Rotating machinery fault diagnosis method and system based on multichannel data

Abstract

The invention belongs to the technical field of intelligent manufacturing and equipment fault diagnosis, and provides a rotating machine fault diagnosis method and system based on multichannel data, and the method comprises the steps: obtaining the basic probability distribution of each channel according to the frequency domain energy significance adjustment, and obtaining the basic probability distribution of each channel in the frequency domain energy significance adjustment. Signal weights of the X axis and the Y axis are dynamically adjusted according to importance, contribution of low-frequency periodic signals is strengthened, and weights corresponding to high-frequency impact signals of the Z axis are increased; the similarity between different channel signals is measured through the consistency index of each channel signal, and the consistency of the signals is enhanced through consistency adjustment; quantizing the conflict degree between different channels; for conflicts between X-axis signals, Y-axis signals and Z-axis signals, according to the frequency domain energy, significance and consistency of the signals, the weight of the signals with large conflicts is reduced, the weight of each channel can be dynamically adjusted, and therefore it is ensured that key signals are fully utilized, and the influence of redundancy and conflict information is effectively restrained.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and equipment fault diagnosis technology, and in particular relates to a method and system for diagnosing rotating machinery faults based on multi-channel data. Background Technology

[0002] Rotating machinery is widely used in industrial systems, and its fault detection capability is directly related to the safety of equipment operation and overall production efficiency. However, rotating equipment is often affected by complex operating conditions during operation, such as high noise, load variations, and uncertainties in multi-source information, which makes fault diagnosis a challenging task.

[0003] Current methods for diagnosing rotating machinery faults face the following problems: the information expression capability of a single sensor is insufficient and it is difficult to adapt to complex working conditions; the multi-source information fusion mechanism lacks effective handling of data conflicts and uncertainties, and the feature extraction and expression capabilities are insufficient in multi-scale environments. For example, local faults such as pitting and cracks usually produce high-frequency impact characteristics in the vertical direction, while systemic faults such as rotor imbalance and misalignment are more often manifested as periodic vibrations, mainly occurring in the horizontal direction. The differences between these faults in different directions lead to signal redundancy and conflict problems. Redundancy problems are mainly manifested in the similarity of vibration information in different channels of the signal, while conflict problems are manifested in the mutual interference of signals in different channels. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a rotating machinery fault diagnosis method and system based on multi-channel data. This invention assigns lower weights to signals with significant conflicts based on factors such as the frequency domain energy, significance, and consistency of the signals, thereby reducing the interference of conflicting signals on the final fusion result. It can dynamically adjust the weights of each channel, ensuring that key signals are fully utilized and effectively suppressing the influence of redundant and conflicting information.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions: In a first aspect, the present invention provides a method for diagnosing rotating machinery faults based on multi-channel data, comprising: Acquire vibration signals of rotating machinery in three channels: X-axis, Y-axis, and Z-axis; The vibration signals from three channels are dynamically weighted and fused. Specifically, the basic probability allocation of each channel is obtained by adjusting the frequency domain energy significance. In the frequency domain energy significance adjustment, the signal weights of the X-axis and Y-axis are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals and increasing the weight of high-frequency impact signals on the Z-axis. The similarity between signals from different channels is measured by the consistency index of each channel signal, and the consistency is enhanced through consistency adjustment. The degree of conflict between different channels is quantified. For the conflict between the X-axis and Y-axis signals and the Z-axis signal, the weight of the conflicting signal is reduced according to the frequency domain energy, significance, and consistency of the signal. Based on the dynamically weighted fused signal and the preset fault diagnosis model, the diagnosis result is obtained; wherein, the fault diagnosis model is a model based on multi-scale convolutional neural network and channel attention mechanism.

[0006] Furthermore, the continuous time series signal of the vibration signal is segmented with a preset window length and the window interval is set to be non-overlapping. A preset number of samples are extracted for each type of fault state, and each sample consists of time series sampling points of a preset window length.

[0007] Furthermore, basic probability allocation Frequency significance Consistency indicators and conflict : ; ; ; ; in, f For frequency; This is for the channel at frequency f The basic probability allocation at the location; α is the significance adjustment factor emphasized by the control weight. For local faults such as cracks and pitting, the significance factor is dynamically adjusted so that the high-frequency components on the Z-axis receive more attention during the processing. and They represent Variance and mean across all channels; ε is the positive constant; β is the parameter; B and C Channels k and j The focal length; The intersection of the sets of focal elements; is an empty set; γ is a parameter.

[0008] Furthermore, dynamic weights The calculation is as follows: ; in, N This represents the total number of channels.

[0009] Furthermore, data fusion for: ; Where ⨁ represents the splicing operation, For channel k The frequency domain characteristics.

[0010] Furthermore, the multi-scale feature extraction module in the fault diagnosis model uses parallel convolution kernels of different sizes to extract features at different scales, and concatenates the feature vectors extracted at multiple scales into a unified feature matrix.

[0011] Furthermore, in the channel attention enhancement module of the fault diagnosis model, the dynamic adjustment formula for the weight of each channel is as follows: ; in, Indicates the first k Feature vectors of each channel; and It is a trainable weight matrix; ReLU and These represent the ReLU and Sigmoid activation functions, respectively. L It is the feature length.

[0012] Secondly, the present invention also provides a rotating machinery fault diagnosis system based on multi-channel data, comprising: The data acquisition module is configured to acquire vibration signals of the rotating machinery in three channels: X-axis, Y-axis, and Z-axis. The data fusion module is configured to dynamically weight and fuse vibration signals from three channels. Specifically, it obtains the basic probability allocation for each channel based on frequency domain energy saliency adjustment. In this adjustment, the signal weights for the X and Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals and increasing the weight of high-frequency impact signals on the Z axis. The similarity between different channel signals is measured using a consistency index, and signal consistency is enhanced through consistency adjustment. The degree of conflict between different channels is quantified. For conflicts between X and Y axis signals and the Z axis signal, the weight of conflicting signals is reduced based on their frequency domain energy, saliency, and consistency. The fault diagnosis module is configured to obtain a diagnosis result based on the dynamically weighted fused signal and a preset fault diagnosis model; wherein the fault diagnosis model is a model based on a multi-scale convolutional neural network and a channel attention mechanism.

[0013] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the rotating machinery fault diagnosis method based on multi-channel data described in the first aspect.

[0014] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the rotating machinery fault diagnosis method based on multi-channel data described in the first aspect.

[0015] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the rotating machinery fault diagnosis method based on multi-channel data described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention first generates a basic probability allocation for each channel based on frequency domain energy calculation. Periodic vibration signals on the X and Y axes typically exhibit a more prominent performance in the low-frequency range. Therefore, during frequency domain energy saliency adjustment, the signal weights of the X and Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals. Meanwhile, the high-frequency impact signal on the Z axis is prominent in the high-frequency range, thus receiving a higher weight in frequency domain energy saliency adjustment to ensure sufficient attention is paid to high-frequency impact information. Next, the similarity between signals from different channels is measured by calculating the consistency index of each channel signal. For example, under crack fault conditions, the impact characteristics of the Z axis may have a high similarity to the periodic signals of the X / Y axes. Consistency adjustment enhances the consistency of these signals, enabling more accurate focus on reliable information and reducing misjudgments caused by conflicting signals. Finally, the degree of conflict between different channels is quantified through the calculation of the conflict index. For conflicts between X / Y axis and Z axis signals, lower weights are assigned to signals with greater conflicts based on factors such as frequency domain energy, significance, and consistency of the signals. This reduces the interference of conflicting signals on the final fusion result. The weights of each channel can be dynamically adjusted to ensure that key signals are fully utilized and to effectively suppress the influence of redundant and conflicting information.

[0017] 2. This invention proposes a multi-scale convolutional structure, which extracts features in parallel by setting receptive fields of different sizes. It can simultaneously capture features of minute local faults and overall motion anomalies, demonstrating good adaptability and generalization ability in rotating machinery fault diagnosis. After multi-scale feature extraction, a channel attention mechanism is introduced to automatically increase the response weight of key fault features.

[0018] 3. The method of this invention addresses the conflict and inconsistency issues of multi-channel signals by introducing a dynamic weighting mechanism to overcome their uncertainty and conflict characteristics, effectively improving the robustness of data fusion and feature representation capabilities. In the feature extraction stage, the combination of multi-scale convolution and channel attention mechanisms significantly enhances the model's ability to perceive complex fault modes. Validation on multiple bearing fault datasets shows that the method of this invention outperforms some existing comparative methods in terms of accuracy, robustness, and adaptability to low signal-to-noise ratio, demonstrating good promotional value and engineering application potential. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a flowchart of the intelligent fault diagnosis process in Embodiment 1 of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0023] Example 1: Rotating machinery is widely used in industrial systems, and its fault detection capability directly affects the safety of equipment operation and overall production efficiency. However, rotating equipment is often affected by complex operating conditions, such as high noise, load variations, and uncertainties from multiple sources, making fault diagnosis a challenging task. Traditional fault diagnosis methods often rely on a combination of signal processing techniques and classical machine learning algorithms. They extract time-frequency features using methods such as Fourier transform and wavelet transform, and then use models such as support vector machines or random forests for classification. While these methods have some effectiveness, they are highly dependent on manual feature design, significantly influenced by subjective experience, and perform poorly in high-dimensional, nonlinear, and low signal-to-noise ratio environments, exhibiting limited robustness.

[0024] With the development of deep learning, models such as convolutional neural networks, deep belief networks, and long short-term memory networks have been widely used in fault diagnosis, possessing automatic feature extraction capabilities and reducing the subjectivity of feature design. However, most current deep learning methods are still based on single sensor data, making it difficult to comprehensively perceive system-level fault information, and they are quite sensitive to strong noise or sensor faults, limiting their practical industrial application value.

[0025] In recent years, multi-sensor fusion has gradually become a research hotspot. Data fusion can be divided into three categories: data-layer fusion, feature-layer fusion, and decision-layer fusion. Data-layer fusion directly integrates the original sensor signals, which can reduce information loss, but it suffers from problems such as data redundancy, conflicts, and uncertainties. Feature-layer fusion extracts features from each sensor before fusing them, which can improve representation capabilities, but it has high computational complexity and requires reasonable weight allocation to avoid information interference. Decision-layer fusion improves robustness by integrating the outputs of multiple classifiers. Common methods include voting mechanisms and DS evidence theory, but its performance is highly dependent on the fusion strategy and has certain limitations.

[0026] In summary, current methods for diagnosing rotating machinery faults still face the following problems: the information expression capability of a single sensor is insufficient and it is difficult to adapt to complex working conditions; the multi-source information fusion mechanism lacks effective handling of data conflicts and uncertainties; and the feature extraction and expression capabilities are insufficient in multi-scale environments.

[0027] To address at least one of the aforementioned problems, this embodiment provides a rotating machinery fault diagnosis method based on multi-channel data. Based on multi-channel vibration signals, dynamic weighted fusion is performed using Dempster–Shafer evidence theory to improve the stability and robustness of signal representation. The fused signal is then subjected to deep feature extraction via a multi-scale convolutional neural network and a channel attention mechanism, thereby achieving high-precision identification of multiple fault states. Experimental results show that this method maintains good diagnostic performance even in low signal-to-noise ratio environments, with high identification accuracy and strong adaptability. Compared to existing technologies, the method in this embodiment improves diagnostic reliability and reduces the need for manual intervention, possessing significant engineering application value. Specifically, it includes: S1. Signal Acquisition: Optionally, a multi-functional rotor test platform can be used for fault signal acquisition. The platform consists of a drive motor, shaft, coupling, test bearing housing, load device, sensor assembly, and data acquisition system. The component under test is installed in the bearing housing, and the shaft is connected to the load through the coupling, which can simulate various speeds and load conditions.

[0028] The sampling speeds were 1000 rpm, 2000 rpm, and 3000 rpm, with torques of 0 Nm, 3.65 Nm, and 7.3 Nm, respectively. The sampling frequency was set to 12 kHz, with 200,000 sampling points per group and a sampling duration of 16 seconds per session. Two groups were collected, totaling 512,000 data points. This covered various typical bearing and rotor fault conditions, including: normal, inner ring crack, outer ring crack, inner ring pitting, outer ring pitting, rolling element pitting, cage fracture, inner and outer ring pitting, imbalance, misalignment, rotor rubbing, rotor crack, and rotor deformation. Pitting was performed using laser etching, with each pitting point being etched 5 times; cracks were simulated using wire cutting.

[0029] The acquisition system is configured with three channels to simultaneously record multimodal physical signals. The channel configuration is as follows: Channel 1: X-axis acceleration sensor (mounted in the test bearing housing); Channel 2: Y-axis acceleration sensor (mounted in the test bearing housing); Channel 3: Z-axis acceleration sensor (mounted in the test bearing housing).

[0030] The data recorded by the acquisition system includes the sensitivity configuration of each channel, sensor type identification, channel coupling relationship, channel characteristic values, and the full physical quantity sampling sequence. Ultimately, raw multi-channel vibration signal data with clear labels are obtained, which can be used as input for subsequent fault identification modeling and verification.

[0031] S2, Data Fusion Processing: In the fault diagnosis of rotating machinery, different types of faults will produce differentiated vibration responses in the three axial directions. For example, localized faults (such as pitting and cracks) usually produce high-frequency impact characteristics in the vertical direction (Z-axis), while systemic faults (such as rotor imbalance and misalignment) are more likely to exhibit periodic vibrations, mainly occurring in the horizontal direction (X / Y axes). These differences in faults in different directions lead to signal redundancy and conflict issues.

[0032] Redundancy issues mainly manifest as similar vibration information across different signal channels. For example, localized faults (such as cracks or pitting) typically present as significant high-frequency impact signals in the Z-axis direction, while systemic faults (such as rotor imbalance) exhibit periodic vibrations in the X / Y-axis directions. This repetitive signal information can lead to redundancy between different signal channels, thus affecting the accuracy of fault diagnosis models.

[0033] The conflict problem manifests as mutual interference between signals from different channels. For example, the high-frequency impact signal on the Z-axis may be masked by the periodic vibration signal on the X / Y axes, preventing critical signals from being fully reflected. In this case, simple signal splicing methods may cause redundant signals to mask critical signals or introduce conflicting information from different directions, further affecting the accuracy of fault identification.

[0034] Therefore, to effectively address redundancy and conflict issues, this embodiment proposes a dynamic weighted fusion method based on DS evidence theory. By quantifying the saliency, consistency, and conflict of signals from different channels, the weights of each channel can be dynamically adjusted, thereby ensuring that key signals are fully utilized and effectively suppressing the impact of redundant and conflicting information.

[0035] Specifically, the basic probability assignment (BPA) for each channel is first calculated based on frequency domain energy. Periodic vibration signals along the X and Y axes are typically more prominent in the low-frequency range; therefore, during frequency domain energy saliency adjustment, the signal weights of the X / Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals. High-frequency impact signals along the Z axis, however, are prominent in the high-frequency range, thus receiving a higher weight in frequency domain energy saliency adjustment to ensure sufficient attention is paid to high-frequency impact information. Next, the similarity between signals from different channels is measured by calculating a consistency index. For example, under crack fault conditions, the impact characteristics of the Z-axis may have high similarity to the periodic signals of the X / Y axes. Through consistency adjustment, the model enhances the consistency of these signals, helping the algorithm to more accurately focus on reliable information and reducing misjudgments caused by conflicting signals. Finally, the degree of conflict between different channels is quantified by calculating a conflict index. For conflicts between X / Y axis and Z axis signals, the model will assign lower weights to signals with greater conflicts based on factors such as frequency domain energy, significance, and consistency, thereby reducing the interference of conflicting signals on the final fusion result.

[0036] To fully extract key fault features from different channels and improve the quality of the fused data, this embodiment uses a fixed-length sliding window to divide the samples for each type of fault data during the data preprocessing stage. Specifically, the acquired continuous time series signals are segmented with a window length of 1024, and the window intervals are set to be non-overlapping to ensure the independence and representativeness of the samples. 200 samples are extracted for each type of fault state, and each sample consists of 1024 time-series sampling points. This results in a multi-class dataset covering various fault conditions for subsequent data fusion. Furthermore, to further address the potential directional redundancy, conflicts, and noise interference issues in multi-channel vibration signal fusion, this embodiment proposes a weighted fusion method based on DS evidence theory, specifically including the following steps: S2.1 Generating Basic Probability Assignment (BPA): Let the first... k The frequency domain energy of each channel is Then its frequency f The basic probability assignment is defined as follows: ; In the process of BPA generation, high-frequency impact signals along the Z-axis (such as cracks and pitting) are given higher weight in frequency domain energy calculation, thereby enhancing the identification of local faults.

[0037] S2.2, Frequency Significance: ; in, Is the channel at frequency f The BPA is used to control the significance adjustment factor, while α is the significance adjustment factor emphasized by the control weight. For local faults such as cracks and pitting, we dynamically adjust the significance factor so that the high-frequency components on the Z-axis receive more attention during the processing, while the periodic signals on the X / Y axes are kept in balance during the processing.

[0038] S2.3 Consistency Indicators: ; in, and They represent The variance and mean across all channels. ε is a small positive constant to avoid division by zero. The parameter β is used to adjust for the impact of consistency. For localized faults such as cracks and pitting, we enhance consistency to ensure that similar information across different channels is reinforced, thereby helping the model focus on reliable information and reducing interference from redundant information.

[0039] S2.4, Conflict: ; in, B and C These are channels k and j Jiao Yuan, This represents the intersection of sets of focal elements, while This represents the empty set. The parameter γ is used to adjust the sensitivity to collisions. Under localized faults such as pitting and cracks, the Z-axis signal may collide with signals from other channels due to its high-frequency characteristics. By adjusting the collision sensitivity, we can quantify the collisions and adjust the signal weights to ensure that high-frequency fault signals such as cracks and pitting still receive high weights even under conditions of significant collisions.

[0040] S2.5 Dynamic Weight Calculation: ; in, N The total number of channels is given. By calculating the dynamic weight of each channel, we comprehensively consider frequency significance, consistency, and conflict, and adjust the weights according to the characteristics of different fault types (such as cracks and pitting) to ensure effective signal fusion.

[0041] S2.6 Data Fusion: The calculated dynamic weights are used to fuse the BPA and frequency domain features. ; Where ⨁ represents the splicing operation; For channel kThe frequency domain characteristics are analyzed. In this embodiment, dynamic weighted fusion can highlight the high-frequency impact characteristics of local faults such as cracks and pitting in multi-channel signals, while suppressing the influence of redundant directional information. The final fusion result can effectively improve the accuracy and robustness of fault diagnosis.

[0042] Although the acquisition platform in this embodiment supports simultaneous acquisition and extended analysis of multi-source signals such as current and temperature, this embodiment focuses on the differences in rotating machinery faults across multiple channels, employing the DS evidence theory to perform dynamic weighted fusion of X / Y / Z three-channel vibration signals. Under fault conditions such as bearing pitting, rotor imbalance, and cracks, fixed weighting or simple splicing often ignores directional sensitivity, reducing the model's ability to identify faults. Therefore, this embodiment represents the importance of each channel through frequency domain energy saliency, consistency, and conflict, and performs dynamic data fusion accordingly. This enhances key directional information and suppresses redundant directional information. This method is not only applicable to data fusion of similar signals but also has the ability to be extended to heterogeneous signals such as current and temperature, enabling true multi-source fusion diagnosis in future applications.

[0043] S3. Feature Extraction and Feature Fusion: Rotating machinery often exhibits complex vibration response characteristics under different fault states, including both high-frequency, low-amplitude local impact signals and systemic abnormal signals with strong periodicity and low frequency. The coexistence of these multi-scale features limits the extraction range and expressive power of traditional single-scale convolutional structures in feature extraction. Furthermore, features fused from multiple channels often exhibit high dimensionality and mixed information, making it difficult for the model to automatically focus on key feature dimensions, thus affecting the robustness and accuracy of fault diagnosis.

[0044] To address the aforementioned issues, this embodiment proposes a fusion method based on multi-scale convolutional neural networks (Multi-Scale CNN) and a channel attention mechanism (Squeeze-and-Excitation). This method aims to comprehensively extract fault features at different scales from the fused signal and automatically increase the weights of key features through the attention mechanism, suppressing redundancy and noise interference, thereby improving the model's diagnostic performance. This method mainly includes the following two sub-modules: S3.1 Multi-scale Feature Extraction Module (MSCNN): MSCNN uses parallel convolutional kernels of different sizes (e.g., 3, 5, 7, 9, 11) to extract features at different scales: ; in, Indicates the use of size k Features extracted by convolution kernels and These are the weights and biases of the convolution kernel, respectively. Represents the activation function ReLU; Feature vectors extracted at multiple scales are concatenated into a unified feature matrix: ; in, Indicates by the first i The feature vectors obtained from the convolutional kernels. This multi-scale concatenation generates a high-dimensional representation that integrates multiple receptive fields.

[0045] In rotating machinery, different fault types often exhibit vibration characteristics at different scales, making it difficult for single-scale convolution operations to effectively extract features simultaneously. Therefore, this embodiment proposes a multi-scale convolutional structure (MSCNN), which extracts features in parallel by setting receptive fields of different sizes. This allows for the simultaneous capture of features from minute local faults and overall motion anomalies, demonstrating good adaptability and generalization ability in rotating machinery fault diagnosis.

[0046] S3.2 Channel Attention Enhancement Module (SE): In the SE module, the weight of each channel is dynamically adjusted using the following formula: ; in, Indicates the first k Feature vectors of each channel and It is a trainable weight matrix, ReLU and These represent the ReLU and Sigmoid activation functions, respectively. L It is the feature length.

[0047] To enhance the model's ability to perceive key fault features, this embodiment introduces a channel attention mechanism (Squeeze-and-Excitation, SE module) after multi-scale feature extraction, which automatically increases the response weights of key fault features.

[0048] S4. In the fault identification phase, when the system experiences an abnormal signal again, multi-channel vibration data is first collected and divided into multiple signal samples using a sliding window. Subsequently, based on the same dynamic weighted fusion strategy used in the training phase, fusion feature vectors of the signal under frequency domain saliency, consistency, and conflict degree are extracted. This feature vector is directly input into the trained MSCNN-SE model, which uses multi-scale convolutional kernels to extract local and global features, and the SE module to adaptively allocate importance weights between channels, thereby strengthening the response to key fault modes and suppressing the influence of interference terms. Finally, the model automatically outputs the fault type label for each sample, including normal state, typical bearing faults (such as inner ring cracks, pitting, cage fracture), and rotor faults (such as imbalance, misalignment, cracks, etc.), achieving fully automated identification without manual intervention.

[0049] Compared to traditional methods, the method in this embodiment addresses the conflict and inconsistency issues of multi-channel signals by introducing a dynamic weighting mechanism to overcome their uncertainty and conflict characteristics, effectively improving the robustness of data fusion and feature representation capabilities. The feature extraction stage combines multi-scale convolution and channel attention mechanisms, significantly enhancing the model's ability to perceive complex fault modes. Validation on multiple bearing fault datasets shows that the method in this embodiment outperforms some existing comparative methods in terms of accuracy, robustness, and adaptability to low signal-to-noise ratios, demonstrating good generalization value and engineering application potential.

[0050] To verify the performance of the intelligent fault diagnosis method in this embodiment, fault diagnosis was first performed on the data collected from the rotating machinery test bench using this method. In this embodiment, the learning rate was set to 0.01, the loss function to cross-entropy, and the maximum number of training iterations to 100. To simulate common interferences in actual industrial environments, Gaussian white noise with different signal-to-noise ratios (SNR) was injected into the data. The fault diagnosis method of this embodiment was compared with several neural network model methods (CNN, ResNet, AlexNet, SAE, MLP, VAFCNN, MSCNN-FRDF, and MIFMN) in terms of classification accuracy. The comparison data of classification accuracy are shown in Table 1. CNN, ResNet, AlexNet, SAE, and MLP are five traditional fault diagnosis methods, with the following specific structures: The MLP architecture contains three hidden layers with 512, 256, and 128 neurons respectively, the input batch size is set to 64, and all layers use the ReLU activation function. The SAE model is designed with an input layer containing 1024 neurons, followed by three hidden layers containing 768, 512, and 128 neurons respectively. The basic CNN architecture consists of three convolutional layers (kernel sizes of 15, 10, and 5), two max-pooling layers, and one global average pooling layer. The ResNet model employs a modified deep architecture, utilizing skip connections to maintain gradient flow during training. The model configuration is based on the standard ResNet guidance scheme. The AlexNet model contains five convolutional layers (each with a different kernel size) and a series of max-pooling layers.

[0051] Table 1. Comparison of classification accuracy results for fault diagnosis classification (%)

[0052] As can be seen from the experimental results in Table 1, the fault diagnosis method proposed in this embodiment exhibits extremely high fault identification accuracy under different signal-to-noise ratio conditions, especially maintaining a classification performance of 99.61% even in a -10dB strong noise environment, significantly outperforming other comparative models. This indicates that this embodiment not only effectively improves the robustness of fault identification in low signal-to-noise ratio scenarios but also greatly enhances the accuracy of fault diagnosis. Compared to traditional methods that rely on a large amount of feature construction and redundant sensor data, this embodiment achieves automatic identification of key fault features through multi-scale feature extraction and dynamic fusion mechanisms, thereby simplifying the diagnostic process and reducing industrial operation and maintenance costs. For industrial equipment maintenance personnel, the fault category can be quickly determined simply by using the fused features extracted in this embodiment, without relying on complex signal analysis or expensive manual experience intervention. Therefore, this embodiment greatly improves the efficiency and practicality of rotating machinery fault diagnosis while ensuring equipment operation safety, and has good engineering application prospects.

[0053] Example 2: This embodiment provides a rotating machinery fault diagnosis system based on multi-channel data, including: The data acquisition module is configured to acquire vibration signals of the rotating machinery in three channels: X-axis, Y-axis, and Z-axis. The data fusion module is configured to dynamically weight and fuse vibration signals from three channels. Specifically, it obtains the basic probability allocation for each channel based on frequency domain energy saliency adjustment. In this adjustment, the signal weights for the X and Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals and increasing the weight of high-frequency impact signals on the Z axis. The similarity between different channel signals is measured using a consistency index, and signal consistency is enhanced through consistency adjustment. The degree of conflict between different channels is quantified. For conflicts between X and Y axis signals and the Z axis signal, the weight of conflicting signals is reduced based on their frequency domain energy, saliency, and consistency. The fault diagnosis module is configured to obtain a diagnosis result based on the dynamically weighted fused signal and a preset fault diagnosis model; wherein the fault diagnosis model is a model based on a multi-scale convolutional neural network and a channel attention mechanism.

[0054] The working method of the system is the same as that of the rotating machinery fault diagnosis method based on multi-channel data in Embodiment 1, and will not be repeated here.

[0055] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the rotating machinery fault diagnosis method based on multi-channel data described in Embodiment 1.

[0056] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the rotating machinery fault diagnosis method based on multi-channel data described in Embodiment 1.

[0057] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the rotating machinery fault diagnosis method based on multi-channel data described in Embodiment 1.

[0058] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for diagnosing rotating machinery faults based on multi-channel data, characterized in that, include: Acquire vibration signals of rotating machinery in three channels: X-axis, Y-axis, and Z-axis; Dynamically weighted and fused vibration signals from three channels; Specifically, the basic probability allocation for each channel is obtained based on frequency domain energy saliency adjustment. In this adjustment, the signal weights of the X and Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals and increasing the weight of high-frequency impulse signals on the Z axis. The similarity between different channel signals is measured using the consistency index of each channel signal, and the consistency is enhanced through consistency adjustment. The degree of conflict between different channels is quantified. For conflicts between X and Y axis signals and Z axis signals, the weight of conflicting signals is reduced based on their frequency domain energy, saliency, and consistency. Based on the dynamically weighted fused signal and the preset fault diagnosis model, the diagnosis result is obtained; wherein, the fault diagnosis model is a model based on multi-scale convolutional neural network and channel attention mechanism.

2. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 1, characterized in that, The continuous time series signal of the vibration signal is segmented with a preset window length and the window interval is set to be non-overlapping. A preset number of samples are extracted for each type of fault state, and each sample consists of time series sampling points of a preset window length.

3. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 1, characterized in that, Basic probability allocation Frequency significance Consistency indicators and conflict : ； ； ； ； in, f For frequency; This is for the channel at frequency f The basic probability allocation at the location; α is the significance adjustment factor emphasized by the control weight. For local faults such as cracks and pitting, the significance factor is dynamically adjusted so that the high-frequency components on the Z-axis receive more attention during the processing. and They represent Variance and mean across all channels; ε is the positive constant; β is the parameter; B and C Channels k and j The focal length; The intersection of the sets of focal elements; is an empty set; γ is a parameter.

4. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 3, characterized in that, Dynamic weights The calculation is as follows: ； in, N This represents the total number of channels.

5. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 4, characterized in that, Data fusion for: ； Where ⨁ represents the splicing operation, For channel k The frequency domain characteristics.

6. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 1, characterized in that, The multi-scale feature extraction module in the fault diagnosis model uses parallel convolution kernels of different sizes to extract features at different scales, and concatenates the feature vectors extracted from multiple scales into a unified feature matrix.

7. The rotating machinery fault diagnosis method based on multi-channel data as described in claim 6, characterized in that, In the channel attention enhancement module of the fault diagnosis model, the dynamic adjustment formula for the weight of each channel is as follows: ； in, Indicates the first k Feature vectors of each channel; and It is a trainable weight matrix; ReLU and These represent the ReLU and Sigmoid activation functions, respectively. L It is the feature length.

8. A rotating machinery fault diagnosis system based on multi-channel data, characterized in that, include: The data acquisition module is configured to acquire vibration signals of the rotating machinery in three channels: X-axis, Y-axis, and Z-axis. The data fusion module is configured to dynamically weight and fuse vibration signals from three channels. Specifically, it obtains the basic probability allocation for each channel based on frequency domain energy saliency adjustment. In this adjustment, the signal weights for the X and Y axes are dynamically adjusted according to their importance, strengthening the contribution of low-frequency periodic signals and increasing the weight of high-frequency impact signals on the Z axis. The similarity between different channel signals is measured using a consistency index, and signal consistency is enhanced through consistency adjustment. The degree of conflict between different channels is quantified. For conflicts between X and Y axis signals and the Z axis signal, the weight of conflicting signals is reduced based on their frequency domain energy, saliency, and consistency. The fault diagnosis module is configured to obtain a diagnosis result based on the dynamically weighted fused signal and a preset fault diagnosis model; wherein the fault diagnosis model is a model based on a multi-scale convolutional neural network and a channel attention mechanism.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the steps of the rotating machinery fault diagnosis method based on multi-channel data as described in any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the rotating machinery fault diagnosis method based on multi-channel data as described in any one of claims 1-7.

Images