Equipment fault diagnosis method based on multi-scale image convolution

Abstract

The invention provides an equipment fault diagnosis method based on multi-scale image convolution, and relates to the technical field of industrial equipment intelligent fault diagnosis, and the method comprises the steps: extracting the time domain, frequency spectrum domain and time-frequency features of a vibration signal through a multi-scale input layer, and generating a 32-dimensional feature vector through the fusion of a cross-scale feature coupling module; and constructing an inter-equipment relation perception graph convolution model, generating a dynamic adjacency matrix in combination with a physical distance and a real-time working condition, and extracting space-time fusion features through space-time convolution. Transient and periodic features are enhanced through an accidental fault sensitive time sequence module, and time sequence features are output in combination with dual-channel fusion and a self-attention mechanism. And finally, fault identification and positioning are realized by adopting a double-threshold detection and equipment comparison enhancement strategy, an interpretable diagnosis evidence chain containing multi-scale feature contribution is generated, and the weak fault detection rate and the diagnosis credibility are improved.

Description

Technical Field

[0001] The present application belongs to the technical field of intelligent fault diagnosis of industrial equipment, and specifically relates to an equipment fault diagnosis method based on multi-scale graph convolution. Background Art

[0002] In modern industrial production, rotating machinery (such as fans, pumps, and compressors) is a key component. The stability and reliability of its operating status are crucial to the safety and efficiency of the entire production process. With the development of the Industrial Internet of Things (IIoT), big data, and artificial intelligence technologies, fault diagnosis methods based on predictive maintenance have gradually become a research hotspot. Traditional fault diagnosis methods mainly rely on single-scale feature extraction and static models, which have difficulty effectively addressing issues such as multi-scale feature coupling, dynamic correlation modeling between devices, and transient event capture. Especially in complex industrial environments, weak fault signals are easily drowned out by factors such as background noise and vibration interference, making early fault detection difficult.

[0003] The inventors discovered that existing methods often use simple splicing or weighted fusion to process multi-scale features, lacking an effective cross-scale feature coupling mechanism, making it difficult to identify weak fault signatures. Traditional graph convolutional networks (GCNs) typically use a static adjacency matrix, which fails to reflect the dynamic adjustment requirements of vibration transmission relationships between devices as operating conditions change. Traditional dual-threshold detection methods rely on manual experience to set thresholds, lacking adaptability and prone to false positives and missed detections in complex industrial environments. Summary of the Invention

[0004] This application provides a device fault diagnosis method based on multi-scale graph convolution to solve one of the above technical problems.

[0005] The technical solutions adopted in this application are: The present application provides a device fault diagnosis method based on multi-scale graph convolution, comprising: The vibration signal of the group control equipment is decomposed at multiple scales through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal and generate multi-dimensional feature data; Dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector; Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed through the occasional fault sensitive timing module, the Teager energy operator is used to enhance the transient impact features, and the adaptive comb filter is combined to extract the periodic fault features. The dual-channel weight fusion and self-attention mechanism are used to output the timing features. Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0006] According to one embodiment of the present application, the vibration signal of the group control device is multi-scale decomposed through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal to generate multi-dimensional feature data, specifically: Use IEPE accelerometers to collect triaxial vibration signals and calculate transient impact characteristics such as peak value, pulse factor, and impact duration; waveform factor and skewness waveform distribution characteristics; high-frequency energy proportion and zero-crossing rate energy change characteristics; Extract peak frequency bandwidth, sideband energy ratio, and sideband frequency interval modulation characteristics through envelope analysis; The wavelet transform method is used to generate multi-scale energy entropy features, sub-band energy features and time-frequency domain dynamic features.

[0007] According to one embodiment of the present application, the multi-dimensional feature data is dynamically interactively fused through the cross-scale feature coupling module, and the recognition of weak fault features is enhanced by a bidirectional gating mechanism, and the fused 32-dimensional feature vector is output, specifically: performing standardization processing on the multidimensional feature data; The standardized multi-dimensional feature data is spliced ​​in the channel dimension, and the importance weights of the features at each scale are learned through one-dimensional convolution and softmax function; The bidirectional gating mechanism is used to achieve complementary enhancement of features between scales, including: Strengthen the correlation between time domain features and frequency domain feature frequencies; Strengthen the relationship between frequency domain features and time-frequency features; The residual connection method is used to retain the key information of the original features and output the fused 32-dimensional feature vector.

[0008] According to one embodiment of the present application, the 32-dimensional feature vector is used to construct a convolutional model of the relationship perception graph between devices, and a dynamic adjacency matrix is ​​generated based on the physical distance of the devices and the similarity of the real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features, specifically: The sensor of each device is a graph node, and the node feature is the 32-dimensional feature vector; Dynamically adjust the adjacency matrix weights based on the physical distance between devices and the similarity of real-time working conditions; The physical distance factor is multiplied by the speed difference factor to generate a physical similarity matrix; The dynamic fusion coefficient α is adjusted according to the physical distance between devices. The smaller the α value, the higher the adjacency weight; The vibration transmission law between devices is captured through spatial convolution operation, and the correlation features in the time series are extracted through temporal convolution operation. The feature flow is controlled by the gating unit to alleviate the over-smoothing problem.

[0009] According to one embodiment of the present application, the spatiotemporal fusion features are processed by the sporadic fault sensitive timing module, the transient impact features are enhanced by the Teager energy operator, the periodic fault features are extracted by combining the adaptive comb filter, and the timing features are output by dual-channel weight fusion and self-attention mechanism, specifically: The transient impact energy is calculated by the Teager energy operator, and the high-frequency characteristics are quantified by combining the short-time zero-crossing rate to generate transient features. Extract periodic fault features through adaptive comb filtering, where the center frequency of the filter is dynamically adjusted according to the fault type; The weights of transient features and periodic features are learned through one-dimensional convolution, and the features of the two pathways are fused according to the weights. The attention weight at the moment of fault occurrence is calculated through the self-attention mechanism to improve the temporal resolution.

[0010] According to one embodiment of the present application, the dual-threshold detection mechanism and device contrast enhancement strategy are used to identify and locate faults on the time series features, generating an interpretable diagnostic evidence chain that includes multi-scale feature contributions, specifically: Set transient thresholds and continuous thresholds to correspond to the abnormal judgment of short-term impact characteristics and periodic characteristics respectively; Thresholds are dynamically updated every 24 hours based on the latest normal data to adapt to baseline drift caused by equipment aging; Through the device comparison enhancement strategy, the deviation between the abnormality of a single device and the average abnormality of the group of devices is calculated, and the abnormal offset is generated to enhance the characteristics of weak faults.

[0011] According to one embodiment of the present application, it further includes: Calculate the fault propagation impact based on the dynamic adjacency matrix and determine the fault source device; Generate an interpretable evidence chain that includes multi-scale feature contributions, including the impact energy exceedance multiple, characteristic frequency peak and its change multiple, and correlation changes with neighboring equipment; Output a structured diagnostic report including fault location, evidence chain, and repair suggestions.

[0012] According to one embodiment of the present application, generating the dynamic adjacency matrix includes the following steps: Obtain the physical distance data between group control devices and build a basic distance matrix; Collect the operating condition data of each device in real time and calculate the speed difference and load change similarity between two devices; The distance matrix and the working condition similarity matrix are weightedly fused to generate a dynamic adjacency matrix, in which the weight coefficients are pre-set or adjusted through online learning according to the equipment type and installation environment.

[0013] A second aspect of the present application provides a computer-readable storage medium having a program stored thereon, which implements the steps in the method when executed by a processor.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the steps in the method described above when executing the program.

[0015] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: This application uses a multi-scale input layer to extract the time, spectral, and time-frequency characteristics of vibration signals, comprehensively covering all types of fault characteristics during equipment operation. By utilizing the bidirectional gating mechanism in the cross-scale feature coupling module, dynamic interaction and complementary enhancement between features at different scales are achieved, improving the recognition of weak fault characteristics and achieving a 95.6% detection rate for occasional faults in harsh environments with a signal-to-noise ratio of -10dB.

[0016] This application builds a convolutional model for device-to-device relationship perception graphs. This model generates a dynamic adjacency matrix based on the physical distance between devices and the similarity of their real-time operating conditions, accurately modeling the spatiotemporal relationships between devices. By capturing the mutual influence between devices through spatiotemporal convolution, the application improves fault location accuracy to within 0.5 meters, effectively eliminating positioning errors caused by varying operating conditions.

[0017] This application's sporadic fault-sensitive timing module combines the Teager energy operator and adaptive comb filter to enhance the ability to extract transient impact features. It also uses dual-channel weight fusion and a self-attention mechanism to focus on the moment of fault occurrence, effectively capturing sporadic faults occurring within seconds. A transient-persistent dual-threshold detection mechanism with adaptive threshold adjustment is designed, significantly reducing false alarm rates and missed detections in complex industrial environments.

[0018] This application uses a dual-threshold detection mechanism and device contrast enhancement strategy to identify and locate faults using time series features, generating an interpretable diagnostic evidence chain that incorporates contributions from multi-scale features. The resulting structured diagnostic report facilitates rapid understanding and locating of fault sources by industrial engineers, further enhancing the method's practicality and industrial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of a flow chart of a device fault diagnosis method based on multi-scale graph convolution provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0020] Reference numerals: 810 , processor; 820 , communication interface; 830 , memory; 840 , communication bus. DETAILED DESCRIPTION

[0021] In order to more clearly illustrate the overall concept of the present application, a detailed description is given below in an illustrative manner in conjunction with the accompanying drawings.

[0022] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application may also be implemented in other ways than those described herein, and therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below. It should be noted that the embodiments of the present application and the features of each embodiment may be combined with each other unless there is a conflict.

[0023] In this application, unless otherwise expressly specified and limited, a first feature "above" or "below" a second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples.

[0024] Example 1 like Figure 1 As shown, a device fault diagnosis method based on multi-scale graph convolution includes: S100 , performing multi-scale decomposition on the vibration signal of the group control device through a multi-scale input layer, extracting the time domain features, spectrum domain features, and time-frequency features of the vibration signal, and generating multi-dimensional feature data.

[0025] As mentioned above, the multi-scale input layer is the starting point of the entire fault diagnosis method. Its main purpose is to extract features at different scales from the vibration signals of group-controlled equipment (such as fans and pumps). Specifically, this step involves decomposing the original vibration signal to obtain time domain features, spectral domain features, and time-frequency features.

[0026] Time-domain features: These features are extracted directly from the time series of the original vibration signal and include peak value, pulse factor, impact duration, form factor, skewness, high-frequency energy proportion, and zero-crossing rate. They reflect the signal's characteristics in the time dimension, such as transient impact intensity, waveform distribution, and energy variation.

[0027] Spectral Domain Features: Envelope analysis techniques are used to extract modulation features from vibration signals, including fundamental frequency characteristics (such as peak frequency bandwidth), sideband characteristics (such as sideband energy ratio), and spectral distribution characteristics. These features reveal key information about the signal in the frequency domain and help identify specific fault modes, such as wear or cracks in bearings or gearboxes.

[0028] Time-frequency features: Wavelet transforms are used to perform multi-scale decomposition of the signal, generating time-frequency features that describe sudden energy changes during faults. These features include multi-scale energy entropy features, sub-band energy features, and time-frequency dynamic features. Time-frequency features capture signal changes in both time and frequency, making them particularly useful for detecting non-stationary signals.

[0029] For example, suppose you are monitoring a fan in an industrial plant equipped with a triaxial vibration sensor (X / Y / Z directions). The vibration data collected by the sensor is first fed into the multi-scale input layer: Time Domain Feature Extraction: Calculates parameters such as peak value, pulse factor, and impact duration within each time window during wind turbine operation. For example, if a significant increase in peak value and a prolonged impact duration are observed within a certain time window, this may indicate a serious mechanical failure within the wind turbine.

[0030] Spectral Domain Feature Extraction: By performing a Fast Fourier Transform (FFT) on the vibration signal, we can obtain its spectrum and extract features such as peak frequency bandwidth and sideband energy ratio. For example, if the energy ratio of a sideband is detected to increase suddenly, it may be due to a crack or eccentricity in a fan component.

[0031] Time-frequency feature extraction: Wavelet transform is used to decompose the signal into multiple sub-bands and analyze the energy distribution of each sub-band. For example, if the energy of a sub-band suddenly drops, it may indicate damage or other abnormalities in the wind turbine blades.

[0032] S200 , dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector.

[0033] As mentioned above, the cross-scale feature coupling module is a key step in the entire fault diagnosis method. It aims to enhance the recognition of weak fault signatures by dynamically and interactively fusing features at different scales (time, spectrum, and time-frequency domains). Specifically, this module uses a bidirectional gating mechanism to process multidimensional feature data, ensuring effective complementarity and enhancement between features at each scale, ultimately outputting a 32-dimensional fused feature vector.

[0034] Standardization processing: First, the extracted time domain, spectrum domain and time-frequency domain features are standardized to make them consistent in magnitude, which facilitates subsequent feature fusion.

[0035] Feature concatenation and fusion: The standardized features are concatenated in the channel dimension, and the importance weights of each scale feature are learned through 1D convolution operations. These weights determine the contribution of each feature to the final fused feature.

[0036] Bidirectional gating mechanism: This mechanism enhances inter-scale feature complementarity. For example, it strengthens the correlation between temporal impact features and spectral feature frequencies, as well as the interaction between spectral and time-frequency domain features. Furthermore, a residual connection is used to preserve key information from the original features, preventing loss of important features.

[0037] Output fused feature vector: After the above processing, a fused feature vector containing 32 dimensions is generated. This vector integrates feature information at different scales and can more effectively identify weak fault signals.

[0038] For example, assume that feature data in the time domain, spectrum domain, and time-frequency domain have been extracted from the vibration signal of a wind turbine: Normalization: Normalize time-domain features (such as peak value and impulse factor), spectral-domain features (such as peak frequency bandwidth and sideband energy ratio), and time-frequency features (such as multi-scale energy entropy and sub-band energy features) to maintain consistency in their numerical ranges. The normalized features are then concatenated into a high-dimensional feature matrix.

[0039] A 1D convolutional layer is used to learn the importance weights of features at each scale. For example, it is found that peaks in time domain features are particularly important for identifying transient shocks, while sideband energy in spectrum domain features is more critical for detecting modulation faults.

[0040] A bidirectional gating mechanism is used to achieve feature complementarity between scales. For example, within a certain time window, if the time domain features show a large peak but the spectral domain features do not show significant changes, the bidirectional gating mechanism can be used to enhance the influence of the time domain features to better capture potential transient shocks.

[0041] The residual connection method is used to retain the key information of the original features, ensuring that important fault signals are not lost during the feature fusion process.

[0042] Output fused feature vector: After the above processing, a 32-dimensional fused feature vector is generated. This vector not only contains feature information at all scales, but also enhances the recognition of weak fault features through a bidirectional gating mechanism.

[0043] S300. Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between the devices and the similarity between the real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features.

[0044] As mentioned above, the device-to-device relationship-aware graph convolutional model is built based on the generated 32-dimensional feature vector. It aims to capture the spatiotemporal relationships between group-controlled devices through a dynamic adjacency matrix. This model uses the physical distance between devices and the similarity of their real-time operating conditions (such as speed and load) to adjust the connection weights between devices and extract fused features through spatiotemporal convolution.

[0045] Modeling inter-device relationships: Each device's sensor is treated as a node in a graph, with node features represented by 32-dimensional feature vectors after cross-scale coupling. Based on the physical distance between devices and the similarity of their real-time operating conditions, a dynamic adjacency matrix is ​​constructed to reflect the vibration transmission relationships between devices.

[0046] Dynamic adjacency matrix generation: The dynamic adjacency matrix not only considers the physical distance between devices but also incorporates real-time operating conditions (such as speed differences, load changes, etc.). For example, two physically close devices with similar operating conditions may have a higher association weight, indicating a strong vibration transmission effect between them.

[0047] Spatiotemporal convolution: After constructing the dynamic adjacency matrix, a graph convolutional network (GCN) performs spatial convolution to capture local correlation patterns between devices. Next, a one-dimensional temporal convolution is applied to the feature sequence after spatial convolution to extract temporal correlations. This entire process uses gating units to optimize feature flow and prevent oversmoothing, ultimately yielding spatiotemporal fusion features.

[0048] As an example, suppose you are monitoring an industrial plant with several wind turbines: Each wind turbine is considered a node, and its node features are 32-dimensional feature vectors after cross-scale coupling processing. These features reflect the current operating status of the wind turbine, including information in the time domain, spectrum domain, and time-frequency domain.

[0049] The physical distance between each wind turbine is calculated, and the speed and load data of each wind turbine is collected in real time. For example, if two wind turbines are physically close and have similar speeds, they are considered to be strongly associated and are assigned a higher adjacency matrix weight.

[0050] Dynamically adjust the weights in the adjacency matrix to accommodate changes in different operating conditions. For example, if the load on a particular wind turbine suddenly increases, the vibration transmission relationship between it and its adjacent wind turbines will also change. The system will automatically update the adjacency matrix to reflect this change.

[0051] First, a spatial convolution operation is performed, aggregating node features using a dynamic adjacency matrix to generate spatial fusion features. This step can capture the mutual influence between wind turbines, such as how the fault signal of one wind turbine propagates to other wind turbines.

[0052] Then, a temporal convolution operation is performed, applying a one-dimensional convolution to the spatially fused feature sequence to extract the changing patterns in the time series. This helps identify the development trend of faults over time, such as whether the fault of a particular wind turbine is gradually increasing.

[0053] The gating unit is used to control the flow of features, ensuring that important features are effectively transmitted and avoiding information loss or over-smoothing.

[0054] S400, processing the spatiotemporal fusion features through an occasional fault sensitive timing module, using a Teager energy operator to enhance transient impact features, combining an adaptive comb filter to extract periodic fault features, and outputting timing features through dual-channel weight fusion and self-attention mechanism.

[0055] As mentioned above, the sporadic fault-sensitive timing module is designed to address the characteristics of sporadic faults in group control equipment. It aims to improve fault detection accuracy by enhancing transient impact features and extracting periodic fault features. This module utilizes the Teager Energy Operator (TEO) to enhance transient impact features and incorporates an adaptive comb filter to extract periodic fault features. Finally, time series features are generated through dual-channel weight fusion and a self-attention mechanism.

[0056] Transient impact feature enhancement: The Teager energy operator is used to process spatiotemporal fusion features, highlighting the energy changes of transient impacts. TEO can more sensitively capture the transient impact characteristics in the signal, helping to identify short-term, sudden faults.

[0057] Periodic Fault Feature Extraction: An adaptive comb filter is used to extract periodic fault features. This filter dynamically adjusts its center frequency based on the actual speed of the equipment to accurately capture the periodic vibration characteristics of components such as bearings and gearboxes.

[0058] Dual-channel weighted fusion: Transient impact features and periodic fault features are processed as two independent channels. 1D convolution is used to learn the weights of each channel, and then the features of the two channels are fused according to the weights. This dual-channel design can better separate different types of fault features.

[0059] Self-attention mechanism: Utilizes the self-attention mechanism to focus on features at the moment of fault occurrence, improving temporal resolution. Through the query / key / value generation mechanism and attention weight calculation method, the model can accurately identify the moment of fault occurrence in complex time series.

[0060] For example, suppose you are monitoring a cluster of fans in an industrial plant: The Teager energy operator is used to process the spatiotemporal fusion features of the wind turbine, enhancing transient impact characteristics. For example, if a wind turbine blade is suddenly impacted within a certain time window, resulting in a significant transient impact in the vibration signal, TEO can significantly amplify this feature, making it easier for the system to identify this abnormal event.

[0061] An adaptive comb filter is used to extract periodic fault characteristics of wind turbines. For example, if a wind turbine bearing experiences localized damage, periodic shocks will appear in its vibration signal. The adaptive comb filter dynamically adjusts its center frequency based on the wind turbine's current speed to accurately capture these periodic shocks.

[0062] Transient impact features and periodic fault features are processed as two independent pathways. A 1D convolution is used to learn the weight of each pathway. For example, if transient impact features are found to be more important within a certain time period, a higher weight is assigned, and vice versa. The features of the two pathways are then fused according to the weights to generate a comprehensive time series feature.

[0063] The self-attention mechanism focuses on features at the moment a fault occurs. For example, during wind turbine operation, a minor mechanical failure may occur, causing a small change in the vibration signal. The self-attention mechanism automatically identifies and amplifies the features at that moment, helping the system more accurately locate the time of the fault.

[0064] S500: Based on a dual-threshold detection mechanism and a device contrast enhancement strategy, fault identification and location are performed on the time series features to generate an interpretable diagnostic evidence chain including contributions from multi-scale features.

[0065] As described above, the dual-threshold detection mechanism and device contrast enhancement strategy are the core means for fault identification and location in this invention. They aim to improve the detection and location accuracy of sporadic faults by combining multi-scale feature information. This step, after obtaining the timing features output by the sporadic fault-sensitive timing module, further analyzes and determines the features, generating a physically interpretable diagnostic evidence chain.

[0066] Dual-threshold detection mechanism: Two independent but collaborative thresholds—a transient threshold and a sustained threshold—are set to identify short-duration, impact-type faults (such as instantaneous impact and looseness) and periodic / persistent faults (such as bearing wear and gear cracks), respectively. When the timing characteristics within a certain period exceed both thresholds simultaneously, the system identifies a true fault event. If only the transient threshold is exceeded, it is considered a non-fault disturbance, avoiding false alarms.

[0067] Device comparison enhancement strategy: This strategy compares the abnormality level of the current device with the average abnormality level of other similar devices in its group and calculates the deviation value. This strategy effectively distinguishes true individual device failures from group operating fluctuations (such as sudden load changes and environmental vibrations), thereby enhancing the representation of weak fault signals and improving diagnostic sensitivity.

[0068] Generation of an explainable diagnostic evidence chain: After fault identification and location, the system automatically generates a structured, multi-layered diagnostic evidence chain, including: The multiple times the impact energy exceeded the standard at the time of the fault; the characteristic frequency peak and its changing trend; the correlation offset with other equipment; the contribution weight of multi-scale features (time domain, spectrum domain, time-frequency domain) in the decision-making; this information together constitutes a diagnostic basis with clear logic and physical meaning, which is easy for operation and maintenance personnel to understand and verify.

[0069] For example, suppose in a fan group control system, a fan experiences occasional vibration shock due to a small crack in the bearing: The system detected that the wind turbine had an obvious transient impact at a certain moment, and its transient energy characteristics exceeded the preset "transient threshold"; At the same time, after adaptive comb filtering, it was found that the device also showed periodic feature enhancement in the corresponding time period, and the amplitude exceeded the "persistence threshold"; Because both conditions are met at the same time, the system determines it as a real fault event and triggers an alarm; if only the transient exceeds the standard, it may be identified as external interference (such as foreign object impact) and will not trigger a high-level alarm.

[0070] Calculate the anomaly score for the wind turbine (based on the degree to which the time series characteristics deviate from the normal model); At the same time, the average abnormality scores of the remaining 10 wind turbines in the same group were calculated; If the abnormal score of the wind turbine is significantly higher than the group mean (for example, higher than 2 standard deviations), it is further confirmed that it is an isolated fault source rather than caused by changes in the overall operating environment; This deviation result is used to improve the fault confidence of the device and assist in precise positioning.

[0071] Diagnostic evidence chain output example: "Failure time: April 5, 2025, 14:23:17"; “Main features: impact energy exceeded the standard by 3.2 times, and the amplitude of the characteristic frequency (BPFO) increased by 2.8 times”; "Correlation analysis: The correlation with adjacent equipment decreased by 40%, indicating strong vibration independence"; “Feature contribution: time domain features account for 45%, spectrum domain accounts for 35%, and time-frequency domain accounts for 20%”; "Recommended action: It is recommended to shut down and inspect the outer ring of #3 bearing for early cracks."

[0072] This chain of evidence not only provides judgment results, but also reveals the basis for judgment, greatly enhancing the credibility and practicality of the system.

[0073] According to one embodiment of the present application, the vibration signal of the group control device is multi-scale decomposed through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal to generate multi-dimensional feature data, specifically: Use IEPE accelerometers to collect triaxial vibration signals and calculate transient impact characteristics such as peak value, pulse factor, and impact duration; waveform factor and skewness waveform distribution characteristics; high-frequency energy proportion and zero-crossing rate energy change characteristics; Extract peak frequency bandwidth, sideband energy ratio, and sideband frequency interval modulation characteristics through envelope analysis; The wavelet transform method is used to generate multi-scale energy entropy features, sub-band energy features and time-frequency domain dynamic features.

[0074] As mentioned above, the IEPE accelerometer installed on the group control equipment first collects vibration signals in three orthogonal directions during operation, known as triaxial vibration signals. This sensor has excellent frequency response and anti-interference capabilities, enabling it to accurately capture the vibration state of the equipment under different operating conditions.

[0075] After obtaining the original vibration signal, time domain feature extraction is performed. The peak value of the signal is calculated to reflect the maximum vibration amplitude; the pulse factor, that is, the ratio of the peak value to the absolute average value, is calculated to identify whether there is impact vibration; the impact duration is extracted to characterize the duration of the impact event. The above constitutes transient impact features. At the same time, the waveform factor, that is, the ratio of the root mean square value to the absolute average value, is extracted to describe the steepness of the signal waveform; the skewness is calculated to reflect the asymmetry of the signal amplitude distribution relative to the normal distribution, which is used to determine whether the vibration is abnormally skewed. In addition, energy change features are extracted, including the high-frequency energy ratio, that is, the ratio of energy in the high-frequency band (such as above 2kHz) to the total energy, which is used to identify the increase in high-frequency components caused by early weak faults; the zero-crossing rate is calculated, that is, the number of times the signal crosses the zero point per unit time, which reflects the fluctuation frequency of the signal.

[0076] Next, spectral domain feature extraction is performed. Envelope analysis is performed on the original vibration signal, specifically by performing bandpass filtering and Hilbert transform on the signal to obtain the envelope signal, followed by spectral analysis of the envelope signal. The peak frequency bandwidth is extracted, i.e., the frequency range occupied by the main frequency peak in the envelope spectrum, reflecting the frequency concentration of the fault impact. The sideband energy ratio is extracted, i.e., the ratio of the energy of the sideband frequency component to the energy of the center frequency, which is used to identify modulation faults, such as gear meshing faults. The sideband frequency interval is extracted, i.e., the frequency difference between adjacent sidebands, which corresponds to the fault characteristic frequency and helps determine the fault type and location.

[0077] Finally, time-frequency domain feature extraction is performed. Wavelet transforms are used to perform multi-scale decomposition of the original vibration signal, breaking it down into multiple sub-signals in different frequency bands. Based on the signal energy of each sub-band, multi-scale energy entropy is calculated to measure the uniformity of energy distribution at different scales. Energy concentration is often seen when a fault occurs. The energy values ​​of each sub-band are extracted to form sub-band energy signatures, which are used to identify frequency bands with abnormally concentrated energy. Simultaneously, the energy variation trends over time at each scale are analyzed to extract dynamic features in the time-frequency domain, which are used to capture the evolution of non-stationary and transient faults.

[0078] According to one embodiment of the present application, the multi-dimensional feature data is dynamically interactively fused through the cross-scale feature coupling module, and the recognition of weak fault features is enhanced by a bidirectional gating mechanism, and the fused 32-dimensional feature vector is output, specifically: performing standardization processing on the multidimensional feature data; The standardized multi-dimensional feature data is spliced ​​in the channel dimension, and the importance weights of the features at each scale are learned through one-dimensional convolution and softmax function; The bidirectional gating mechanism is used to achieve complementary enhancement of features between scales, including: Strengthen the correlation between time domain features and frequency domain feature frequencies; Strengthen the relationship between frequency domain features and time-frequency features; The residual connection method is used to retain the key information of the original features and output the fused 32-dimensional feature vector.

[0079] As mentioned above, the multidimensional feature data consisting of time-domain features, spectral-domain features, and time-frequency-domain features extracted from the multi-scale input layer is first normalized. This process involves normalizing each feature type by either mean or minimum and maximum values, ensuring that features of different dimensions and value ranges remain consistent in numerical scale. This prevents a single feature type from dominating the fusion process due to excessive amplitude, ensuring comparability and fairness among features at all scales in subsequent fusion.

[0080] The normalized multidimensional feature data is then concatenated along the channel dimension to form a high-dimensional feature vector. This concatenation operation integrates features from different time, spectrum, and time-frequency domains into a unified input representation, providing a foundation for subsequent weight learning and feature fusion.

[0081] Next, the concatenated features are processed through a one-dimensional convolution operation to extract local correlations across scales and generate importance scores for features at each scale. These scores are then input into a softmax function to calculate a normalized weight for each feature scale. This weight reflects the relative importance of each scale for fault identification under the current operating conditions, enabling adaptive weighting of different features.

[0082] On this basis, a bidirectional gating mechanism is introduced to achieve dynamic interaction and complementary enhancement between features at different scales. Specifically, this involves two gating processes: first, strengthening the correlation between time-domain and spectral-domain features. Through the gating structure, the modulation characteristics of strong impulse signals in the frequency domain are enhanced, improving the response to transient faults. Second, strengthening the correlation between spectral-domain and time-frequency-domain features. Using the gating mechanism, sideband information in the frequency domain is dynamically coupled with subband energy from wavelet decomposition, enhancing the ability to capture non-stationary fault characteristics. By controlling the flow of information, the gating mechanism selectively retains or amplifies characteristic responses related to faults.

[0083] Furthermore, a residual connection is used during feature fusion to directly introduce the original multi-dimensional feature information or the backbone features of the intermediate layer into the output. This structure effectively preserves the key information in the input features, preventing the loss of important features due to nonlinear transformations or gating operations during deep network transmission, thereby improving the stability and expressiveness of the model.

[0084] Finally, through normalization, concatenation, weight learning, bidirectional gated fusion, and residual connections, the high-dimensional features are compressed and mapped into a 32-dimensional low-dimensional fused feature vector. This vector integrates key information from the time, spectral, and time-frequency domains, and through dynamic weighting and interactive enhancement mechanisms, it highlights the expression of weak fault features, providing highly discriminative input features for subsequent graph convolutional modeling and fault diagnosis.

[0085] According to one embodiment of the present application, the 32-dimensional feature vector is used to construct a convolutional model of the relationship perception graph between devices, and a dynamic adjacency matrix is ​​generated based on the physical distance of the devices and the similarity of the real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features, specifically: The sensor of each device is a graph node, and the node feature is the 32-dimensional feature vector; Dynamically adjust the adjacency matrix weights based on the physical distance between devices and the similarity of real-time working conditions; The physical distance factor is multiplied by the speed difference factor to generate a physical similarity matrix; The dynamic fusion coefficient α is adjusted according to the physical distance between devices. The smaller the α value, the higher the adjacency weight; The vibration transmission law between devices is captured through spatial convolution operation, and the correlation features in the time series are extracted through temporal convolution operation. The feature flow is controlled by the gating unit to alleviate the over-smoothing problem.

[0086] As mentioned above, we first construct the basic topology of the graph model by treating the sensor of each device in the group control system as a node in the graph structure. The feature input of each node is the 32-dimensional feature vector output from the previous step. This vector contains the comprehensive feature information of the device's vibration signal in the time domain, spectral domain, and time-frequency domain, and can effectively represent the current operating status of the device.

[0087] On this basis, a dynamic adjacency matrix is ​​constructed to describe the connection relationships and weights between device nodes. This adjacency matrix is ​​not fixed but dynamically adjusted based on the physical distance between devices and their real-time operating conditions. Specifically, the physical distance between devices is first calculated to form a physical distance factor. Simultaneously, the real-time speed data of each device is obtained, and the speed difference between each pair of devices is calculated to form a speed difference factor. The physical distance factor is multiplied by the speed difference factor to obtain a physical similarity matrix, which reflects the comprehensive proximity of devices in terms of spatial location and operating status.

[0088] To further optimize the distribution of adjacency weights, a dynamic fusion coefficient α is introduced, which adaptively adjusts based on the physical distance between devices. When two devices are physically close, α is set to a smaller value, indicating a higher likelihood of vibration transmission between them, thus assigning a higher adjacency weight. Conversely, when the devices are farther apart, α increases, and the adjacency weight decreases accordingly. This mechanism enables the adjacency matrix to more realistically reflect the actual strength of connections between devices under different installation layouts and operating environments.

[0089] After constructing the dynamic adjacency matrix, a spatiotemporal convolution operation is performed. First, spatial convolution is performed, using a graph convolutional network (GCN) to aggregate node features on the graph structure. This means that each node's features are updated to the weighted sum of its own and its neighboring node features, with the weights determined by the dynamic adjacency matrix. This process captures the vibration transmission patterns between devices and identifies the potential propagation paths and impact ranges of fault signals.

[0090] Subsequently, a temporal convolution operation is performed along the time dimension. The node feature sequence obtained after spatial convolution is subjected to a one-dimensional convolution along the time axis to extract the pattern of feature evolution over time and identify the dynamic trend of device status. This operation effectively captures the evolution of features before and after a fault occurs, enhancing the time sensitivity of occasional faults.

[0091] Throughout the spatiotemporal convolution process, a gating unit is introduced to control the transmission of features. By adjusting the ratio of information updated and retained, the gating unit prevents feature smoothing caused by multi-layer graph convolution. This prevents features from becoming similar after multiple aggregations, leading to a loss of discriminative power. This gating mechanism ensures that key features are effectively retained and enhanced during the propagation process.

[0092] Finally, after the joint processing of spatial convolution and temporal convolution, spatiotemporal fusion features containing spatial correlation and temporal evolution information between devices are obtained, providing sufficient contextual support for subsequent fault-sensitive feature extraction and diagnostic decisions.

[0093] According to one embodiment of the present application, the spatiotemporal fusion features are processed by the sporadic fault sensitive timing module, the transient impact features are enhanced by the Teager energy operator, the periodic fault features are extracted by combining the adaptive comb filter, and the timing features are output by dual-channel weight fusion and self-attention mechanism, specifically: The transient impact energy is calculated by the Teager energy operator, and the high-frequency characteristics are quantified by combining the short-time zero-crossing rate to generate transient features. Extract periodic fault features through adaptive comb filtering, where the center frequency of the filter is dynamically adjusted according to the fault type; The weights of transient features and periodic features are learned through one-dimensional convolution, and the features of the two pathways are fused according to the weights. The attention weight at the moment of fault occurrence is calculated through the self-attention mechanism to improve the temporal resolution.

[0094] As mentioned above, first, for the transient impact signals commonly seen in sporadic faults, the Teager energy operator is used to process the time series in the spatiotemporal fusion features. The Teager energy operator can effectively reflect the instantaneous energy changes of the signal, and is particularly sensitive to short-term, high-intensity impact components. The transient impact energy at each time point is calculated using this operator, highlighting the energy mutation characteristics at the moment the fault occurs. At the same time, the high-frequency vibration characteristics are quantified in combination with the short-term zero-crossing rate, that is, the number of times the signal crosses the zero point is counted within the sliding time window to identify high-frequency oscillation behavior caused by local damage. The Teager energy output is combined with the short-term zero-crossing rate result to form a feature vector describing the transient impact characteristics, which serves as one of the inputs for subsequent analysis.

[0095] Secondly, to capture the periodic impact characteristics common in equipment failures (such as bearing fault frequency and gear meshing frequency), an adaptive comb filter is used to process the spatiotemporal fusion features. This comb filter has multiple equally spaced passbands, capable of simultaneously enhancing multiple harmonic frequency components. Its key parameters—center frequency and frequency spacing—are dynamically adjusted based on the equipment type and fault mode. For example, when diagnosing a rolling bearing, the system calculates the theoretical fault characteristic frequency based on the bearing geometry and real-time speed and uses this as the comb filter's baseband frequency. For gears, the meshing frequency and its sidebands are determined based on the number of teeth and speed. The filter passband is updated in real time with operating conditions, ensuring accurate extraction of periodic fault components and generation of the corresponding periodic fault feature sequence even under varying speed conditions.

[0096] Subsequently, the generated transient features and periodic fault features are treated as two independent processing paths and subjected to dual-channel fusion. The features of the two paths are weighted and learned through a one-dimensional convolutional network, automatically identifying which type of features are more discriminative under different fault stages or operating conditions. For example, when transient shocks dominate the initial stage of the fault, the system assigns a higher weight to transient features; when the fault develops into a stable periodic shock stage, the fusion ratio of periodic features is increased. Finally, based on the learned weight coefficients, the features of the two paths are weighted and fused to obtain a comprehensive feature representation that combines transient sensitivity with periodic recognition capabilities.

[0097] Finally, a self-attention mechanism is introduced based on the fused features to further enhance the ability to locate the moment of fault occurrence. The self-attention mechanism generates an attention weight distribution by calculating the correlation between each time step in the feature sequence. This weight reflects the importance of each time point to the feature expression at the current moment, enabling the model to automatically focus on the period when the abnormal event occurs. For example, when a significant energy surge or periodic increase occurs at a certain moment, the attention weight at that moment will be significantly amplified, thereby enhancing the expression strength of the feature at that moment and improving the temporal resolution. After the attention weighted processing, the final time series feature is output. This feature not only contains rich fault mode information but also highlights the abnormal performance at key time points, providing high-precision input for subsequent fault identification and location.

[0098] According to one embodiment of the present application, the dual-threshold detection mechanism and device contrast enhancement strategy are used to identify and locate faults on the time series features, generating an interpretable diagnostic evidence chain that includes multi-scale feature contributions, specifically: Set transient thresholds and continuous thresholds to correspond to the abnormal judgment of short-term impact characteristics and periodic characteristics respectively; Thresholds are dynamically updated every 24 hours based on the latest normal data to adapt to baseline drift caused by equipment aging; Through the device comparison enhancement strategy, the deviation between the abnormality of a single device and the average abnormality of the group of devices is calculated, and the abnormal offset is generated to enhance the characteristics of weak faults.

[0099] As mentioned above, a dual-threshold detection mechanism is first established to determine whether the equipment's operating status is abnormal. This mechanism includes two independent but complementary thresholds: a transient threshold and a sustained threshold. The transient threshold is used to detect short-term, high-intensity impact anomalies, corresponding to the transient impact characteristics output by the Teager energy operator. When the transient energy exceeds this threshold within a certain period of time, it indicates the possibility of sudden mechanical shock or local damage. The sustained threshold is used to identify periodic or persistent fault characteristics, corresponding to the periodic fault components extracted by the adaptive comb filter. When the amplitude of the periodic characteristic remains above this threshold for a certain period of time (such as for multiple consecutive time windows), the equipment is considered to have a stable fault trend. Only when both the transient and sustained characteristics exceed their respective thresholds does the system determine it as a valid fault event, avoiding misjudging a single transient interference as a true fault.

[0100] To accommodate vibration baseline drift caused by factors such as aging, lubrication changes, and loose installation during long-term equipment operation, the threshold is not fixed but rather dynamically updated. The system recalculates the threshold every 24 hours based on the most recent normal operating data collected. This method involves selecting time series feature data marked as "normal," analyzing the distribution characteristics of its transient energy and periodicity (such as mean and standard deviation), and updating the transient and sustained thresholds accordingly. For example, the mean plus twice the standard deviation can be used as the new threshold. This ensures that the threshold adjusts adaptively as the equipment's health status slowly changes, maintaining detection sensitivity and stability.

[0101] On this basis, a device comparison enhancement strategy is introduced to further improve the ability to identify weak faults. This strategy identifies individual anomalies by comparing the operating status differences between the target device and other devices in the same group. Specifically, the abnormality index of each device is first calculated. This index is determined by the degree to which its time series characteristics deviate from the normal model. For example, it can be calculated based on the characteristic amplitude, rate of change, or Euclidean distance from the historical benchmark. Then, the average value of the abnormality of all devices in the group is calculated as the group benchmark under the current working conditions. The abnormality of a single device is compared with the average value to calculate its deviation value, that is, the abnormal offset. If the abnormal offset of a device is significantly higher than that of other devices (such as more than 1.5 times the group standard deviation), it is considered that the device has an independent fault risk rather than being affected by the overall environmental fluctuations. This abnormal offset is used to enhance the fault confidence of the device, thereby effectively amplifying the performance of weak fault signals in the diagnosis results.

[0102] Finally, upon completing fault identification and location, the system generates an interpretable diagnostic evidence chain encompassing the contributions of multi-scale features. This evidence chain is output in a structured format and includes: the time of fault occurrence, the magnitude of the transient impulse energy exceeding the specified value, the periodic characteristic frequency and its amplitude variation, the magnitude of the abnormal offset, and the relative contributions of time-domain, spectral-domain, and time-frequency-domain features to the diagnosis. This evidence chain not only provides the fault diagnosis result but also reveals the basis for the judgment, facilitating operations and maintenance personnel's understanding of the diagnostic logic and supporting subsequent maintenance decisions, thus achieving a transition from "black-box early warning" to "transparent diagnosis."

[0103] According to one embodiment of the present application, it further includes: Calculate the fault propagation impact based on the dynamic adjacency matrix and determine the fault source device; Generate an interpretable evidence chain that includes multi-scale feature contributions, including the impact energy exceedance multiple, characteristic frequency peak and its change multiple, and correlation changes with neighboring equipment; Output a structured diagnostic report including fault location, evidence chain, and repair suggestions.

[0104] As mentioned above, first, based on the dynamic adjacency matrix constructed in the previous step, the degree of propagation of the fault between the device groups is calculated to determine the most likely fault source device. Specifically, after detecting an abnormality in one or more devices, the correlation strength and directionality information between devices described by the dynamic adjacency matrix is ​​used to reversely analyze the propagation path of the abnormal characteristics. By comparing the chronological order of the abnormalities of each device, the trend of changes in the characteristic intensity, and its connection relationship in the graph structure, the possibility score of each device as the source of the fault is calculated. For example, if a device has an abnormality the earliest in time, and its abnormal characteristics are propagated to multiple downstream devices through high-weight edges, then the device is judged to be a high-impact node and is very likely to be the source of the fault. This process combines the graph propagation model with time series correlation analysis to improve the accuracy of fault location.

[0105] Secondly, an interpretable diagnostic evidence chain is generated to clearly present the basis for fault judgment. This evidence chain not only includes the final diagnostic results, but also integrates key indicators from the original signal to the feature fusion process to ensure that the diagnostic process is transparent and reliable. Specific contents include: the impact energy exceedance multiple, that is, the multiple of the current transient energy relative to the normal baseline, reflecting the severity of the fault impact; the characteristic frequency peak and its change multiple, describing the increase in the amplitude of the frequency component related to the specific fault type in the frequency domain, used to support the judgment of the fault type; and the change in correlation with adjacent equipment. By comparing the correlation coefficient or synergy index of the vibration signals between the equipment before and after the fault, the coupling relationship change caused by the fault is revealed, which assists in judging whether the fault has local independence. The above multi-scale feature contributions are sorted by weight or significance to form a logically coherent evidence sequence.

[0106] Finally, the system outputs a structured diagnostic report, serving as a direct basis for human-machine interaction and operational decision-making. This report, presented in a standardized format, includes three core components: First, the fault location, which clearly identifies the device number, installation location, and possible faulty components (such as bearings and gearboxes); second, the diagnostic evidence chain, which lists supporting indicators and their values ​​in an itemized format to facilitate technicians' review of the judgment logic; and third, remediation recommendations, which recommend appropriate actions based on the fault type and severity, such as "Stop and inspect fan #4 drive-end bearing" or "Prioritize lubrication maintenance and monitor subsequent impact energy trends." The report can be output in a variety of formats, including text, PDF, or integrated into industrial monitoring platforms to meet the needs of different scenarios.

[0107] According to one embodiment of the present application, generating the dynamic adjacency matrix includes the following steps: Obtain the physical distance data between group control devices and build a basic distance matrix; Collect the operating condition data of each device in real time and calculate the speed difference and load change similarity between two devices; The distance matrix and the working condition similarity matrix are weightedly fused to generate a dynamic adjacency matrix, in which the weight coefficients are pre-set or adjusted through online learning according to the equipment type and installation environment.

[0108] As mentioned above, first, obtain the physical distance data between each device in the group control system. This data can be obtained from device installation drawings, on-site measurements, or positioning systems (such as UWB and RFID), reflecting the relative positions of the devices in the spatial layout. Based on this distance data, a basic distance matrix is ​​constructed, where each element of the matrix represents the straight-line distance or installation path distance between any two devices. This matrix serves as a basic measure of the physical connection between devices, reflecting the physical possibility of vibration signal propagation between them. Closer devices generally have stronger mechanical coupling and vibration transmission relationships.

[0109] Secondly, the operating condition data of each device is collected in real time, including key parameters such as speed, load, and operating temperature. Based on this data, the similarity of the operating status between any two devices is calculated. Specifically, by analyzing the time series of the device speed, the degree of difference is calculated. For example, the normalized difference or dynamic time warping method is used to measure the consistency of the speed change; at the same time, the similarity of the load change is evaluated to determine whether the devices are in the same or similar operating mode. The similarity indicators of speed difference and load change are combined into a working condition similarity matrix, which reflects the degree of coordination of the equipment in its operating behavior. Equipment with more similar working conditions may show stronger correlation in vibration characteristics.

[0110] Then, the basic distance matrix and the operating condition similarity matrix are weightedly fused to generate the final dynamic adjacency matrix. During the fusion process, the elements of the two matrices are weighted and combined item by item, and the weight coefficient is used to adjust the relative importance of physical distance and operating conditions in the overall correlation assessment. The weight coefficient can be pre-set according to the equipment type and installation environment. For example, in scenarios where vibration transmission mainly depends on the physical structure (such as a shared foundation), the weight of the distance term is increased; and in scenarios where operating synchronization dominates the correlation (such as parallel operation of pump groups), the weight of the operating condition similarity is increased. In addition, the weight coefficient can also be automatically adjusted according to historical fault data or normal operation samples through an online learning mechanism to achieve adaptive optimization, so that the adjacency matrix can dynamically evolve with the system operating status.

[0111] The resulting dynamic adjacency matrix not only reflects the static spatial relationship between devices but also incorporates the influence of real-time operating status. It can more accurately characterize the dynamic correlation structure of vibration characteristics between group-controlled devices and provide reliable spatial topology support for subsequent graph convolution operations.

[0112] An embodiment of a second aspect of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method of any one of the embodiments of the first aspect when executing the program.

[0113] Figure 2 An example of a physical structure diagram of an electronic device is shown below. Figure 2 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the method in any embodiment of the first aspect above, the method including: The vibration signal of the group control equipment is decomposed at multiple scales through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal and generate multi-dimensional feature data; Dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector; Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed through the occasional fault sensitive timing module, the Teager energy operator is used to enhance the transient impact features, and the adaptive comb filter is combined to extract the periodic fault features. The dual-channel weight fusion and self-attention mechanism are used to output the timing features. Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0114] Furthermore, the logic instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as standalone products, stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for causing a computer device (such as a personal computer, server, or network device) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories, random access memories, magnetic disks, or optical disks.

[0115] On the other hand, the present invention further provides a computer program product, comprising a computer program. The computer program may be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the method provided by each of the above methods, including: The vibration signal of the group control equipment is decomposed at multiple scales through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal and generate multi-dimensional feature data; Dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector; Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed through the occasional fault sensitive timing module, the Teager energy operator is used to enhance the transient impact features, and the adaptive comb filter is combined to extract the periodic fault features. The dual-channel weight fusion and self-attention mechanism are used to output the timing features. Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0116] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for cigarette box image recognition provided by the above methods is implemented, and the method includes: The vibration signal of the group control equipment is decomposed at multiple scales through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal and generate multi-dimensional feature data; Dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector; Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed through the occasional fault sensitive timing module, the Teager energy operator is used to enhance the transient impact features, and the adaptive comb filter is combined to extract the periodic fault features. The dual-channel weight fusion and self-attention mechanism are used to output the timing features. Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0117] Example 2 Step 1: Multi-scale input layer construction: The multi-scale input layer is designed with a three-stream parallel processing structure to address the multi-scale characteristics of rotating machinery fault signals, such as transient impact, modulation characteristics, and steady-state vibration. This structure aims to comprehensively cover various fault characteristics during equipment operation and provide rich and effective original feature data for subsequent feature coupling and analysis.

[0118] Scale 1: Original signal branch (time domain features): An IEPE accelerometer (±50g range) was used to collect the equipment's three-axis vibration signals (X / Y / Z directions). The sampling frequency was set to 12.8 kHz. This sampling frequency was chosen based on the Nyquist criterion to fully cover the critical fault frequencies of rotating machinery within the 0-5 kHz range.

[0119] The collected original signal is processed using a second-order Butterworth high-pass filter (cut-off frequency 1Hz) to eliminate the DC component interference caused by the sensor zero drift; the 3σ criterion is used to detect and replace abnormal points in the signal. When the signal value meets When , the abnormal point is replaced by the mean value of the signal, where is the mean of the signal, σ is the standard deviation of the signal; the above signal is normalized and mapped to the range of [-1,1].

[0120] The signal time domain features are extracted from three dimensions: impact intensity, distribution form, and energy change. The specific composition is as follows: 1. Transient impact characteristics (4 dimensions): Peak value: The maximum value of the signal within the time window, which directly reflects the instantaneous impact intensity (such as the instantaneous vibration peak value of a bearing crack).

[0121] Kurtosis: By formula Calculate and quantify the degree of deviation of the impulse component in the signal (normal signal kurtosis is ≈3, and can reach more than 5 in the case of fault).

[0122] Pulse Factor: The ratio of peak value to signal average value, which highlights high-impact signals in a low-amplitude background.

[0123] Impact duration: The number of consecutive sampling points where the signal exceeds the threshold (3σ of normal signal), reflecting the time span of the fault impact.

[0124] 2. Waveform distribution characteristics (3D): Form factor: The ratio of the root mean square value to the average value, which describes the smoothness of the signal waveform (the waveform of wear fault is relatively smooth, and the waveform of crack fault is steep).

[0125] Skewness: describes the symmetry of the signal amplitude distribution. During a fault, the absolute value of the skewness increases due to the unidirectionality of the impulse signal.

[0126] Peak frequency: The dominant frequency of the impact signal obtained by short-time Fourier transform, which is associated with the natural frequency of the faulty component (such as the impact frequency of a bearing outer ring fault).

[0127] 3. Energy change characteristics (3D): High-frequency energy ratio: The ratio of the energy in the 2-5kHz frequency band to the total energy in the 0-5kHz frequency band, highlighting the high-frequency impact components (such as high-frequency vibrations caused by crack propagation).

[0128] Energy entropy: calculates the energy value within a 10ms window , where To normalize the signal value, the entropy value is calculated based on the energy distribution of 100 sub-windows to reflect the uniformity of energy distribution (energy concentration during a fault causes a decrease in entropy value).

[0129] Zero-crossing rate: The number of signal zero-crossing points per unit time, indirectly reflecting the proportion of high-frequency components in the signal (the zero-crossing rate of the impact signal is significantly higher than that of the normal signal).

[0130] After feature extraction, the output dimension of this branch is F1=Nsensors×T×10, where Nsensors is the number of sensors, T is the time step, and 10 is the frequency feature dimension.

[0131] Scale 2: Envelope spectrum branches (frequency domain characteristics): This branch mainly focuses on the modulation characteristics of fault signals (such as sideband modulation of bearing faults) and extracts characteristic frequency components through envelope analysis, thereby achieving accurate identification of equipment faults.

[0132] Convert the real signal into an analytical signal, separate the envelope signal and the carrier signal, and the calculation formula is:

[0133] Where, is the original signal, is the Hilbert transform result, The envelope signal is an envelope signal. Through Hilbert transform, the envelope information in the signal can be effectively extracted, highlighting the modulation characteristics caused by the fault.

[0134] A 5th-order Butterworth low-pass filter (cut-off frequency 2kHz) is used to smooth the envelope signal and remove the interference of high-frequency noise, so that the envelope signal can more clearly reflect the fault characteristics. A 2048-point fast Fourier transform (FFT) is performed on the envelope signal after low-pass filtering to obtain the envelope spectrum. .

[0135] The characteristic frequency band intercepts the 0-2kHz frequency band (this frequency band covers key fault frequencies such as bearing BPFO / BPFI and gear meshing frequency) and extracts the following features: 1. Basic frequency characteristics (3D): Characteristic frequency peak (1 dimension): , which directly corresponds to the core frequency of the fault (such as bearing BPFO and gear meshing frequency) and is the core basis for judging the fault type.

[0136] Peak frequency energy (1D): energy value at the peak of the characteristic frequency , quantifies the strength of the fault signal.

[0137] Peak frequency bandwidth (1D): The 3dB bandwidth on both sides of the characteristic frequency peak reflects the degree of dispersion of the fault frequency (wear-related faults have a wider bandwidth, while crack-related faults have a narrower bandwidth).

[0138] 2. Sideband features (2D): Sideband energy ratio (1D): The ratio of the energy within the characteristic frequency range ±100 Hz to the total energy. This is used to evaluate the modulation depth (such as the sideband energy enhancement caused by gear eccentricity).

[0139] Sideband frequency interval (1D): The frequency difference between adjacent sidebands corresponds to the device rotation frequency (such as the modulation frequency related to motor speed) and helps locate the source of the fault (such as faults synchronized with the speed).

[0140] 3. Spectrum distribution characteristics (3D): Spectral entropy (1 dimension): The calculation formula is , where Frequency When a fault occurs, the spectral entropy value will decrease because the fault will cause the energy to be concentrated at a specific frequency.

[0141] After feature extraction, the output dimension of this branch is F2=Nsensors×T×6, where Nsensors is the number of sensors, T is the time step, and 8 is the frequency feature dimension.

[0142] Scale 3: Wavelet energy entropy branch (time-frequency characteristics): Wavelet packet decomposition is used to capture the non-stationary characteristics of the signal, which is suitable for analyzing the time-frequency distribution of sporadic faults. The wavelet packet basis function uses the db4 wavelet basis as the decomposition basis function, and a three-layer decomposition is used to decompose the signal into 8 sub-bands. The center frequencies of each sub-band are 312.5Hz, 937.5Hz, 1562.5Hz, 2187.5Hz, 2812.5Hz, 3437.5Hz, 4062.5Hz, and 4687.5Hz. Wavelet packet reconstruction is performed on each sub-band to obtain ,in, is the number of decomposition layers, Number the sub-band.

[0143] In the wavelet energy entropy branch of scale 3, the output dimension is 16-dimensional time-frequency features, which are designed from three dimensions: energy distribution, frequency resolution, and time evolution. The specific composition is as follows: 1. Multi-scale energy entropy characteristics (8 dimensions): Based on the 8 sub-bands obtained by 3-layer wavelet packet decomposition, the energy entropy of each sub-band is calculated: ,in, For the Tier The energy proportion of the sub-band. When a device fails, energy is concentrated in a specific sub-band, resulting in a decrease in energy entropy. 2. Sub-band energy characteristics (6 dimensions): Filter out the six key frequency bands that are most strongly associated with the fault from the eight sub-bands (determined by mutual information analysis) and extract their energy values: These six-dimensional features quantify the vibration energy intensity at different frequency bands. 312.5Hz and 937.5Hz correspond to gear meshing frequencies and harmonics, with energy changes reflecting gear wear. 2187.5Hz and 2812.5Hz cover the passing frequency of bearing rolling elements, with sudden energy increases indicating bearing defects. 4062.5Hz and 4687.5Hz are used to capture high-frequency impacts (such as instantaneous crack growth).

[0144] 3. Time-frequency domain dynamic characteristics (2D): Energy entropy change rate: Calculate the average change rate of 8 energy entropies in adjacent time windows. The formula is:

[0145] This feature can sensitively capture sudden changes in energy distribution when a fault occurs (such as a sudden drop in entropy caused by an occasional shock).

[0146] High-frequency energy ratio: The ratio of the sum of the energies of the two high-frequency sub-bands (4062.5 Hz and 4687.5 Hz) to the total energy, highlighting the high-frequency characteristics of transient impacts.

[0147] After feature extraction, the output dimension of this branch is F3=Nsensors×T×16, where Nsensors is the number of sensors, T is the time step, and 16 is the frequency feature dimension.

[0148] Step 2: Cross-scale feature coupling module: The cross-scale feature coupling module is the key to solving the problem of weak fault feature drowning. Its core function is to achieve dynamic interaction and complementary enhancement of features at different scales through a bidirectional gating mechanism, breaking through the limitations of "feature conflict" and "information loss" of traditional multi-scale methods and improving the recognition of fault features.

[0149] To ensure that the features of the three scales are consistent in magnitude and facilitate subsequent feature coupling processing, the features of each scale are standardized using the formula:

[0150] Where, is the s-th scale feature, is the mean of the s-th scale feature, is the standard deviation of the s-th scale feature, and the standardized feature It will serve as the input of the cross-scale coupling module.

[0151] Since the feature dimensions of the three scales are different, in order to achieve effective interaction between the features, their dimensions need to be unified. 、 、 The dimensions are unified to C=32 through 1D convolution (kernel size 3, stride 1), and the calculation formula is:

[0152] Where, is a 1D convolution operation, Indicates the number of channels of the convolution output.

[0153] The three scale features after dimension alignment are concatenated in the channel dimension to obtain:

[0154] Where, is the number of devices, is the time step, and 96 is the feature dimension after concatenation (32×3).

[0155] The importance weight of each scale feature is learned through 1D convolution and softmax function. The formula is:

[0156] Where, , 、 、 They are the corresponding three weights, and satisfy .

[0157] The learned weights are used to couple features of different scales in a bidirectional manner to achieve complementary enhancement of features, where: Scale 1-2 coupling: Strengthen the correlation between the time domain impact characteristics and the frequency domain characteristic frequencies, the formula is:

[0158] Scale 2-3 coupling: Strengthen the correlation between frequency domain features and non-stationary time-frequency features. The formula is:

[0159] Where, Represents the element-wise multiplication operation. By adjusting the weights, important features can dominate the coupling process, achieving effective feature enhancement.

[0160] In order to avoid losing the key information of the original scale features during the feature coupling process, the residual connection method is used to add the original scale features and the coupled features to obtain the final output features. The formula is:

[0161] After cross-scale feature coupling processing, the dimension of the output feature is Nsensors×T×32.

[0162] Step 3: Device-to-Device Relationship-Aware Graph Convolution: The core function of the convolutional layer of the device relationship perception graph is to model the vibration transmission relationship between group control devices by constructing a dynamic adjacency matrix, solving the defect that traditional static graph models cannot adapt to changes in equipment operating conditions, thereby more accurately capturing the spatiotemporal correlation between devices and improving the accuracy of fault diagnosis. 1. Graph model definition: The sensor of each device is regarded as a node of the graph, and the node feature is the feature after cross-scale coupling, that is,

[0163] Where, For the The characteristics of the nodes, For the Cross-scale coupling characteristics of the device, T is the time step, and 32 is the feature dimension.

[0164] 2. Dynamic adjacency matrix construction: The vibration transmission intensity between devices is calculated using a dynamic adjacency matrix. , where N is the number of devices, The value range of is [0,1], The larger the value, the better the device With equipment The stronger the vibration correlation between them.

[0165] Here, the dynamic adjacency matrix By physical similarity matrix Similarity matrix with features The weighted fusion is:

[0166] Where, is the dynamic fusion coefficient, whose value is adjusted in real time with the operating status of the device, and is used to balance the contribution of physical similarity and feature similarity in the adjacency matrix.

[0167] Physical similarity matrix: The main considerations are the impact of physical distance and rotation speed differences between devices on vibration transmission.

[0168] Physical distance factor: based on the device's installation coordinates computing devices With equipment Euclidean distance between , and then quantify the effect of physical distance on vibration transmission through the exponential decay model, the formula is:

[0169] Where, is the average distance between devices in the factory (5m in the experiment). This formula shows that the closer the distance between devices, the larger the value of the physical distance factor and the stronger the vibration transmission intensity.

[0170] Speed ​​difference factor: the speed difference of the equipment This will affect the vibration coupling strength between devices. The closer the speeds are, the stronger the vibration coupling will be. The speed difference factor is modeled using the Sigmoid function, and the formula is:

[0171] Where, is the speed difference coefficient, usually taken as 0.01, and Equipment and equipment Real-time speed.

[0172] Multiply the physical distance factor and the speed difference factor to obtain the physical similarity matrix:

[0173] Feature similarity matrix: It mainly considers the influence of characteristic similarity and load difference between devices on vibration correlation, based on cross-scale coupling characteristics.

[0174] Where, represents the inner product of vectors, Represents the norm of a vector. The closer the feature cosine similarity value is to 1, the more similar the features of the two devices are, and the stronger the vibration correlation is.

[0175] Load difference filtering: When the load difference between devices When , the reflection effect of feature similarity on vibration correlation will be affected, so the indicator function To filter, the formula is:

[0176] Where, It is an indicator function. When the condition in the brackets is met, the value is 1, otherwise it is 0.

[0177] Dynamic fusion coefficient α: The value of α is dynamically adjusted by the physical distance between devices. The formula is:

[0178] Where, is the maximum distance between equipment in the factory. is the sensitivity control coefficient. 3. Spatiotemporal convolution operation: Based on dynamic adjacency matrix , the spatiotemporal correlation between devices is captured through a two-layer gated graph convolution (Gated GCN). Among them, the spatial graph convolution is calculated as:

[0179] Where, , is the degree matrix, is the identity matrix, , For the Node type of the layer, Representation layer normalization operation, For the The spatial convolution weights of the layer (of dimension 32×64). is the activation function (ReLU function is selected).

[0180] Temporal convolution: Perform 1D temporal convolution (kernel size 3) on the features output by spatial convolution to capture the correlation relationship in the time series. The formula is:

[0181] Where, represents a 1D convolution operation, Indicates the convolution kernel size.

[0182] The flow of features is controlled by the gating unit to alleviate the common over-smoothing problem in graph convolutional networks. The formula is:

[0183] Where, and are the weights and biases of the gate unit, is the sigmoid activation function, and ⊙ is the element-wise multiplication operation. After two layers of convolution operations, the spatiotemporal fusion features are obtained. .

[0184] Step 4: Timing modules that are sensitive to occasional faults: The sporadic fault-sensitive timing module is designed specifically for the "short-term transient" characteristics of sporadic faults (usually lasting 0.5-2s). Through dual-channel feature enhancement and gating mechanisms, it enhances the extraction of transient impact features while suppressing noise interference, solving the problem of delayed response to sporadic faults in traditional timing models. 1. Transient-continuous dual threshold detection: The module consists of two parallel pathways and a gated fusion unit. The specific structure is as follows: Teager Energy Operator (TEO): processes the feature H output by the spatiotemporal graph convolution to enhance the transient impact feature. The formula is:

[0185] Compared with traditional energy calculation methods, TEO is more sensitive to instantaneous energy changes.

[0186] Short-time zero-crossing rate: Calculate the number of zero-crossing points of the signal in each 10ms window to quantify the high-frequency characteristics of the impact. The formula is:

[0187] Where sign() is the sign function and T is the window size (here T = 10ms).

[0188] The feature H output by the spatiotemporal graph convolution is processed to enhance the transient impact feature. The results of TEO and short-time zero-crossing rate are added to obtain the formula:

[0189] In view of the periodicity of rotating machinery faults (such as the periodic impact of bearing faults), a comb filter with adjustable center frequency is designed. The formula is: The module consists of two parallel pathways and a gated fusion unit. The specific structure is as follows:

[0190] Where, is the fundamental frequency of the device, B is the filter bandwidth, K is the harmonic order, The control coefficient is 0-1.

[0191] After processing by the adaptive comb filter, the characteristics of the cycle-holding path are obtained:

[0192] The gating parameters are learned through 1D convolution to balance the contributions of transient features and periodic features. The formula is:

[0193] Where, represents a 1D convolution operation, Represents the number of output channels, and the sigmoid function maps the weight value to the range of [0,1].

[0194] The features of the two pathways are fused according to the learned weights. The formula is:

[0195] When a device fails, , at this time, transient features dominate the fusion results; when the equipment is operating normally, , with cyclical characteristics as the dominant factor.

[0196] 2. Temporal attention focus: The self-attention mechanism focuses on the features at the time of fault occurrence and improves the temporal resolution: Query / Key / Value generation:

[0197] Where, is the weight matrix.

[0198] Attention weight calculation:

[0199] Where, is the dimension of Query / Key, is a scaling factor used to prevent the weight value from being too large. Multiply the attention weight by Value to obtain the temporal enhancement feature .

[0200] Step 5: Troubleshooting Strategy: The fault diagnosis strategy combines dual-threshold detection with device contrast enhancement to achieve accurate identification and location of sporadic faults and generate an explainable diagnostic evidence chain. 1. Dual threshold detection mechanism: Timing Enhancement Features Dimensionality reduction to device-time anomaly vector through global average pooling , the abnormality degree is calculated based on the characteristic distribution of the equipment during normal operation. The formula is:

[0201] Where, For the The device is in The dimensionality reduction features of the moment, For the The characteristic mean value of each device during normal operation, is the corresponding standard deviation.

[0202] Dual threshold setting: High quantile threshold : Take the 95% quantile of the abnormality of normal data; Low quantile threshold : Take the 80% quantile of the abnormality of normal data; Dynamic threshold update: To adapt to baseline drift in normal operating characteristics caused by factors such as equipment aging, the threshold is recalculated every 24 hours using the latest normal data.

[0203] Decision logic: ; Three consecutive time steps satisfy ; : Determined as a "suspected occasional fault" (manual review required); : Judged as "normal".

[0204] This decision logic can effectively solve the problem of missed detection of occasional faults by a single threshold. 2. Device contrast enhancement: By comparing the abnormality levels of group-controlled devices, the characteristics of minor faults on a single device can be enhanced: Average abnormality of group devices:

[0205] Exception offset:

[0206] Final Rating

[0207] When the abnormality of a single device is significantly higher than the average abnormality of the group of devices, , the final score will be amplified 1-2 times, thereby enhancing the recognition of subtle faults.

[0208] 3. Fault location and evidence generation: Correlation impact: Calculates the impact of fault propagation on other devices based on the dynamic adjacency matrix. The formula is:

[0209] Location rules: The faulty device is the one with the highest abnormality and significant associated impact, namely:

[0210] Where, is the association weight.

[0211] Explainable evidence chain generation: The output contains diagnostic evidence of multi-scale feature contributions, as follows: Scale 1 evidence: the impact energy exceeds the standard multiple (such as "+12.7dB"); Scale 2 evidence: characteristic frequency peaks (e.g., “BPFO = 156 Hz enhanced 3 times”); Evidence of association: Change in correlation with adjacent equipment (e.g., “correlation with M4 vibration decreased by 40%”); Output format: Structured JSON is used to facilitate industrial engineers to quickly understand and locate the source of the fault.

[0212] Anything not described in this application can be achieved by adopting or drawing on existing technologies.

[0213] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0214] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A device fault diagnosis method based on multi-scale graph convolution, characterized in that: include: The vibration signal of the group control equipment is decomposed at multiple scales through the multi-scale input layer to extract the time domain features, spectrum domain features and time-frequency features of the vibration signal and generate multi-dimensional feature data; Dynamically and interactively fuse the multi-dimensional feature data through a cross-scale feature coupling module, enhance the recognition of weak fault features using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector; Using the 32-dimensional feature vector, a convolutional model of the relationship perception graph between devices is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed through the occasional fault sensitive timing module, the Teager energy operator is used to enhance the transient impact features, and the adaptive comb filter is combined to extract the periodic fault features. The dual-channel weight fusion and self-attention mechanism are used to output the timing features. Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

2. The method according to claim 1, characterized in that The multi-scale decomposition of the vibration signal of the group control device is performed through the multi-scale input layer, and the time domain features, spectrum domain features and time-frequency features of the vibration signal are extracted to generate multi-dimensional feature data, specifically: Use IEPE accelerometers to collect triaxial vibration signals and calculate transient impact characteristics such as peak value, pulse factor, and impact duration; waveform factor and skewness waveform distribution characteristics; high-frequency energy proportion and zero-crossing rate energy change characteristics; Extract peak frequency bandwidth, sideband energy ratio, and sideband frequency interval modulation characteristics through envelope analysis; The wavelet transform method is used to generate multi-scale energy entropy features, sub-band energy features and time-frequency domain dynamic features.

3. The method according to claim 1, characterized in that The multi-dimensional feature data is dynamically interactively fused through the cross-scale feature coupling module, and the recognition of weak fault features is enhanced by using a bidirectional gating mechanism, and the fused 32-dimensional feature vector is output, specifically: performing standardization processing on the multidimensional feature data; The standardized multi-dimensional feature data is spliced ​​in the channel dimension, and the importance weights of the features at each scale are learned through one-dimensional convolution and softmax function; The bidirectional gating mechanism is used to achieve complementary enhancement of features between scales, including: Strengthen the correlation between time domain features and frequency domain feature frequencies; Strengthen the relationship between frequency domain features and time-frequency features; The residual connection method is used to retain the key information of the original features and output the fused 32-dimensional feature vector.

4. The method according to claim 1, wherein The 32-dimensional feature vector is used to construct a convolutional model of the relationship perception graph between devices. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time working conditions. The dynamic adjacency matrix is ​​subjected to a spatiotemporal convolution operation to obtain spatiotemporal fusion features, specifically: The sensor of each device is a graph node, and the node feature is the 32-dimensional feature vector; The adjacency matrix weight is dynamically adjusted based on the physical distance of the equipment and the similarity of the real-time working conditions. The formula is: The physical distance factor is multiplied by the speed difference factor to generate a physical similarity matrix; The dynamic fusion coefficient α is adjusted according to the physical distance between devices. The smaller the α value, the higher the adjacency weight; The vibration transmission law between devices is captured through spatial convolution operation, and the correlation features in the time series are extracted through temporal convolution operation. The feature flow is controlled by the gating unit to alleviate the over-smoothing problem.

5. The method according to claim 1, wherein The spatiotemporal fusion features are processed by the sporadic fault sensitive timing module, the transient impact features are enhanced by the Teager energy operator, the periodic fault features are extracted by combining the adaptive comb filter, and the timing features are output by dual-channel weight fusion and self-attention mechanism. Specifically: The transient impact energy is calculated by the Teager energy operator, and the high-frequency characteristics are quantified by combining the short-time zero-crossing rate to generate transient features. Extract periodic fault features through adaptive comb filtering, where the center frequency of the filter is dynamically adjusted according to the fault type; The weights of transient features and periodic features are learned through one-dimensional convolution, and the features of the two pathways are fused according to the weights. The attention weight at the moment of fault occurrence is calculated through the self-attention mechanism to improve the temporal resolution.

6. The method according to claim 1, wherein Based on the dual-threshold detection mechanism and device contrast enhancement strategy, the time series features are used to identify and locate faults, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated. Specifically, Set transient thresholds and continuous thresholds to correspond to the abnormal judgment of short-term impact characteristics and periodic characteristics respectively; Thresholds are dynamically updated every 24 hours based on the latest normal data to adapt to baseline drift caused by equipment aging; Through the device comparison enhancement strategy, the deviation between the abnormality of a single device and the average abnormality of the group of devices is calculated, and the abnormal offset is generated to enhance the characteristics of weak faults.

7. The method according to claim 6, characterized in that Also includes: Calculate the fault propagation impact based on the dynamic adjacency matrix and determine the fault source device; Generate an interpretable evidence chain that includes multi-scale feature contributions, including the impact energy exceedance multiple, characteristic frequency peak and its change multiple, and correlation changes with neighboring equipment; Output a structured diagnostic report including fault location, evidence chain, and repair suggestions.

8. The method according to claim 4, characterized in that The generation of the dynamic adjacency matrix comprises the following steps: Obtain the physical distance data between group control devices and build a basic distance matrix; Collect the operating condition data of each device in real time and calculate the speed difference and load change similarity between two devices; The distance matrix and the working condition similarity matrix are weightedly fused to generate a dynamic adjacency matrix, in which the weight coefficients are pre-set or adjusted through online learning according to the equipment type and installation environment.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method according to any one of claims 1 to 8 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.

Images