression loss is calculated only based on the regression head output value and damage degree label corresponding to the fault category. Other regression heads do not participate in the calculation of the regression loss. When the true state category label is normal state, the normal sample does not participate in the regression loss calculation, and its corresponding regression loss term is 0.

7. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 6, characterized in that, In step S4, the classification loss uses cross-entropy loss, and the regression loss uses smoothing L1 loss; The smoothing parameter for the smoothed L1 loss is set to 0.2; The regression loss weights for different regression heads are adaptively adjusted using learnable parameters.

8. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, In step S4, the bearing condition category includes at least normal condition, inner ring fault condition, and outer ring fault condition; Each of the described regression heads includes at least an inner-circle dedicated regression head and an outer-circle dedicated regression head; When the classification head identifies the damage as normal, the damage level output is 0; When an inner ring fault is identified, the damage level is output by calling the inner ring dedicated regression head. When an outer ring fault is identified, the damage level is output by calling the dedicated outer ring regression head.

9. The bearing fault depth regression method based on cross-channel gating and residual shrinkage network according to claim 1, characterized in that, In step S4, the degree of damage is the etching damage depth, which is measured in millimeters; the method adopts an end-to-end joint training approach, using an optimizer to train the network parameters.

10. A bearing fault depth regression system based on cross-channel gating and residual shrinkage network, characterized in that, include: The data acquisition and preprocessing module is used to acquire bearing operating status monitoring signals synchronously collected by multiple channels of sensors, and to preprocess the monitoring signals to obtain multi-channel time-series samples. A cross-channel gating module is used to receive the multi-channel time-series samples, calculate the time-domain energy description value of each channel, construct a time-domain energy description vector from the time-domain energy description values ​​of each channel, generate dynamic weights based on the time-domain energy description vectors, and use the dynamic weights to perform weighted fusion of the signals of each channel to obtain a weighted multi-channel signal. The deep residual shrinking network encoder module includes a deep residual shrinking network, which contains an adaptive soft threshold shrinking layer for receiving the weighted multi-channel signal, suppressing background noise and extracting fault features, and outputting feature representations. The multi-task diagnostic module includes a classification head and multiple regression heads for receiving the feature representation. The classification head is used to identify the bearing's state category, and each regression head corresponds to a different fault type for estimating the degree of damage for that type. Based on the identification results of the classification head, the corresponding regression head is selected from each regression head to output the damage level estimate.