A device fault diagnosis method based on multi-scale graph convolution

Abstract

The application provides a device fault diagnosis method based on multi-scale graph convolution, relates to the technical field of intelligent fault diagnosis of industrial equipment, and comprises the following steps: extracting time domain, frequency domain and time-frequency characteristics of a vibration signal through a multi-scale input layer, and generating a 32-dimensional feature vector by fusion through a cross-scale feature coupling module. A device relationship perception graph convolution model is constructed, a dynamic adjacency matrix is generated in combination with a physical distance and a real-time working condition, time-space fusion features are extracted through time-space convolution, transient and periodic characteristics are enhanced through an occasional fault sensitive time sequence module, time sequence features are output in combination with double-channel fusion and a self-attention mechanism. Finally, fault identification and positioning are realized through a double-threshold detection and a device comparison enhancement strategy, and an explainable diagnosis evidence chain containing multi-scale feature contributions is generated, so that the weak fault detection rate and the diagnosis reliability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent fault diagnosis of industrial equipment, and particularly relates to a device fault diagnosis method based on multi-scale graph convolution. BACKGROUND

[0002] In modern industrial production, rotating machinery equipment (such as fans, pumps, compressors, etc.) as a key component, the stability and reliability of its running state are crucial to the safety and efficiency of the entire production process. With the development of industrial Internet of Things (IIoT), big data and artificial intelligence technology, 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 are difficult to effectively handle problems such as multi-scale feature coupling, dynamic correlation modeling between devices, and transient event capture. Especially in complex industrial environments, due to the influence of background noise, vibration interference and other factors, weak fault signals are easily submerged, making early fault detection difficult.

[0003] The inventors found that existing methods mostly use simple concatenation or weighted fusion to process multi-scale features, lack effective cross-scale feature coupling mechanisms, and make it difficult to identify weak fault features. Traditional graph convolution networks (GCN) usually use static adjacency matrices, which cannot reflect the dynamic adjustment needs of vibration conduction relationships between devices as operating conditions change. Traditional double-threshold detection methods rely on manual experience to set thresholds and lack adaptive ability, which can easily lead to false positives and missed detections in complex industrial environments. SUMMARY

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

[0005] The technical solution adopted by the application is as follows:

[0006] The application provides a device fault diagnosis method based on multi-scale graph convolution, comprising:

[0007] The multi-scale input layer performs multi-scale decomposition on the vibration signal of the group control device to extract time domain features, frequency domain features and time-frequency features of the vibration signal, and generates multi-dimensional feature data;

[0008] The cross-scale feature coupling module dynamically interacts and fuses the multi-dimensional feature data, uses a bidirectional gating mechanism to enhance the recognition of weak fault features, and outputs a 32-dimensional feature vector after fusion;

[0009] A device relationship perception graph convolution model is constructed based on the 32-dimensional feature vector, a dynamic adjacency matrix is generated based on the physical distance between devices and the real-time operating condition similarity, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix.

[0010] The spatio-temporal fusion features are processed by the accidental fault sensitive timing module, the transient impact features are enhanced by using a Teager energy operator, the periodic fault features are extracted by combining an adaptive comb filter, and the timing features are output by a double-channel weight fusion and a self-attention mechanism;

[0011] Based on a double-threshold detection mechanism and a device contrast enhancement strategy, the timing features are subjected to fault identification and positioning to generate an explainable diagnostic evidence chain containing multi-scale feature contributions.

[0012] According to one embodiment of the present application, the vibration signals of the group control device are subjected to multi-scale decomposition by a multi-scale input layer, the time domain features, the spectral domain features and the time-frequency features of the vibration signals are extracted, and multi-dimensional feature data is generated, specifically:

[0013] IEPE acceleration sensors are used to collect three-axis vibration signals, and the peak value, the pulse factor, the impact duration transient impact features, the waveform factor, the skewness waveform distribution features, the high-frequency energy proportion, and the zero-crossing rate energy change features are calculated.

[0014] The peak frequency bandwidth, the sideband energy ratio, and the sideband frequency interval modulation features are extracted by envelope analysis.

[0015] Multi-scale energy entropy features, sub-band energy features, and time-frequency domain dynamic features are generated by using wavelet transform methods.

[0016] According to one embodiment of the present application, the multi-dimensional feature data is subjected to dynamic interactive fusion by a cross-scale feature coupling module, the recognition degree of weak fault features is enhanced by using a bidirectional gate mechanism, and a fused 32-dimensional feature vector is output, specifically:

[0017] The multi-dimensional feature data is subjected to standardization processing.

[0018] The standardized multi-dimensional feature data is spliced in the channel dimension, and one-dimensional convolution and a softmax function are used to learn the importance weights of the features of each scale.

[0019] The feature complementation and enhancement between scales are realized by a bidirectional gate mechanism, specifically including:

[0020] The correlation between the time domain features and the frequency domain features is strengthened.

[0021] The correlation between the frequency domain features and the time-frequency features is strengthened.

[0022] The residual connection method is used to retain the key information of the original features, and a fused 32-dimensional feature vector is output.

[0023] According to one embodiment of the present application, the device-to-device relationship perception graph convolution model is constructed based on the 32-dimensional feature vector, a dynamic adjacency matrix is generated based on the physical distance between devices and the real-time working condition similarity, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix. Specifically,

[0024] The sensors of each device are taken as graph nodes, and the node features are the 32-dimensional feature vectors.

[0025] The adjacency matrix weight is dynamically adjusted based on the physical distance between devices and the real-time working condition similarity.

[0026] The physical distance factor and the rotational speed difference factor are multiplied to generate a physical similarity matrix.

[0027] The dynamic fusion coefficient α is adjusted according to the physical distance between devices. The smaller the α value is, the higher the adjacency weight is.

[0028] The vibration conduction law between devices is captured by a spatial convolution operation, the correlation features on the time sequence are extracted by a time convolution operation, and the feature flow is controlled by a gating unit to alleviate the over-smoothing problem.

[0029] According to one embodiment of the present application, the spatio-temporal fusion feature is processed by an accidental fault sensitive time sequence module, a Teager energy operator is used to enhance the transient impact feature, a self-adaptive comb filter is used to extract the periodic fault feature, and a double-channel weight fusion and a self-attention mechanism are used to output the time sequence feature. Specifically,

[0030] 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 the transient feature.

[0031] The periodic fault feature is extracted by the self-adaptive comb filter, and the center frequency of the filter is dynamically adjusted according to the fault type.

[0032] The weights of the transient feature and the periodic feature are learned by one-dimensional convolution, and the features of the two paths are fused according to the weights.

[0033] The attention weight at the time of fault occurrence is calculated by the self-attention mechanism to improve the time resolution.

[0034] According to one embodiment of the present application, the time sequence feature is subjected to fault recognition and positioning based on a double-threshold detection mechanism and a device comparison enhancement strategy, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated. Specifically,

[0035] The transient threshold and the continuous threshold are set, which correspond to the abnormal judgment of the short-time impact feature and the periodic feature, respectively.

[0036] The threshold is dynamically updated based on the latest normal data every 24 hours to adapt to baseline drift caused by device aging factors;

[0037] By comparing the enhancement strategy of the device, the deviation of the single device anomaly degree and the average anomaly degree of the group device is calculated, and the abnormal offset is generated to strengthen the weak fault characteristics.

[0038] According to one embodiment of the present application, further comprising:

[0039] Based on the dynamic adjacency matrix, the fault propagation influence degree is calculated, and the fault source device is determined;

[0040] Generate an interpretable evidence chain, including multi-scale feature contributions, including impact energy exceeding multiple times, feature frequency peak value and its change multiple, and correlation change with adjacent devices;

[0041] Output a structured diagnostic report, including fault location, evidence chain and repair suggestion.

[0042] According to one embodiment of the present application, the generation of the dynamic adjacency matrix comprises the following steps:

[0043] Obtain the physical distance data between the group control devices, and construct a basic distance matrix;

[0044] Real-time acquisition of the running condition data of each device, calculation of the speed difference and load change similarity between each two devices;

[0045] The distance matrix and the working condition similarity matrix are weighted and fused to generate a dynamic adjacency matrix, wherein the weight coefficients are pre-set or adjusted online according to the device type and installation environment.

[0046] The second aspect embodiment of the present application provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the steps in the method.

[0047] The third aspect embodiment of the present application provides an electronic device, which includes a memory, a processor, and a program stored in the memory and executable on the processor, and the processor executes the program to realize the steps in the method.

[0048] Due to the adoption of the above technical solutions, the present application has the following beneficial effects:

[0049] The present application extracts time domain, frequency domain and time-frequency features of vibration signals through a multi-scale input layer, comprehensively covering various fault features in the device running process. By using the bidirectional gating mechanism in the cross-scale feature coupling module, dynamic interaction and complementary enhancement between different scale features are realized, the recognition degree of weak fault features is improved, and the accidental fault detection rate can reach 95.6% in a harsh environment with a signal-to-noise ratio of-10dB.

[0050] The application constructs an inter-device relationship perception graph convolution model, generates a dynamic adjacency matrix based on device physical distance and real-time working condition similarity, and realizes accurate modeling of the spatio-temporal correlation relationship between devices. Through spatio-temporal convolution operation, the mutual influence between devices is captured, and the fault positioning accuracy is improved to within 0.5 meters, effectively eliminating the positioning deviation caused by working condition changes.

[0051] The application combines the Teager energy operator and the adaptive comb filter in the occasional fault sensitive timing module, enhances the extraction ability of transient impact features, and focuses on the fault occurrence time through double-channel weight fusion and self-attention mechanism, realizes efficient capture of second-level occasional faults. A transient-continuous double threshold detection mechanism is designed, which has the function of adaptive threshold adjustment, significantly reduces the false positive rate in complex industrial environments, and reduces the missed detection situation.

[0052] Based on the double threshold detection mechanism and the device contrast enhancement strategy, the application identifies and locates faults based on timing features, generates an interpretable diagnostic evidence chain containing multi-scale feature contributions. Output structured diagnostic report, easy for industrial engineers to quickly understand and locate the fault source, further enhancing the practicality and industrial application value of the method. BRIEF DESCRIPTION OF DRAWINGS

[0053] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The illustrative embodiments of the application and their descriptions are used to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0054] Figure 1 A flowchart of a multi-scale graph convolution-based device fault diagnosis method provided by an embodiment of the application is shown in the drawings.

[0055] Figure 2 A structural diagram of an electronic device provided by an embodiment of the application is shown in the drawings.

[0056] Reference signs:

[0057] 810, processor; 820, communication interface; 830, memory; 840, communication bus. DETAILED DESCRIPTION

[0058] In order to more clearly explain the overall concept of the application, the following will be described in detail in an exemplary manner with reference to the drawings.

[0059] Many specific details are set forth in the following description in order to provide a thorough understanding of the present application. However, the present application can be practiced according to other embodiments that can not be explicitly described herein, and the present application is not limited to the embodiments described herein. It is to be understood that the embodiments and features of the present application can be combined with each other, if not contradictory.

[0060] In the present application, unless specifically stated and limited otherwise, the first feature is "on" or "under" the second feature can be that the first and second features are in direct contact, or the first and second features are indirectly in contact through an intermediate medium. In the description of the specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples.

[0061] Embodiment 1

[0062] As shown in Figure 1 A multi-scale graph convolution-based equipment fault diagnosis method includes:

[0063] S100, multi-scale decomposition of the vibration signal of the group control equipment is performed through a multi-scale input layer, time domain features, frequency domain features and time-frequency features of the vibration signal are extracted, and multi-dimensional feature data is generated.

[0064] As described above, the multi-scale input layer is the starting point of the entire fault diagnosis method, and its main purpose is to extract features at different scales from the vibration signal of the group control equipment (such as fans, pumps, etc.). Specifically, this step includes decomposing the original vibration signal to obtain time domain features, frequency domain features and time-frequency features.

[0065] Time domain features: These features are directly extracted from the time series of the original vibration signal, including peak value, pulse factor, impact duration, waveform factor, skewness, high-frequency energy proportion and zero-crossing rate, etc. They reflect the characteristics of the signal in the time dimension, such as transient impact strength, waveform distribution form and energy change.

[0066] Frequency domain features: modulation features are extracted from the vibration signal through envelope analysis techniques, including basic frequency features (such as peak frequency bandwidth), sideband features (such as sideband energy ratio) and spectral distribution features. These features reveal the key information of the signal in the frequency domain, which helps to identify specific types of fault patterns, such as wear or cracks in bearings or gearboxes.

[0067] Time-frequency features: Utilize wavelet transform to perform multi-scale decomposition on the signal, generating time-frequency features that describe the sudden change in energy distribution when a fault occurs. This includes multi-scale energy entropy features, sub-band energy features, and time-frequency domain dynamic features. Time-frequency features can capture changes in both time and frequency dimensions of the signal, making them particularly useful for detecting non-stationary signals.

[0068] For example, suppose we are monitoring a fan in an industrial plant, which is equipped with a three-axis vibration sensor (X / Y / Z directions). The vibration data collected by the sensor is first sent to the multi-scale input layer:

[0069] Time-domain feature extraction: Calculate parameters such as peak value, pulse factor, and impact duration within each time window during the fan's operation. For example, if a significant increase in peak value and a long impact duration are found within a certain time window, it may indicate a serious mechanical failure inside the fan.

[0070] Spectrum domain feature extraction: By performing Fast Fourier Transform (FFT) on the vibration signal, its frequency spectrum can be obtained, and features such as peak frequency bandwidth and sideband energy ratio can be extracted. For example, if a sudden increase in energy ratio of a certain sideband is detected, it may be due to a crack or eccentricity problem in a component of the fan.

[0071] Time-frequency feature extraction: Use wavelet transform to decompose the signal into multiple sub-bands and analyze the energy distribution of each sub-band. For example, when the energy of a certain sub-band suddenly decreases, it may indicate damage or other abnormal conditions of the fan blades.

[0072] S200, dynamically interactively fuse the multi-dimensional feature data through the cross-scale feature coupling module, enhance the recognition degree of weak fault features by using a bidirectional gating mechanism, and output a fused 32-dimensional feature vector.

[0073] As mentioned above, the cross-scale feature coupling module is a key step in the entire fault diagnosis method, aiming to enhance the recognition degree of weak fault features by dynamically interactively fusing features at different scales (time domain, spectrum domain, and time-frequency domain). Specifically, this module uses a bidirectional gating mechanism to process multi-dimensional feature data, ensuring effective complementation and enhancement between features at different scales, and finally outputs a 32-dimensional fused feature vector.

[0074] Standardization: First, standardize the extracted time-domain, spectrum-domain, and time-frequency-domain features to keep them consistent in magnitude, facilitating subsequent feature fusion.

[0075] Feature concatenation and fusion: The standardized features are concatenated in the channel dimension, and 1D convolution operations are used to learn the importance weights of each scale feature. These weights determine the contribution of each feature in the final fused feature.

[0076] Bidirectional gating mechanism: The complementary enhancement of features between scales is achieved through a bidirectional gating mechanism. For example, the correlation between time-domain impact features and spectral domain feature frequencies can be strengthened, as well as the interaction between spectral domain features and time-frequency domain features. In addition, residual connection is used to preserve key information of original features, avoiding loss of important features.

[0077] Output fusion feature vector: After the above processing, a 32-dimensional fusion feature vector is generated, which integrates feature information at different scales and can more effectively identify weak fault signals.

[0078] For example, assume that time-domain, spectral domain, and time-frequency domain feature data have been extracted from the vibration signals of a fan:

[0079] Standardization: Standardize the time-domain features (such as peak value, pulse factor, etc.), spectral domain features (such as peak frequency bandwidth, sideband energy ratio, etc.), and time-frequency features (such as multi-scale energy entropy, sub-band energy features, etc.) to make them consistent in numerical range. The standardized features are concatenated into a high-dimensional feature matrix.

[0080] Use 1D convolution layers to learn the importance weights of each scale feature. For example, it is found that the peak value in the time-domain feature is particularly important for identifying transient impacts, while the sideband energy ratio in the spectral domain feature is more critical for detecting modulation faults.

[0081] Use a bidirectional gating mechanism to achieve complementary features between scales. For example, within a certain time window, if the time-domain feature shows a large peak value but the spectral domain feature does not show significant changes, the bidirectional gating mechanism can be used to enhance the influence of the time-domain feature to better capture potential transient impacts.

[0082] Use residual connection to preserve key information of original features, ensuring that important fault signals are not lost during feature fusion.

[0083] Output fusion feature vector: After the above processing, a 32-dimensional fusion feature vector is generated, which not only contains feature information at all scales, but also enhances the recognition of weak fault features through the bidirectional gating mechanism.

[0084] S300, construct an inter-device relationship perception graph convolution model based on the 32-dimensional feature vector, generate a dynamic adjacency matrix based on device physical distance and real-time operating condition similarity, and obtain a spatio-temporal fusion feature by performing spatio-temporal convolution operations on the dynamic adjacency matrix.

[0085] As mentioned above, the inter-device relationship-aware graph convolutional model is constructed based on the generated 32-dimensional feature vectors, aiming to capture the spatio-temporal correlation between group control devices through a dynamic adjacency matrix. This model uses the physical distance between devices and the real-time operating condition similarity (such as speed, load, etc.) to adjust the connection weight between devices, and extracts the fused features through spatio-temporal convolution operation.

[0086] Inter-device relationship modeling: First, the sensors of each device are taken as nodes in the graph, and the node features are the 32-dimensional feature vectors after cross-scale coupling. Based on the physical distance between devices and the real-time operating condition similarity, a dynamic adjacency matrix is constructed to reflect the vibration conduction relationship between devices.

[0087] Dynamic adjacency matrix generation: The dynamic adjacency matrix not only considers the physical distance between devices, but also combines the real-time operating condition (such as speed difference, load change, etc.). For example, two devices that are physically close and have similar operating conditions may have a higher correlation weight, indicating that they have a strong vibration conduction influence on each other.

[0088] Spatio-temporal convolution operation: After the dynamic adjacency matrix is constructed, a spatial convolution operation is performed through the graph convolution network (GCN) to capture the local correlation pattern between devices. Then, a one-dimensional temporal convolution is applied to the spatially convolved feature sequence to extract the correlation relationship in the time sequence. The entire process optimizes the feature flow through the gating unit to prevent over-smoothing problems, and finally obtains the spatio-temporal fused features.

[0089] For example, assume that an industrial plant containing multiple fans is being monitored:

[0090] Each fan is taken as a node, and the node features are the 32-dimensional feature vectors after cross-scale coupling processing. These features reflect the current operating state of the fan, including time domain, frequency domain, and time-frequency domain information.

[0091] The physical distance between each fan is calculated, and the speed and load data of each fan are collected in real time. For example, if two fans are physically close and have similar speeds, it is considered that they have a strong correlation, and a higher adjacency matrix weight is assigned.

[0092] The weight values in the adjacency matrix are dynamically adjusted to adapt to changes under different operating conditions. For example, when the load of a fan suddenly increases, the vibration conduction relationship between it and its neighboring fans will also change, and the system will automatically update the adjacency matrix to reflect this change.

[0093] First, a spatial convolution operation is performed to aggregate the node features using the dynamic adjacency matrix, obtaining spatially fused features. This step can capture the mutual influence between fans, such as how a fault signal of one fan propagates to other fans.

[0094] Then a time convolution operation is performed, applying a one-dimensional convolution to the spatial fusion feature sequence to extract the change pattern over time. This helps identify the development trend of the fault over time, such as whether the fault of a certain fan gradually intensifies.

[0095] The feature flow is controlled by the gating unit to ensure that important features are effectively passed on and avoid information loss or over-smoothing phenomena.

[0096] S400, the spatio-temporal fusion features are processed by the occasional fault sensitive time sequence module, the transient impact features are enhanced using the Teager energy operator, the periodic fault features are extracted using the adaptive comb filter, and the time sequence features are output by the dual-channel weight fusion and self-attention mechanism.

[0097] As mentioned above, the occasional fault sensitive time sequence module is designed for the characteristics of occasional faults in group control equipment, aiming to improve the accuracy of fault detection by enhancing transient impact features and extracting periodic fault features. This module uses the Teager energy operator (TEO) to enhance transient impact features and combines with the adaptive comb filter to extract periodic fault features. Finally, the time sequence features are generated by dual-channel weight fusion and self-attention mechanism.

[0098] Transient impact feature enhancement: Teager energy operator is used to process spatio-temporal 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 and sudden faults.

[0099] Periodic fault feature extraction: an adaptive comb filter is used to extract periodic fault features. This filter can dynamically adjust the center frequency according to the actual speed of the device to accurately capture the periodic vibration characteristics of bearings or gearboxes and other components.

[0100] Dual-channel weight fusion: the transient impact features and periodic fault features are processed as two independent channels, the weights of each channel are learned by 1D convolution, 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.

[0101] Self-attention mechanism: the self-attention mechanism is used to focus on the features at the time of fault occurrence, improving the time resolution. Through the query / key / value generation mechanism and attention weight calculation method, the model can accurately identify the time of fault occurrence in complex time series.

[0102] For example, suppose we are monitoring a fan cluster in an industrial plant:

[0103] The Teager energy operator is used to process the time-space fusion characteristics of the fan, and the transient impact characteristics are enhanced. For example, if the fan blade is suddenly impacted by external impact within a certain time window, causing obvious transient impact in the vibration signal, TEO can significantly amplify this feature, making it easier for the system to identify this abnormal event.

[0104] An adaptive comb filter is used to extract the periodic fault characteristics of the fan. For example, if the fan bearing has local damage, periodic impact characteristics will appear in the vibration signal. The adaptive comb filter can dynamically adjust the center frequency according to the current speed of the fan, accurately capturing these periodic impacts.

[0105] The transient impact characteristics and periodic fault characteristics are processed as two independent paths. The weights of each path are learned by 1D convolution, for example, if the transient impact characteristics are more important in a certain time period, a higher weight is given; otherwise. Then, according to the weight, the features of the two paths are fused to generate comprehensive time sequence features.

[0106] The self-attention mechanism is used to focus on the features at the time of fault occurrence. For example, during the operation of the fan, a slight mechanical fault occurs at a certain time, causing a slight change in the vibration signal. The self-attention mechanism can automatically identify and amplify the features at this time, helping the system to more accurately locate the time point of fault occurrence.

[0107] S500, based on the double threshold detection mechanism and the device comparison enhancement strategy, the time sequence characteristics are fault identified and located, and the interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0108] As mentioned above, the double threshold detection mechanism and the device comparison enhancement strategy are the core means for fault identification and positioning in the present application, aiming to improve the detection accuracy and positioning precision of occasional faults by combining multi-scale feature information. After obtaining the time sequence features output by the occasional fault sensitive time sequence module, this step further analyzes and judges, and generates a diagnostic evidence chain with physical interpretability.

[0109] Double threshold detection mechanism: set two independent but cooperative thresholds - transient threshold and continuous threshold, respectively, to identify short-time impact type faults (such as transient impact, loosening) and periodic / continuous faults (such as bearing wear, gear crack). When the time sequence characteristics in a certain period exceed both thresholds, the system determines that it is a real fault event; if only the transient threshold is exceeded, it is considered as a non-fault disturbance to avoid false positives.

[0110] Device contrast enhancement strategy: compare the abnormality degree of the current device with the average abnormality degree of other similar devices in the same group, and calculate the deviation value. This strategy can effectively distinguish between individual device failures and group operation fluctuations (such as load mutations and environmental vibrations), thereby enhancing the performance of weak fault signals and improving diagnostic sensitivity.

[0111] Interpretable diagnostic evidence chain generation: after completing fault identification and positioning, the system automatically generates a structured and multi-level diagnostic evidence chain, including:

[0112] The impact energy exceeds the threshold multiple times at the time of fault occurrence; the characteristic frequency peak value and its trend; the correlation deviation with other devices; the contribution weight of multi-scale features (time domain, frequency domain, and time-frequency domain) in decision-making; these information together constitute a clear logical and explicit physical meaning of diagnostic evidence, which is easy for operation and maintenance personnel to understand and verify.

[0113] For example, assume that in a fan group control system, a fan has a small crack in the bearing and produces occasional vibration impact:

[0114] The system monitors a significant transient impact on the fan at a certain time, and its transient energy characteristics exceed the preset "transient threshold";

[0115] At the same time, after adaptive comb filter processing, it is found that the device also has periodic feature enhancement in the corresponding time period, and the amplitude exceeds the "continuous threshold";

[0116] Since both conditions are met, the system determines that it is a real fault event and triggers an alarm; if only the transient exceeds, it may be identified as external interference (such as foreign object impact), and no high-level alarm is triggered.

[0117] Calculate the abnormal score of the fan (based on the degree of deviation from the normal model of time series characteristics);

[0118] At the same time, the average abnormal score of the remaining 10 fans in the same group is calculated;

[0119] If the abnormal score of the fan is significantly higher than the group average (for example, 2 times the standard deviation), it is further confirmed as an isolated fault source, rather than a change in the overall operating environment;

[0120] This deviation result is used to enhance the fault confidence of the device and assist in precise positioning.

[0121] Diagnostic evidence chain output example:

[0122] "Fault time: April 5, 2025, 14:23:17";

[0123] "Main features: impact energy exceeds 3.2 times, characteristic frequency (BPFO) amplitude rises by 2.8 times";

[0124] "Correlation analysis: correlation with adjacent equipment decreases by 40%, indicating strong vibration independence";

[0125] "Feature contribution: time domain features account for 45%, frequency domain features account for 35%, and time-frequency domain features account for 20%";

[0126] "Suggested measures: suggest stopping the machine to check whether there is an early crack in the outer ring of bearing #3".

[0127] This evidence chain not only provides the judgment result, but also reveals the judgment basis, greatly enhancing the credibility and practicality of the system.

[0128] According to one embodiment of the present application, the vibration signal of the group control device is multi-scale decomposed by the multi-scale input layer, the time domain features, frequency domain features and time-frequency features of the vibration signal are extracted, and multi-dimensional feature data is generated, specifically:

[0129] Use IEPE acceleration sensor to collect three-axis vibration signals, and calculate peak value, pulse factor, impact duration transient impact features; waveform factor, skewness waveform distribution features; high-frequency energy ratio, zero-crossing rate energy change features;

[0130] Extract peak frequency bandwidth, sideband energy ratio, and sideband frequency interval modulation features through envelope analysis;

[0131] Generate multi-scale energy entropy features, sub-band energy features and time-frequency domain dynamic features using wavelet transform method.

[0132] As described above, first, the IEPE acceleration sensor installed on the group control device is used to collect the vibration signals of the device in three orthogonal directions during operation, i.e. three-axis vibration signals. The sensor has good frequency response characteristics and anti-interference ability, and can accurately capture the vibration state of the device under different working conditions.

[0133] After obtaining the original vibration signal, time-domain feature extraction is performed. The peak value of the signal is calculated to reflect the maximum degree of vibration amplitude; the pulse factor, i.e., the ratio of the peak value to the absolute average value, is calculated to identify whether there is an impact vibration; the impact duration is extracted to represent the duration of the impact event. The above constitutes the transient impact class feature. At the same time, the waveform factor, i.e., 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 judge whether the vibration is skewed abnormal. In addition, energy change class features are also extracted, including high-frequency energy proportion, i.e., the proportion of energy in the high-frequency band (such as above 2 kHz) to the total energy, which is used to identify the increase of high-frequency components caused by early weak faults; the zero-crossing rate, i.e., the number of times the signal crosses zero per unit time, is calculated to reflect the frequency of signal fluctuations.

[0134] Then, frequency domain feature extraction is performed. Envelope analysis is performed on the original vibration signal, specifically including band-pass filtering, Hilbert transform to obtain the envelope signal, and then frequency spectrum analysis is performed on the envelope signal. The peak frequency bandwidth, i.e., the frequency range occupied by the main frequency peak in the envelope spectrum, is extracted to reflect the frequency concentration degree of the fault impact; the sideband energy ratio, i.e., the ratio of the energy of the sideband frequency component to the energy of the center frequency, is extracted to identify modulation type faults, such as gear meshing faults; the sideband frequency interval, i.e., the frequency difference between adjacent sidebands, is extracted, which corresponds to the fault characteristic frequency and helps to judge the fault type and location.

[0135] Finally, time-frequency domain feature extraction is performed. The original vibration signal is decomposed into multiple sub-signals of different frequency bands using wavelet transform method. Based on the signal energy of each sub-band, the multi-scale energy entropy is calculated to measure the uniformity of energy distribution at different scales, and energy concentration often occurs when a fault occurs; the energy value of each sub-band is extracted to form a sub-band energy feature for identifying the frequency band where the energy is abnormally concentrated; at the same time, the time-frequency domain dynamic feature is extracted by analyzing the energy change trend at each scale, which is used to capture the evolution process of non-stationary and transient faults.

[0136] According to one embodiment of the present application, the multi-dimensional feature data is dynamically interactively fused by the cross-scale feature coupling module, the recognition degree of weak fault features is enhanced by using a bidirectional gating mechanism, and a fused 32-dimensional feature vector is output, specifically:

[0137] The multi-dimensional feature data is standardized;

[0138] The standardized multi-dimensional feature data is spliced in the channel dimension, and the importance weight of each scale feature is learned by one-dimensional convolution and softmax function;

[0139] The feature complementation and enhancement between scales is realized through a bidirectional gating mechanism, specifically including:

[0140] The correlation between the time domain feature and the frequency domain feature is strengthened;

[0141] The correlation between the frequency domain feature and the time-frequency feature is strengthened;

[0142] The residual connection method is used to retain the key information of the original feature, and a fused 32-dimensional feature vector is output.

[0143] As described above, first, the multi-dimensional feature data composed of the time domain feature, the frequency domain feature and the time-frequency domain feature extracted by the multi-scale input layer is standardized. This processing includes mean normalization or maximum and minimum normalization for each type of feature, so that features of different dimensions and different value ranges remain consistent in numerical scale, avoiding the dominance of a certain type of feature due to excessive amplitude, and ensuring the comparability and fairness of each scale feature in subsequent fusion.

[0144] Then, the standardized multi-dimensional feature data is spliced in the channel dimension to form a high-dimensional feature vector. This splicing operation integrates features from different time domains, frequency domains and time-frequency domains into a unified input representation, providing a basis for subsequent weight learning and feature fusion.

[0145] Next, the spliced features are processed through one-dimensional convolution operation to extract cross-scale local correlation and generate importance scores for each scale feature. Then, these scores are input into the softmax function to calculate the normalized weight values corresponding to each feature scale. This weight reflects the relative importance of each scale feature to fault identification under the current working condition, achieving adaptive weighting of different features.

[0146] On this basis, a bidirectional gating mechanism is introduced to realize dynamic interaction and complementary enhancement between different scale features. Specifically, there are two directions of gating process: first, the correlation between the time domain feature and the frequency domain feature is strengthened, and the modulation characteristics of strong impact signals in the frequency domain are enhanced through the gating structure, improving the response ability to transient faults; second, the correlation between the frequency domain feature and the time-frequency domain feature is strengthened, and the sideband information in the frequency domain is dynamically coupled with the sub-band energy in the wavelet decomposition through the gating mechanism, enhancing the ability to capture non-stationary fault features. The gating mechanism selectively retains or amplifies the fault-related feature response by controlling the flow path of information.

[0147] Meanwhile, in the feature fusion process, a residual connection method is adopted to directly introduce the original multi-dimensional feature information or the backbone feature of the intermediate layer to the output end. This structure can effectively retain the key information in the input features, prevent important features from being lost due to nonlinear transformation or gating operation during deep-layer transmission, and improve the stability and expression ability of the model.

[0148] Finally, after standardization, splicing, weight learning, bidirectional gate fusion and residual connection, the high-dimensional features are compressed and mapped into a 32-dimensional low-dimensional fusion feature vector. This vector integrates the key information in the time domain, frequency domain and time-frequency domain, and highlights the expression of weak fault features through dynamic weighting and interactive enhancement mechanism, providing high-discriminative input features for subsequent graph convolution modeling and fault diagnosis.

[0149] According to one embodiment of the present application, the 32-dimensional feature vector is used to construct an inter-device relationship-aware graph convolution model, a dynamic adjacency matrix is generated based on the physical distance and real-time working condition similarity of the devices, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix, specifically:

[0150] The sensors of each device are taken as graph nodes, and the node features are the 32-dimensional feature vectors;

[0151] The adjacency matrix weight is dynamically adjusted based on the physical distance and real-time working condition similarity of the devices;

[0152] The physical distance factor and the speed difference factor are multiplied to generate a physical similarity matrix;

[0153] The dynamic fusion coefficient a is adjusted according to the physical distance between devices, and the smaller the a value, the higher the adjacency weight;

[0154] The spatial convolution operation is used to capture the vibration conduction law between devices, the time convolution operation is used to extract the correlation features on the time sequence, and the gating unit is used to control the feature flow to alleviate the over-smoothing problem.

[0155] As described above, first, the sensors of each device in the group control system are taken as a node in the graph structure to construct the basic topology structure of the graph model. The feature input of each node is the 32-dimensional feature vector output by the previous step, which already contains the comprehensive feature information of the device vibration signal in the time domain, frequency domain and time-frequency domain, and can effectively represent the current running state of the device.

[0156] On this basis, a dynamic adjacency matrix is constructed to describe the connection relationship and its weight between device nodes. The adjacency matrix is not fixed but dynamically adjusted according to the physical distance between devices and the real-time running condition. Specifically, first, the physical distance between devices is calculated to form a physical distance factor; at the same time, the real-time speed data of each device is obtained to calculate the speed difference between each pair of devices to form a speed difference factor. Multiply the physical distance factor and the speed difference factor to get the physical similarity matrix, which reflects the comprehensive closeness of devices in spatial position and running state.

[0157] In order to further optimize the allocation of adjacency weight, a dynamic fusion coefficient α is introduced, which is adaptively adjusted according to the physical distance between devices. When the physical distance between two devices is close, the value of α is set to be small, indicating that the possibility of vibration conduction between them is higher, so a higher adjacency weight is given; on the contrary, when the distance between devices is far, the value of α increases, and the adjacency weight decreases accordingly. Through this mechanism, the adjacency matrix can more truly reflect the actual correlation strength between devices under different installation layouts and running environments.

[0158] After the dynamic adjacency matrix is constructed, a spatio-temporal convolution operation is performed. First, the spatial convolution operation is performed, and the graph convolution network (GCN) is used to aggregate the node features on the graph structure, that is, the feature of each node is updated to the weighted sum of its own and its neighbor node features, and the weight is determined by the dynamic adjacency matrix. This process can capture the vibration conduction law between devices and identify the path and impact range of the possible propagation of fault signals.

[0159] Subsequently, a time convolution operation is performed in the time dimension. The node feature sequence obtained after spatial convolution is processed by one-dimensional convolution along the time axis to extract the pattern of feature evolution over time and identify the dynamic change trend of device state. This operation can effectively capture the evolution process of features before and after the occurrence of faults, enhancing the time sensitivity to sporadic faults.

[0160] During the entire spatio-temporal convolution process, a gating unit is introduced to control the transmission process of features. The gating unit adjusts the update and retention ratio of information to prevent the feature from being over-smoothed by multiple graph convolutions, that is, to avoid the features of different nodes becoming similar after multiple aggregations and losing discriminative ability. Through the gating mechanism, it ensures that the key features are effectively retained and enhanced in the propagation process.

[0161] Finally, after joint processing of spatial convolution and time convolution, spatio-temporal fusion features containing spatial correlation and time evolution information between devices are obtained, providing sufficient context support for subsequent fault-sensitive feature extraction and diagnosis decision-making.

[0162] According to one embodiment of the present application, the spatio-temporal fusion features are processed by the sporadic fault sensitive timing module, which adopts Teager energy operator to enhance transient impact features, combines adaptive comb filter to extract periodic fault features, and outputs timing features through double-channel weight fusion and self-attention mechanism, specifically:

[0163] The transient impact energy is calculated by Teager energy operator, and the high-frequency characteristics are quantified by combining short-time zero-crossing rate to generate transient features.

[0164] Periodic fault features are extracted by adaptive comb filter, wherein the center frequency of the filter is dynamically adjusted according to the fault type.

[0165] The weights of transient features and periodic features are learned by one-dimensional convolution, and the features of the two paths are fused according to the weights.

[0166] The attention weights of the fault occurrence time are calculated by the self-attention mechanism to improve the time resolution.

[0167] As described above, first, for the transient impact signal commonly seen in sporadic faults, Teager energy operator is adopted to process the time series in the spatio-temporal fusion features. Teager energy operator can effectively reflect the instantaneous energy change of the signal, especially for short-time, high-intensity impact components. The transient impact energy at each time point is calculated by the operator to highlight the energy mutation characteristics at the moment of fault occurrence. At the same time, the high-frequency vibration characteristics are quantified by combining short-time zero-crossing rate, i.e. the number of signal zero-crossings is counted within a sliding time window to identify the high-frequency oscillation behavior caused by local damage. The Teager energy output and the short-time zero-crossing rate result are combined to form a feature vector describing the transient impact characteristics, which is one of the inputs for subsequent analysis.

[0168] Secondly, in order to capture the periodic impact features commonly seen in device faults (such as bearing fault frequency, gear meshing frequency, etc.), an adaptive comb filter is used to process the spatio-temporal fusion features. The comb filter has multiple equidistant passbands, which can enhance multiple harmonic frequency components at the same time. Its key parameters, center frequency and frequency interval, are dynamically adjusted according to the device type and fault mode. For example, when the diagnosis object is a rolling bearing, the system calculates the theoretical fault characteristic frequency according to the bearing geometric parameters and real-time speed, and takes it as the fundamental frequency of the comb filter; for gear type devices, the meshing frequency and its sidebands are determined according to the number of teeth and speed. The filter passband is updated in real time according to the operating conditions to ensure that the periodic fault components can be accurately extracted under the condition of speed variation, and the corresponding periodic fault feature sequence is generated.

[0169] Subsequently, the generated transient features and periodic fault features are fused as two independent processing paths. A one-dimensional convolution network is used to learn the features of the two paths and automatically identify which features are more discriminative in different fault stages or operating conditions. For example, when the transient impact is dominant in the early stage of the fault, the system gives higher weight to the transient features. When the fault develops to the stable periodic impact stage, the fusion proportion of the periodic features is increased. Finally, the features of the two paths are fused by weighting according to the learned weight coefficients, and a comprehensive feature representation with both transient sensitivity and periodic recognition ability is obtained.

[0170] Finally, a self-attention mechanism is introduced based on the fused features to further improve the positioning ability of the fault occurrence time. The self-attention mechanism calculates the correlation between each time step in the feature sequence to generate an attention weight distribution. The weight reflects the importance of each time point to the current feature expression, so that the model can automatically focus on the time period when the abnormal event occurs. For example, when there is a significant energy surge or periodic enhancement at a certain time, the attention weight of that time will be significantly amplified, thereby enhancing the expression strength of the feature at that time and improving the time resolution. After attention weighting, the final time sequence feature is output. This feature not only contains rich fault mode information, but also highlights the abnormal performance of key time points, providing high-precision input for subsequent fault recognition and positioning.

[0171] According to one embodiment of the present application, the dual-threshold detection mechanism and the device comparison enhancement strategy are used to perform fault recognition and positioning on the time sequence features, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated, specifically:

[0172] Set transient threshold and duration threshold to determine the abnormality of short-time impact features and periodic features, respectively;

[0173] Update the threshold based on the latest normal data every 24 hours to adapt to the baseline drift caused by device aging factors;

[0174] Through the device comparison enhancement strategy, the deviation between the abnormality degree of a single device and the average abnormality degree of a group of devices is calculated to generate an abnormality offset to strengthen weak fault features.

[0175] As mentioned above, first, a double-threshold detection mechanism is set up to determine whether the equipment operating state is abnormal. The 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 abnormalities, corresponding to the transient impact characteristics of the Teager energy operator output; when the transient energy exceeds the threshold in a certain period, it indicates that there may be a sudden mechanical impact or local damage. The sustained threshold is used to identify periodic or continuous failure characteristics, corresponding to the periodic failure components extracted by the adaptive comb filter; when the periodic feature amplitude is continuously higher than the threshold for a certain time (such as consecutive multiple time windows), it is considered that the equipment has a stable developing failure trend. Only when both the transient feature and the sustained feature exceed their respective thresholds, the system determines that it is a valid failure event, avoiding misjudgment of single transient interference as a real failure.

[0176] To adapt to the vibration baseline drift caused by factors such as aging, lubrication state change, installation looseness, etc. during long-term operation of the equipment, the threshold is not fixed but is dynamically updated. The system recalculates the threshold every 24 hours based on the normal operation data collected in the recent period. The specific method is: select the time series feature data marked as "normal state", and calculate the distribution characteristics (such as mean and standard deviation) of its transient energy and periodic features, and update the transient threshold and sustained threshold accordingly. For example, the mean plus 2 times the standard deviation can be used as the new threshold, ensuring that the threshold is self-adaptive to the slow change of the equipment health state, maintaining the detection sensitivity and stability.

[0177] On this basis, the device contrast enhancement strategy is introduced to further improve the recognition ability of weak faults. This strategy compares the operating state differences of the target device with other devices in the same group to identify individual abnormalities. Specifically, first, calculate the abnormality index of each device, which is determined by the degree of deviation of its time series features from the normal model, for example, it can be calculated based on the feature amplitude, change rate or Euclidean distance from the historical baseline. Then, calculate the average of the abnormality index of all devices in the group as the group baseline under the current working condition. Compare the abnormality index of a single device with the average value to calculate its deviation value, i.e. 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 failure risk, rather than being affected by the overall environmental fluctuations. The abnormal offset is used to enhance the failure confidence of the device, effectively amplifying the performance of weak fault signals in the diagnosis results.

[0178] Finally, based on the completion of fault identification and positioning, the system generates an explainable diagnostic evidence chain containing multi-scale feature contributions. The evidence chain is output in a structured form, including the time point of fault occurrence, the transient impact energy exceeding multiple, the periodic characteristic frequency and its amplitude change, the abnormal offset size, and the relative contribution proportion of time domain, frequency domain, and time-frequency domain features in this diagnosis. The evidence chain not only provides the fault judgment result, but also reveals the judgment basis, which is convenient for operation and maintenance personnel to understand the diagnostic logic, supports subsequent maintenance decisions, and realizes the transformation from "black box warning" to "transparent diagnosis".

[0179] According to one embodiment of the present application, further comprising:

[0180] Based on the dynamic adjacency matrix, the influence degree of fault propagation is calculated to determine the fault source device;

[0181] An explainable evidence chain is generated, containing multi-scale feature contributions, specifically including impact energy exceeding multiple, characteristic frequency peak and its change multiple, and correlation change with adjacent devices;

[0182] A structured diagnostic report is output, including fault location, evidence chain and repair suggestion.

[0183] As described above, first, based on the dynamic adjacency matrix constructed in the previous step, the influence degree of fault propagation between device groups is calculated to determine the most likely fault source device. Specifically, after detecting that one or more devices are abnormal, the correlation strength and directionality information between devices described by the dynamic adjacency matrix are used to analyze the propagation path of abnormal features in reverse. By comparing the time sequence of abnormal occurrence of each device, the trend of feature intensity change, and its connection relationship in the graph structure, the possibility score of each device as the fault source is calculated. For example, if a device is the earliest to appear abnormal in time, and its abnormal features propagate to multiple downstream devices through high-weight edges, then the device is determined as a high-influence node, and it is most likely to be the fault source. This process combines the graph propagation model and time series correlation analysis to improve the accuracy of fault positioning.

[0184] Secondly, an explainable diagnosis evidence chain is generated to clearly present the basis for fault judgment. This evidence chain not only contains the final diagnosis result, but also integrates key indicators from the original signal to the feature fusion process, ensuring that the diagnosis process is transparent and credible. The specific content includes: the impact energy exceeding multiple, which 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 amplitude growth of the frequency component related to a specific fault type in the frequency domain, supporting the judgment of fault type; and the correlation change with adjacent devices, revealing the coupling relationship change caused by the fault by comparing the correlation coefficient or cooperativity index of the vibration signals between devices before and after the fault, assisting 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.

[0185] Finally, the system outputs a structured diagnosis report as a direct basis for human-computer interaction and operation and maintenance decision-making. The report is presented in a standardized format, including three core parts: first, the fault location, which clearly indicates the equipment number, installation location, and possible fault components (such as bearings, gearboxes, etc.) where the fault occurs; second, the diagnosis evidence chain, which lists each supporting indicator and its value in an itemized manner, making it easy for technicians to review and judge the logic; third, the repair recommendation, which recommends appropriate handling measures based on the fault type and severity, such as "recommend stopping #4 fan drive end bearing inspection", "priority lubrication maintenance and monitoring of subsequent impact energy trends", etc. The report can support multiple output forms, such as text, PDF, or integration into the industrial monitoring platform interface, meeting the use requirements in different scenarios.

[0186] According to an embodiment of the present application, the generation of the dynamic adjacency matrix comprises the following steps:

[0187] Obtain the physical distance data between group control devices to construct a basic distance matrix;

[0188] Real-time acquisition of the operating condition data of each device, calculation of the speed difference and load change similarity between each pair of devices;

[0189] Weighted fusion of the distance matrix and the operating condition similarity matrix to generate a dynamic adjacency matrix, wherein the weight coefficients are pre-set according to the device type and installation environment or adjusted online.

[0190] As described above, first, the physical distance data between each device in the group control system is acquired. This data can be acquired through device installation drawings, on-site measurement or positioning systems (such as UWB, RFID, etc.), reflecting the relative position of the devices in the spatial layout. Based on these 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 association between devices, reflecting the physical possibility of vibration signal propagation between devices. The closer the distance, the stronger the mechanical coupling and vibration conduction relationship between the devices.

[0191] Secondly, the operating condition data of each device is collected in real time, including the speed, load, working temperature and other key parameters. Based on these data, the operating state similarity between any two devices is calculated. Specifically, by analyzing the time series of the device speed, the 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 speed difference and load change similarity indicators are integrated into an operating condition similarity matrix, which reflects the degree of cooperation of the devices in the operating behavior. The more similar the operating conditions, the more likely the devices will show stronger correlation in vibration characteristics.

[0192] Then, the basic distance matrix and the operating condition similarity matrix are weighted and fused to generate the final dynamic adjacency matrix. In 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 condition in the overall association evaluation. This weight coefficient can be pre-set according to the device type and installation environment, for example, in the scene where vibration conduction mainly depends on the physical structure (such as sharing the base), the weight of the distance item is increased; while in the scene where the running synchronization dominates the association (such as parallel running pump group), the weight of the operating condition similarity is increased. In addition, this weight coefficient can also be automatically adjusted through an online learning mechanism according to historical fault data or normal running samples, realizing adaptive optimization, so that the adjacency matrix can evolve dynamically with the running state of the system.

[0193] The final generated dynamic adjacency matrix not only reflects the static spatial relationship between devices, but also integrates the influence of real-time operating state, which can more accurately depict the dynamic association structure of vibration characteristics between group control devices, providing reliable spatial topology support for subsequent graph convolution operations.

[0194] The second aspect embodiment of the present application provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the method in any of the embodiments of the first aspect.

[0195] Figure 2An example of a schematic diagram of a physical structure of an electronic device is shown in Figure 2 As shown, the electronic device can 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 through the communication bus 840. The processor 810 can invoke the logical instructions in the memory 830 to execute the method in any of the embodiments of the first aspect described above, which includes:

[0196] The vibration signal of the group control device is multi-scale decomposed by a multi-scale input layer to extract time domain features, spectral domain features, and time-frequency features of the vibration signal, and generate multi-dimensional feature data;

[0197] The multi-dimensional feature data is dynamically interactively fused by a cross-scale feature coupling module, the recognition degree of weak fault features is enhanced by using a bidirectional gate mechanism, and a fused 32-dimensional feature vector is output;

[0198] A device relationship perception graph convolution model is constructed by the 32-dimensional feature vector, a dynamic adjacency matrix is generated based on device physical distance and real-time working condition similarity, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix;

[0199] The spatio-temporal fusion feature is processed by an occasional fault sensitive time sequence module, a Teager energy operator is used to enhance transient impact features, a self-adaptive comb filter is used to extract periodic fault features, and a time sequence feature is output by using a dual-channel weight fusion and a self-attention mechanism;

[0200] Based on a double-threshold detection mechanism and a device comparison enhancement strategy, the time sequence feature is subjected to fault recognition and positioning to generate an explainable diagnostic evidence chain containing multi-scale feature contributions.

[0201] In addition, the logical instructions in the memory 830 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0202] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to enable the computer to execute the method provided by any of the above methods, and the method comprises:

[0203] The vibration signal of the group control device is subjected to multi-scale decomposition through a multi-scale input layer, time domain features, spectral domain features and time-frequency features of the vibration signal are extracted, and multi-dimensional feature data is generated;

[0204] The multi-dimensional feature data is subjected to dynamic interactive fusion through a cross-scale feature coupling module, the recognition degree of weak fault features is enhanced by using a bidirectional gate mechanism, and a fused 32-dimensional feature vector is output;

[0205] A device relationship perception graph convolution model is constructed through the 32-dimensional feature vector, a dynamic adjacency matrix is generated based on device physical distance and real-time working condition similarity, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix;

[0206] The spatio-temporal fusion feature is processed through an occasional fault sensitive time sequence module, a Teager energy operator is used to enhance transient impact features, periodic fault features are extracted by combining an adaptive comb filter, and a time sequence feature is output by using a double-channel weight fusion and a self-attention mechanism;

[0207] Based on a double-threshold detection mechanism and a device comparison enhancement strategy, the time sequence feature is subjected to fault recognition and positioning, and an interpretable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0208] In another aspect, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement a cigarette case image recognition method provided by any of the above methods, and the method comprises:

[0209] The vibration signal of the group control device is subjected to multi-scale decomposition through a multi-scale input layer, time domain features, spectral domain features and time-frequency features of the vibration signal are extracted, and multi-dimensional feature data is generated;

[0210] The multi-dimensional feature data is subjected to dynamic interactive fusion through a cross-scale feature coupling module, the recognition degree of weak fault features is enhanced by using a bidirectional gate mechanism, and a fused 32-dimensional feature vector is output;

[0211] A device relationship perception graph convolution model is constructed through the 32-dimensional feature vector, a dynamic adjacency matrix is generated based on device physical distance and real-time working condition similarity, and a spatio-temporal fusion feature is obtained by performing a spatio-temporal convolution operation on the dynamic adjacency matrix;

[0212] The spatio-temporal fusion features are processed by an accidental fault sensitive timing module, transient impact features are enhanced using a Teager energy operator, periodic fault features are extracted using an adaptive comb filter, and a double-channel weight fusion and a self-attention mechanism are used to output timing features.

[0213] Based on a double-threshold detection mechanism and a device contrast enhancement strategy, the timing features are subjected to fault identification and positioning, and an explainable diagnostic evidence chain containing multi-scale feature contributions is generated.

[0214] Embodiment 2

[0215] Step 1: Multi-scale input layer construction:

[0216] The multi-scale input layer has a three-stream parallel processing structure designed to cover all types of fault features in the operation of rotating machinery, providing rich and effective original feature data for subsequent feature coupling and analysis.

[0217] Scale 1: Raw signal branch (time domain features):

[0218] IEPE acceleration sensors (range ±50g) are used to collect three-axis vibration signals (X / Y / Z directions) of the device, with a sampling frequency of 12.8KHz. According to the Nyquist criterion, this sampling frequency can fully cover the key fault frequencies of rotating machinery in the 0-5kHz range.

[0219] A second-order Butterworth high-pass filter (cutoff frequency 1Hz) is used to process the collected raw signals to eliminate the DC component interference caused by sensor zero drift; a 3σ criterion is used to detect and replace abnormal points in the signal, when the signal value satisfies , the abnormal point is replaced with the mean value of the signal, where is the mean value of the signal, and σ is the standard deviation of the signal; the above signal is normalized to map the signal to the range [-1,1].

[0220] Time domain features are extracted from three dimensions of impact intensity, distribution pattern, and energy change, which are composed as follows:

[0221] 1. Transient impact features (4 dimensions):

[0222] Peak value: the maximum value of the signal within the time window, directly reflecting the instantaneous impact intensity (such as the instantaneous vibration peak of bearing crack).

[0223] Kurtosis: calculated by the formula , quantifying the degree of deviation of the impact component in the signal (normal signal kurtosis ≈3, fault can reach 5 or more).

[0224] Impulse factor: the ratio of peak value to signal average, highlighting high-impact signals in low-amplitude background.

[0225] Impulse duration: the number of consecutive sampling points that the signal exceeds the threshold (3σ of normal signal), reflecting the time span of fault impact.

[0226] 2. Waveform distribution characteristics (3D):

[0227] Waveform factor: the ratio of root mean square value to average value, characterizing the smoothness of signal waveform (wear fault waveform is relatively flat, crack fault waveform is steep).

[0228] Skewness: describes the symmetry of signal amplitude distribution, and the absolute value of skewness increases due to the unidirectionality of impact signals during fault.

[0229] Peak frequency: the dominant frequency of impact signals obtained by short-time Fourier transform, related to the inherent frequency of faulty components (such as the impact frequency of bearing outer ring fault).

[0230] 3. Energy change characteristics (3D):

[0231] High-frequency energy proportion: the ratio of 2-5 kHz band energy to 0-5 kHz total energy, highlighting high-frequency impact components (such as high-frequency vibration of crack propagation).

[0232] Energy entropy: the energy value in a 10 ms window , where is the normalized signal value, the entropy value calculated based on the energy distribution of 100 sub-windows, reflecting the uniformity of energy distribution (the entropy value decreases due to energy concentration during fault).

[0233] 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 impact signals is significantly higher than that of normal signals).

[0234] 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.

[0235] Scale 2: envelope spectrum branch (frequency domain features):

[0236] This branch mainly focuses on the modulation characteristics of fault signals (such as sideband modulation of bearing faults), extracts characteristic frequency components through envelope analysis, thereby achieving accurate identification of equipment faults.

[0237] Convert the real signal to an analytical signal, separate the envelope signal and the carrier signal, and the calculation formula is:

[0238]

[0239] where, is the original signal, is the Hilbert transform result, is the envelope signal. Through Hilbert transform, the envelope information in the signal can be effectively extracted, highlighting the modulation characteristics caused by faults.

[0240] A 5th order Butterworth low-pass filter (cutoff frequency 2kHz) is used to smooth the envelope signal, remove the interference of high-frequency noise, and make the envelope signal more clearly reflect the fault characteristics. The envelope signal after low-pass filtering is subjected to 2048-point fast Fourier transform (FFT) to obtain the envelope spectrum .

[0241] The characteristic frequency band is intercepted from 0-2kHz frequency band (which covers the key fault frequencies such as bearing BPFO / BPFI, gear meshing frequency, etc.), and the following features are extracted:

[0242] 1. Basic frequency characteristics (3D):

[0243] Characteristic frequency peak (1D): that is, , which directly corresponds to the core frequency of the fault (such as bearing BPFO, gear meshing frequency), and is the core basis for judging the fault type.

[0244] Peak frequency energy (1D): the energy value at the characteristic frequency peak , which quantifies the intensity of the fault signal.

[0245] Peak frequency bandwidth (1D): 3db bandwidth on both sides of the characteristic frequency peak, reflecting the dispersion degree of the fault frequency (wear type fault bandwidth is wider, crack type fault bandwidth is narrower).

[0246] 2. Sideband characteristics (2D):

[0247] Sideband energy ratio (1D): the proportion of energy in the range of ±100Hz around the characteristic frequency to the total energy, used to evaluate the modulation depth (such as the increase of sideband energy caused by gear eccentricity).

[0248] Sideband frequency interval (1D): the frequency difference between adjacent sidebands, corresponding to the device rotation frequency (such as the modulation frequency related to motor speed), which assists in locating the fault source (such as the fault synchronized with the rotation speed).

[0249] 3. Spectral distribution characteristics (3D):

[0250] Spectral entropy (1D): the calculation formula is , where, is the frequency The proportion of energy at a given frequency to the total energy. When a fault occurs, the spectral entropy value will decrease because the fault causes energy to concentrate at a specific frequency.

[0251] 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.

[0252] Scale 3: Wavelet energy entropy branch (time-frequency characteristics):

[0253] Wavelet packet decomposition captures the non-stationary characteristics of signals, making it suitable for analyzing the time-frequency distribution of sporadic faults. The db4 wavelet basis is selected as the decomposition basis function, employing a three-level decomposition to divide the signal into eight 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, respectively. Wavelet packet reconstruction is performed on each sub-band to obtain... ,in, The number of decomposition layers, Sub-band numbering.

[0254] In the wavelet energy entropy branch at scale 3, the output time-frequency features are 16-dimensional, designed from three dimensions: energy distribution, frequency resolution, and time evolution. The specific structure is as follows:

[0255] 1. Multi-scale energy entropy characteristics (8 dimensions):

[0256] Based on the eight sub-bands obtained from the three-level wavelet packet decomposition, the energy entropy of each sub-band is calculated: ,in, For the first Layer Energy proportion in sub-bands. When equipment malfunctions, energy concentrates in specific sub-bands, causing a decrease in energy entropy.

[0257] 2. Sub-band energy characteristics (6 dimensions):

[0258] From the eight sub-bands, six key frequency bands with the strongest correlation to the fault were selected (determined through mutual information analysis), and their energy values ​​were extracted: These six features quantify the vibration energy intensity in different frequency bands. Among them, 312.5Hz and 937.5Hz correspond to the gear meshing frequency and harmonics, and the energy change reflects the degree of gear wear; 2187.5Hz and 2812.5Hz cover the bearing rolling element passing frequency, and a sudden increase in energy indicates bearing defects; 4062.5Hz and 4687.5Hz are used to capture high-frequency impacts (such as instantaneous crack propagation).

[0259] 3. Time-frequency domain dynamic feature (2D):

[0260] Energy entropy rate of change: Calculate the average rate of change of 8 energy entropies in adjacent time windows, the formula is:

[0261] This feature can sensitively capture the mutation of energy distribution when the fault occurs (such as sudden drop in entropy value caused by accidental impact).

[0262] High-frequency energy proportion: The ratio of the sum of the energy of two high-frequency subbands (4062.5Hz and 4687.5Hz) to the total energy, highlighting the high-frequency characteristics of transient impact.

[0263] 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.

[0264] Step 2: Cross-scale feature coupling module:

[0265] The cross-scale feature coupling module is the key to solving the problem of weak fault feature submersion. Its core function is to realize the dynamic interaction and complementary enhancement of different scale features through a bidirectional gating mechanism, breaking through the limitations of traditional multi-scale methods "feature conflict" and "information loss", and improving the recognition of fault features.

[0266] To ensure that the features of the three scales are consistent in magnitude, facilitating subsequent feature coupling processing, the features of each scale are standardized, the formula is:

[0267]

[0268] 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 will be used as the input of the cross-scale coupling module.

[0269] Since the feature dimensions of the three scales are different, in order to realize the effective interaction between the features, it is necessary to unify the dimensions. The , , are unified to C=32 through 1D convolution (kernel size 3, step 1), and the calculation formula is:

[0270] where, is the 1D convolution operation, represents the number of channels of the convolution output.

[0271] The three dimensional features after dimension alignment are spliced in the channel dimension to obtain:

[0272]

[0273] In the formula, is the number of devices, is the time step, and 96 is the feature dimension after splicing (32x3).

[0274] The importance weight of each scale feature is learned by 1D convolution and softmax function, and the formula is:

[0275]

[0276] In the formula, , , , The corresponding three weights are respectively, and satisfy .

[0277] The features of different scales are bidirectionally coupled by the learned weights to realize complementary enhancement of the features, wherein:

[0278] Scale 1-2 coupling: strengthen the correlation between the time domain impact feature and the frequency of the frequency domain feature, and the formula is:

[0279]

[0280] Scale 2-3 coupling: strengthen the correlation between the frequency domain feature and the non-stationary time-frequency feature, and the formula is:

[0281]

[0282] In the formula, represents the element multiplication operation. Through weight adjustment, important features can dominate in the coupling process to realize effective enhancement of the features.

[0283] In order to avoid losing key information of the original scale feature in the feature coupling process, a residual connection is used to add the original scale feature and the coupled feature to obtain the final output feature, and the formula is:

[0284]

[0285] After the cross-scale feature coupling process, the dimension of the output feature is NsensorsxTx32.

[0286] Step 3: Device relationship perception graph convolution:

[0287] The core function of the device relationship-aware graph convolutional layer is to model the vibration transmission relationship between group control devices by constructing a dynamic adjacency matrix. This solves the problem 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.

[0288] 1. Definition of graphical model:

[0289] Each device's sensor is treated as a node in the graph, and the node features are those resulting from cross-scale coupling.

[0290]

[0291] In the formula, For the first The characteristics of each node For the first The cross-scale coupling characteristics of the equipment, where T is the time step and 32 is the feature dimension.

[0292] 2. Dynamic adjacency matrix construction:

[0293] The vibration transmission intensity between devices is represented by a dynamic adjacency matrix. Where N is the number of devices. The value range of is [0,1]. The larger the value, the more likely the device is to be larger. With equipment The stronger the vibrational correlation between them.

[0294] Here, the dynamic adjacency matrix From the physical similarity matrix Similarity matrix with features It is a weighted fusion, and the formula is:

[0295]

[0296] In the formula, This is a dynamic fusion coefficient whose value is adjusted in real time according to the device's operating status. It is used to balance the contributions of physical similarity and feature similarity in the adjacency matrix.

[0297] Physical similarity matrix: The main considerations are the impact of physical distance and speed differences between equipment on vibration transmission.

[0298] Physical distance factor: based on the device's installation coordinates computing devices With equipment Euclidean distance between Then, the influence of physical distance on vibration transmission is quantified using an exponential decay model, with the following formula:

[0299]

[0300] where, is the average distance between equipment in the plant (5m in the experiment). This formula shows that the closer the distance between equipment, the greater the value of the physical distance factor, and the stronger the vibration conduction intensity.

[0301] Speed difference factor: the speed difference of equipment will affect the vibration coupling intensity between equipment, the closer the speed, the stronger the vibration coupling. The speed difference factor is modeled by the Sigmoid function, and the formula is:

[0302]

[0303] where, is the speed difference coefficient, usually 0.01, and are the real-time speeds of equipment and equipment respectively.

[0304] Multiply the physical distance factor and the speed difference factor to get the physical similarity matrix:

[0305]

[0306] Characteristic similarity matrix: Mainly consider the influence of characteristic similarity and load difference between equipment on vibration correlation, based on cross-scale coupling characteristics

[0307]

[0308] where, denotes the inner product of vectors, denotes the norm of the vector. The closer the value of the characteristic cosine similarity to 1, the more similar the characteristics of the two equipment, and the stronger the vibration correlation.

[0309] Load difference filtering: when the load difference between equipment , the reflection effect of characteristic similarity on vibration correlation will be affected, so filter through the indicator function , the formula is:

[0310]

[0311] where, is the indicator function, which takes the value of 1 when the condition in the parentheses is satisfied, otherwise 0.

[0312] Dynamic fusion coefficient α: the value of α is dynamically adjusted by the physical distance between equipment, and the formula is:

[0313]

[0314] where, is the maximum distance between equipment in the plant, is the sensitivity control coefficient.

[0315] 3. Spatio-temporal convolution operation:

[0316] Based on the dynamic adjacency matrix , the spatio-temporal correlation between equipment is captured through 2-layer Gated GCN. Among them, the spatial graph convolution calculation is:

[0317]

[0318] where, , is the degree matrix, is the identity matrix, , is the node feature of the layer, denotes the layer normalization operation, is the spatial convolution weight of the layer (dimension 32x64). is the activation function (ReLU function is selected).

[0319] Temporal convolution: 1D temporal convolution (kernel size 3) is performed on the spatial convolution output features to capture the correlation relationship on the time sequence, and the formula is:

[0320]

[0321] where, denotes the 1D convolution operation, denotes the convolution kernel size.

[0322] The flow of features is controlled through the gating unit to alleviate the common oversmoothing problem in graph convolution networks, and the formula is:

[0323]

[0324] where, and are the weights and biases of the gating unit, is the sigmoid activation function, and is the element multiplication operation. After 2-layer convolution operation, the spatio-temporal fusion features are obtained.

[0325] Step 4: Occasional fault sensitive time sequence module:

[0326] ​​​​​​The sporadic fault sensitive timing module is designed for the characteristics of sporadic fault "short transient" (usually lasting for 0.5-2s), and through double-channel feature enhancement and gate mechanism, the extraction of transient impact features is strengthened, and noise interference is inhibited, solving the problem of delayed response of traditional timing model to sporadic faults.

[0327] 1. Transient-duration double threshold detection:

[0328] The module includes two parallel paths and a gated fusion unit, and the specific structure is as follows:

[0329] Teager energy operator (TEO): The feature H output by the spatiotemporal graph convolution is processed to enhance the transient impact feature, and the formula is:

[0330]

[0331] Compared with the traditional energy calculation method, TEO is more sensitive to the instantaneous energy change.

[0332] Short-time zero-crossing rate: The number of zero-crossing points of the signal in each 10ms window is calculated to quantify the high-frequency characteristics of the impact, and the formula is:

[0333]

[0334] In the formula, sign() is the sign function, and T is the window size (T=10ms here).

[0335] The feature H output by the spatiotemporal graph convolution is processed to enhance the transient impact feature, and the formula is:

[0336]

[0337] For the periodicity of rotating machinery faults (such as periodic impact of bearing faults), a comb filter with adjustable center frequency is designed, and the formula is:

[0338] The module includes two parallel paths and a gated fusion unit, and the specific structure is as follows:

[0339]

[0340] In the formula, is the fundamental frequency of the equipment, B is the bandwidth of the filter, K is the harmonic order, is a control coefficient taking 0-1.

[0341] After processing by the adaptive comb filter, the feature of the period-preserving path is obtained:

[0342]

[0343] The gate parameters are learned by 1D convolution to balance the contributions of transient features and periodic features, and the formula is:

[0344]

[0345] wherein, represents a 1D convolution operation, represents the number of output channels, and the sigmoid function maps the weight value to the range [0, 1].

[0346] The features of the two paths are fused according to the learned weights, and the formula is:

[0347]

[0348] When the device fails, at this time, the transient features dominate in the fusion result; when the device is running normally, the periodic features are dominant.

[0349] 2. Time attention focus:

[0350] Focus on the features at the time of failure through self-attention mechanism to improve the time resolution:

[0351] Query / Key / Value generation:

[0352]

[0353] wherein, is a weight matrix.

[0354] Attention weight calculation:

[0355]

[0356] wherein, is the dimension of Query / Key, is a scaling factor to prevent the weight value from being too large. Multiply the attention weight by the Value to get the time-enhanced feature .

[0357] Step 5: Fault diagnosis strategy:

[0358] The fault diagnosis strategy combines double-threshold detection and device comparison enhancement to achieve accurate identification and positioning of occasional faults and generate interpretable diagnosis evidence chain.

[0359] 1. Double-threshold detection mechanism:

[0360] The time-enhanced feature is reduced to a device-time anomaly vector by global average pooling The abnormality degree is calculated based on the feature distribution of normal operation of the device, and the formula is:

[0361]

[0362] In the formula, is the reduced feature of the i-th device at the j-th time point, is the feature mean value of the i-th device in normal operation, is the corresponding standard deviation. Double threshold setting:

[0363] High quantile threshold : Take the 95% quantile of the normal data abnormality degree;

[0364] Low quantile threshold : Take the 80% quantile of the normal data abnormality degree;

[0365] Threshold dynamic update: In order to adapt to the baseline drift of normal operation characteristics caused by device aging and other factors, the threshold is recalculated every 24 hours using the latest normal data.

[0366] Decision logic:

[0367]

[0368] ;

[0369] If the abnormality degree of the device satisfies for 3 consecutive time steps, ;

[0370] :

[0371] Determine as "suspected sporadic failure" (need manual review);

[0372] :

[0373] Determine as "normal".

[0374] This decision logic can effectively solve the problem of missed detection of sporadic failures by single threshold.

[0375] 2. Device comparison enhancement:

[0376] By comparing the abnormality degrees between group control devices, the weak fault characteristics of single device are strengthened:

[0377] Group device average abnormality degree:

[0378]

[0379] Abnormal offset:​​

[0380]

[0381] Final Score

[0382]

[0383] When the anomaly of a single device is significantly higher than the average anomaly of the group, the final score will be amplified by 1-2 times, thus enhancing the recognition of weak faults.

[0384] 3. Fault location and evidence generation:

[0385] Correlation Impact: Calculate the impact of fault propagation on other devices based on dynamic adjacency matrix, formula:

[0386]

[0387] Location Rule: The fault device is the device with the highest anomaly and significant correlation impact, that is,

[0388]

[0389] In the formula, is the correlation weight.

[0390] Interpretable evidence chain generation:

[0391] Output diagnostic evidence containing multi-scale feature contributions, as follows:

[0392] Scale 1 evidence: Impact energy exceeds multiple times (e.g. "+12.7dB");

[0393] Scale 2 evidence: Feature frequency peak value (e.g. "BPFO=156Hz enhanced 3 times");

[0394] Correlation evidence: Changes in correlation with adjacent devices (e.g. "Correlation with M4 vibration decreased by 40% ");

[0395] Output format: Use structured JSON to facilitate quick understanding and fault source positioning for industrial engineers.

[0396] If not mentioned in this application, existing technology can be used or referred to for implementation.

[0397] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0398] The above merely provides an example of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of 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 into multiple scales by a multi-scale input layer, and the time-domain features, spectral domain features and time-frequency features of the vibration signal are extracted to generate multi-dimensional feature data. The multidimensional feature data is dynamically and interactively fused through a cross-scale feature coupling module, and the identification of weak fault features is enhanced by a two-way gating mechanism, outputting a fused 32-dimensional feature vector. Using the 32-dimensional feature vector, a device relationship-aware graph convolutional model is constructed. A dynamic adjacency matrix is ​​generated based on the physical distance between devices and the similarity of real-time operating conditions. The dynamic adjacency matrix is ​​then subjected to spatiotemporal convolution operations to obtain spatiotemporal fusion features. The spatiotemporal fusion features are processed by an intermittent fault sensitive timing module, the transient impact features are enhanced by the Teager energy operator, the periodic fault features are extracted by an adaptive comb filter, and the timing features are output by dual-channel weight fusion and self-attention mechanism. Based on a dual-threshold detection mechanism and a device comparison enhancement strategy, the time-series features are used for fault identification and localization, generating an interpretable diagnostic evidence chain that includes contributions from multi-scale features.

2. The method according to claim 1, characterized in that, The vibration signal of the group control equipment is decomposed into multiple scales through a multi-scale input layer to extract the time-domain features, spectral domain features, and time-frequency features of the vibration signal, generating multi-dimensional feature data. Specifically: The triaxial vibration signal was acquired using an IEPE accelerometer, and the peak value, impulse factor, and impact duration transient impact characteristics were calculated; waveform factor, skewness waveform distribution characteristics; high-frequency energy proportion, and zero-crossing rate energy change characteristics were calculated. Peak frequency bandwidth, sideband energy ratio, and sideband frequency spacing modulation characteristics were extracted using envelope analysis. Multi-scale energy entropy features, sub-band energy features, and time-frequency domain dynamic features are generated using wavelet transform.

3. The method according to claim 1, characterized in that, The process involves dynamically interacting and fusing the multidimensional feature data through a cross-scale feature coupling module, enhancing the identification of weak fault features using a bidirectional gating mechanism, and outputting a fused 32-dimensional feature vector. Specifically: The multidimensional feature data is standardized. The standardized multidimensional feature data is concatenated along the channel dimension, and the importance weights of features at each scale are learned through one-dimensional convolution and softmax function. Inter-scale feature complementarity enhancement is achieved through a bidirectional gating mechanism, specifically including: Strengthen the correlation between time-domain features and frequency-domain features; Strengthen the correlation between frequency domain features and time-frequency features; The key information of the original features is preserved by using residual connection, and the fused 32-dimensional feature vector is output.

4. The method according to claim 1, characterized in that, The process involves constructing a device relationship-aware graph convolutional model using the 32-dimensional feature vectors, generating a dynamic adjacency matrix based on the physical distance between devices and the similarity of real-time operating conditions, and then performing a spatiotemporal convolution operation on the dynamic adjacency matrix to obtain spatiotemporal fusion features, specifically: Each device's sensor is used as a graph node, and the node features are the 32-dimensional feature vectors. The adjacency matrix weights are dynamically adjusted based on the physical distance between devices and the similarity of real-time operating conditions. The formula is as follows: The physical distance factor and the rotational speed difference factor are multiplied 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 patterns between devices are captured by spatial convolution operations, and then correlated features on the time series are extracted by temporal convolution operations. The feature flow is controlled by gating units to alleviate the oversmoothing problem.

5. The method according to claim 1, characterized in that, The spatiotemporal fusion features are processed through an intermittent fault-sensitive time-series module. The Teager energy operator is used to enhance transient impact features, and an adaptive comb filter is combined to extract periodic fault features. Finally, a dual-channel weighted fusion and self-attention mechanism are used to output the time-series features. Specifically: The transient impact energy is calculated using the Teager energy operator, and transient characteristics are generated by combining the high-frequency characteristics of short-time zero-crossing rate quantization. Periodic fault features are extracted using an adaptive comb filter, where the center frequency of the filter is dynamically adjusted according to the fault type. The weights of transient and periodic features are learned through one-dimensional convolution, and the features of the two pathways are fused according to the weights. By calculating the attention weights at the moment of failure through a self-attention mechanism, the temporal resolution is improved.

6. The method according to claim 1, characterized in that, The method based on the dual-threshold detection mechanism and device contrast enhancement strategy is used to identify and locate faults in the time-series features, generating an interpretable diagnostic evidence chain that includes contributions from multi-scale features. Specifically: Set transient and sustained thresholds to identify anomalies in short-term impact characteristics and periodic characteristics, respectively. The threshold is dynamically updated every 24 hours based on the latest normal data to adapt to baseline drift caused by equipment aging factors; By employing a device comparison enhancement strategy, the deviation between the anomaly degree of a single device and the average anomaly degree of the group of devices is calculated, and an anomaly offset is generated to enhance the characteristics of weak faults.

7. The method according to claim 6, characterized in that, Also includes: The fault propagation impact is calculated based on the dynamic adjacency matrix to determine the fault source device; Generate an interpretable chain of evidence, which includes multi-scale feature contributions, specifically including the impact energy exceeding the standard multiple, the peak value of characteristic frequencies and their change multiples, and the correlation changes with neighboring devices; Output a structured diagnostic report, including the location of the fault, the chain of evidence, and repair recommendations.

8. The method according to claim 4, characterized in that, The generation of the dynamic adjacency matrix includes the following steps: Acquire physical distance data between group control devices and construct a basic distance matrix; Real-time acquisition of operating condition data of each device, and calculation of speed differences and load change similarities between pairs of devices; The distance matrix and the working condition similarity matrix are weighted and fused to generate a dynamic adjacency matrix, where the weight coefficients are preset or adjusted online 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 the processor, it implements the steps of the method as described in any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-8.

Images