A fan fault diagnosis method based on multi-modal data fusion

Abstract

This invention discloses a wind turbine fault diagnosis method based on multimodal data fusion, belonging to the field of auxiliary equipment condition monitoring technology for thermal power units. The method includes: collecting and standardizing multimodal operating data, historical operating data, maintenance record data, and structural configuration data to generate status data, maintenance indexes, and component connection relationships; calibrating asynchronous time drift to generate time-series aligned data; performing phase slicing of operating conditions and recalibrating the same-phase load reference to generate phase residuals; verifying modal reliability and removing conflicting modes to generate markers; constructing a component topology fault propagation residual map; generating fault state representations using an improved liquid time constant network; and outputting fault diagnosis results. This invention achieves accurate wind turbine fault diagnosis, improves the ability to identify weak faults, and enhances fault location stability.

Description

Technical Field

[0001] This invention relates to the field of auxiliary equipment condition monitoring technology for thermal power units, and in particular to a method for diagnosing fan faults based on multimodal data fusion. Background Technology

[0002] Fan fault diagnosis involves determining the operating status of the forced draft fan, induced draft fan, and primary air fan in thermal power units. The diagnostic objects include the fan casing, impeller, rotor, main shaft, bearings, coupling, motor, inlet duct, outlet duct, damper, blade adjustment mechanism, lubrication system, and foundation support. Existing technologies for identifying fan anomalies in thermal power units primarily utilize vibration sensors, acoustic sensors, temperature sensors, current sensors, pressure sensors, airflow measurement points, adjustment opening measurement points, lubrication monitoring devices, and unit operation monitoring devices to collect fan status data. Then, threshold judgment, rule matching, ordinary time series prediction models, recurrent neural networks, convolutional neural networks, and conventional liquid time constant networks are used to preprocess the collected data, extract features, classify faults, and output alarms.

[0003] Current technologies still have shortcomings. Multi-source fan status data is often directly input into the diagnostic model, lacking asynchronous time drift calibration, phase consistency slicing of operating conditions, and generation of in-phase load benchmarks. This results in insufficient ability to distinguish between normal operating condition changes caused by unit load variations, speed variations, airflow regulation, air pressure fluctuations, damper opening changes, blade opening changes, and coal mill operation / disabling, and actual fault anomalies. The multi-modal fusion process lacks reliability self-verification and conflict stripping processes, making sensor drift, short-term jumps, and local noise prone to interfering with diagnostic results. The fault location process lacks component topology fault propagation residual maps, making it difficult to express the diffusion characteristics of faults along mechanical connections, pneumatic flow, vibration transmission, heat conduction, and electrical coupling paths. Ordinary liquid time constant networks lack impact energy amplification, polarized complex liquid evolution, and phase-coupled state propagation structures, resulting in insufficient expression of weak impact faults, amplitude and phase joint changes, and component linkage anomalies corresponding to bearing impact, rotor imbalance, coupling misalignment, impeller ash accumulation, duct blockage, regulating mechanism abnormalities, surge, stall, and motor abnormalities.

[0004] Therefore, how to provide a wind turbine fault diagnosis method based on multimodal data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for diagnosing fan faults based on multimodal data fusion. The fans are the forced draft fan, induced draft fan, and primary air fan of a thermal power unit. This invention utilizes operating condition phase residuals, modal confidence markers, conflict mode markers, component topology fault propagation residual maps, and improved liquid time constant networks to achieve fault identification, fault component location, fault level determination, and diagnostic evidence chain generation for thermal power unit fans. It has the advantages of reducing interference from changes in operating conditions, improving the accuracy of weak fault identification, and enhancing the clarity of fault location.

[0006] A wind turbine fault diagnosis method based on multimodal data fusion according to an embodiment of the present invention includes:

[0007] Collect multimodal operation data, historical operation data, maintenance record data, and wind turbine structural configuration data. Standardize the multimodal operation data and historical operation data to generate a standardized multimodal status dataset. Generate maintenance record index data based on maintenance record data and generate wind turbine component connection relationships based on wind turbine structural configuration data.

[0008] Asynchronous time drift calibration is performed based on a standardized multimodal state dataset to generate a time-aligned multimodal dataset;

[0009] Perform phase-consistent slicing of the time-aligned multimodal dataset, remove historical abnormal segments based on maintenance record index data and generate in-phase load benchmarks, calculate the phase residuals of each mode based on the in-phase load benchmarks, and generate a phase residual sequence of the operating conditions.

[0010] Perform multimodal confidence self-verification and conflict stripping on the phase residual sequence of the operating condition to generate modal confidence tags and conflict mode tags;

[0011] Based on the connection relationship of wind turbine components, a component topology graph is constructed. The operating condition phase residual sequence, modal confidence label and conflict mode label are mapped to the component topology graph. Adjacency response residual, reverse source residual and cross-node consistency residual are calculated to generate a component topology fault propagation residual map.

[0012] An improved liquid time constant network is constructed. The phase residual sequence of operating conditions, modal confidence label, conflict mode label and component topology fault propagation residual map are input into the improved liquid time constant network. Impulse fault feature enhancement, complex liquid state evolution and component phase coupling propagation are performed to generate a wind turbine fault state characterization.

[0013] Based on the characterization of wind turbine fault status, output wind turbine fault diagnosis results.

[0014] Optionally, the multi-modal operation data of the fan includes fan type marking, vibration data, acoustic data, temperature data, current data, voltage data, power data, speed data, pressure data, air volume data, regulation opening data, lubricating oil status data, and unit-related operation data, wherein the unit-related operation data includes unit load data and coal mill commissioning status data;

[0015] The historical operating data includes historical data corresponding to the multimodal operating data of the wind turbine;

[0016] The maintenance record data includes fault repair time, downtime alarm time, component replacement time, manual inspection abnormality markers, and repair component markers;

[0017] The fan structure configuration data includes the mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship between the fan body, rotating bearing, drive motor, inlet air duct, outlet air duct, regulating mechanism, lubrication system and foundation support.

[0018] Optionally, generating the standardized multimodal state dataset includes:

[0019] Read the acquisition time, sampling frequency, wind turbine type label, component source, and original characteristic values ​​from the multimodal operation data and historical operation data of the wind turbine;

[0020] Perform outlier removal, missing data completion, and dimensional standardization on the original feature values ​​to generate standardized feature values;

[0021] By associating the collection time, sampling frequency, wind turbine type label, component source, and standardized feature values, a standardized multimodal state dataset is generated.

[0022] Optionally, the generation of the time-aligned multimodal dataset includes:

[0023] Read the acquisition time, sampling frequency, wind turbine type label, component source and standardized feature value from the standardized multimodal state dataset, and establish a unified diagnostic time axis according to the preset diagnostic time window;

[0024] Vibration data, acoustic data, and current data are labeled as high-frequency modal data. Window features are extracted from the high-frequency modal data to generate a high-frequency window feature sequence.

[0025] The multimodal operation data of the wind turbine, excluding high-frequency modal data, is marked as low-frequency modal data and mapped to a preset diagnostic time window according to the acquisition time to generate a low-frequency window state sequence;

[0026] Modal time drift is generated based on the offset between the acquisition time of each modality and the center time of the preset diagnostic time window. The high-frequency window feature sequence and the low-frequency window state sequence are corrected according to the modal time drift to generate a time-aligned multimodal dataset.

[0027] Optionally, the generation of the phase residual sequence for the operating condition includes:

[0028] Read the time-aligned multimodal dataset and generate a sequence of operating condition change features based on the differences in unit load, speed, power, pressure, air volume, regulation opening, and coal mill operation status between adjacent preset diagnostic time windows.

[0029] Based on the sequence of operating condition changes, the starting speed-up phase, load ramp-up phase, stable ventilation phase, air volume adjustment phase, coal mill start-up / shutdown disturbance phase, and shutdown speed-down phase are divided to generate operating condition phase slice data;

[0030] Read the maintenance record index data and mark the historical operation data segments corresponding to fault repair time, downtime alarm time, component replacement time, manual inspection anomaly mark and repair component mark as historical anomaly segments;

[0031] From the historical operating data after removing historical abnormal segments, select data that have the same fan type, operating condition phase, preset unit load range, preset speed range, preset air volume range, preset adjustment opening range, coal mill commissioning status and component source as the current preset diagnostic time window, and generate the same phase load benchmark according to the median of the standardized characteristic values ​​of each mode;

[0032] Subtract the standardized feature value of the corresponding mode in the same phase load benchmark from the standardized feature value of each mode within the current preset diagnostic time window to obtain the vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, pressure phase residual, air volume phase residual, power phase residual and lubricating oil state phase residual. Then, arrange them according to the acquisition time, fan type label and component source to generate the working condition phase residual sequence.

[0033] Optionally, the generation of modal confidence markers and conflict modal markers includes:

[0034] Read the phase residual sequence of the operating condition and establish a modal residual alignment table according to the same preset diagnostic time window, the same fan type label and the same component source;

[0035] Count the number of missing data, the length of consecutive missing data, and the number of valid residuals for each modality, and generate data integrity markers;

[0036] The historical residual fluctuation range is generated based on the residual distribution between historical operating data after removing historical abnormal segments and the load benchmark with the same phase, and the amplitude of the phase residual change of each mode is calculated based on the modal residual alignment table to generate short-term stability markers.

[0037] A trend consistency mark is generated based on the increase or decrease direction of the phase residuals of each mode within the same preset diagnostic time window. A response consistency mark of adjacent components is generated based on the order of residual occurrence and the direction of residual increase or decrease of adjacent components within consecutive preset diagnostic time windows.

[0038] Modal credibility labels are generated based on the combination of data integrity labels, short-term stability labels, trend consistency labels, and adjacent component response consistency labels. Modalities that are missing continuity, have short-term jumps, have opposite trend directions, and have no corresponding responses from adjacent components are labeled as conflicting modes, and conflicting mode labels are generated.

[0039] Optionally, the generation of conflict mode tags includes:

[0040] Integrity scores are generated based on the ratio of the number of valid residuals to the number of expected residuals for each modality, and integrity deductions are generated based on the length of consecutive missing values.

[0041] A stability score is generated based on the comparison between the variation amplitude of phase residuals in each mode and the historical residual fluctuation range.

[0042] A trend score is generated based on the number of times the phase residuals of each mode increase or decrease in the same direction within the same preset diagnostic time window, and an adjacency response score is generated based on the number of times adjacent components exhibit residual responses in the same direction within consecutive preset diagnostic time windows.

[0043] Modal confidence scores are generated based on integrity scores, integrity deductions, stability scores, trend scores, and adjacency response scores. Modal confidence labels are then generated according to the intervals to which the modal confidence scores belong.

[0044] Continuously missing modes, short-term abrupt change modes, and modes with opposite trends and no corresponding response from adjacent components are written into conflict mode tags.

[0045] Optionally, the generation of the component topology fault propagation residual map includes:

[0046] Read the connection relationship of the fan components, set the fan body, rotating bearing, drive motor, inlet air duct, outlet air duct, regulating mechanism, lubrication system and foundation support as component nodes, write component node mark for each component node, set mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship as connection edge, and generate component topology diagram;

[0047] The phase residual sequence of the operating condition is mapped to the component node according to the component source, and the modal confidence label and conflict mode label are mapped to the component node and the connection edge to generate component residual topology data.

[0048] Based on the phase residual difference, residual increase / decrease direction and modal confidence marker of the component nodes at both ends of the connection edge within the same preset diagnostic time window, the adjacent response residual is calculated.

[0049] Starting from the component node with the highest residual magnitude, backtrack along the connecting edge and calculate the reverse source residual based on the order of residual occurrence and the difference in residual intensity between the upstream and downstream component nodes.

[0050] Based on the order of residual occurrence, direction of residual increase / decrease, and type of connection edge of multiple component nodes within a continuous preset diagnostic time window, cross-node consistency residuals are calculated. Then, by combining adjacency response residuals, reverse source residuals, cross-node consistency residuals, modal confidence markers, and conflict modal markers, a component topology fault propagation residual map is generated.

[0051] Optionally, the generation of wind turbine fault state characterization includes:

[0052] An improved liquid time constant network is constructed, which includes a reliable mode selection layer, an impact energy amplification layer, a polarization complex liquid evolution layer, a phase coupling state propagation layer, and a conflict mode bypass verification layer.

[0053] The operating condition phase residual sequence, modal confidence label, and conflict mode label are input into the confidence mode selection layer to extract the confidence residual data and conflict residual data, and generate the confidence mode input sequence and conflict mode bypass sequence.

[0054] The reliable modal input sequence is input into the impact energy amplification layer, and the vibration phase residual sudden increase segment, acoustic phase residual pulse segment, current phase residual period shift segment, pressure phase residual oscillation segment and air volume phase residual fluctuation segment are extracted to generate the impact-enhanced liquid input sequence.

[0055] Phase shift is generated based on the phase residual difference between consecutive preset diagnostic time windows in the impact-enhanced liquid input sequence. The amplitude residual is written into the real part of the complex liquid state, and the phase shift is written into the imaginary part of the complex liquid state. Continuous state evolution is performed along the liquid time constant update path to generate a complex liquid evolution state.

[0056] The complex liquid evolution state and the component topology fault propagation residual map are input into the phase coupling state propagation layer. The component phase coupling propagation is performed along the mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship to generate the topology coupling liquid state.

[0057] The conflict mode bypass sequence is input into the conflict mode bypass verification layer to generate conflict mode verification evidence. Based on the topologically coupled liquid state and the conflict mode verification evidence, a wind turbine fault state characterization is generated.

[0058] Optionally, the output fan fault diagnosis results include:

[0059] Read the topologically coupled liquid state, conflict mode verification evidence, component node markers, and preset diagnostic time window markers from the wind turbine fault state characterization;

[0060] Perform fault category decoding on the topologically coupled liquid state to generate fault categories;

[0061] The location of the faulty component is generated based on the component node markers and the residual concentration locations in the component topology fault propagation residual map.

[0062] Fault levels are classified based on the intensity of anomalies in the fault state characterization of wind turbines and the number of continuously preset diagnostic time windows.

[0063] Based on modal confidence markers, conflict modal verification evidence, adjacent response residuals, reverse source residuals, and cross-node consistency residuals, diagnostic confidence and diagnostic evidence chains are generated. Then, the fault category, fault component location, fault level, diagnostic confidence and diagnostic evidence chains are combined to output the wind turbine fault diagnosis results.

[0064] The beneficial effects of this invention are:

[0065] This invention proposes a multimodal data fusion-based method for fan fault diagnosis, where the fans are the forced draft fan, induced draft fan, and primary air fan of a thermal power unit. By constructing a process for acquiring multimodal operating data of the fans, asynchronous time drift calibration, phase consistency slicing of operating conditions, generation of in-phase load benchmarks, and calculation of phase residuals of operating conditions, vibration data, acoustic data, temperature data, current data, voltage data, power data, speed data, pressure data, air volume data, regulation opening data, lubricating oil status data, and unit-related operating data are transformed into structured residual input data for fault diagnosis. Compared with fixed threshold judgment, single sensor alarm, and ordinary time-series classification methods, this invention can reduce the interference of unit load changes, speed changes, air volume regulation, pressure fluctuations, regulation opening changes, coal mill start-up / shutdown disturbances, and sensor time drift on the diagnostic results, improving the stability of fan fault identification in thermal power units under varying operating conditions.

[0066] This invention generates modal credibility markers and conflict mode markers through multimodal credibility self-verification and conflict stripping. It combines component topology fault propagation residual maps to express the fault propagation relationships between the wind turbine body, rotating bearing, drive motor, inlet duct, outlet duct, regulating mechanism, lubrication system, and foundation support. Furthermore, it utilizes an improved liquid time constant network, incorporating impact energy amplification, polarized complex liquid evolution, and phase-coupled state propagation, to generate wind turbine fault state representations. Compared to conventional liquid time constant networks and ordinary multimodal fusion models, this invention enhances the ability to express weak impact faults, amplitude and phase joint changes, and component linkage anomalies corresponding to bearing anomalies, rotor imbalance, coupling misalignment, impeller anomalies, duct blockage, regulating mechanism anomalies, surge, stall, motor anomalies, and lubrication system anomalies. This improves the accuracy of fault category identification, fault component location, fault level determination, and diagnostic evidence chain generation. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is an overall flowchart of a wind turbine fault diagnosis method based on multimodal data fusion proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the construction of the component topology fault propagation residual map in a wind turbine fault diagnosis method based on multimodal data fusion proposed in this invention;

[0070] Figure 3 This is a schematic diagram of the improved liquid time constant network structure of a wind turbine fault diagnosis method based on multimodal data fusion proposed in this invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0072] refer to Figure 1 , Figure 2 and Figure 3 A wind turbine fault diagnosis method based on multimodal data fusion includes:

[0073] Collect multimodal operation data, historical operation data, maintenance record data, and wind turbine structural configuration data. Standardize the multimodal operation data and historical operation data to generate a standardized multimodal status dataset. Generate maintenance record index data based on maintenance record data and generate wind turbine component connection relationships based on wind turbine structural configuration data.

[0074] Asynchronous time drift calibration is performed based on a standardized multimodal state dataset to generate a time-aligned multimodal dataset;

[0075] Perform phase-consistent slicing of the time-aligned multimodal dataset, remove historical abnormal segments based on maintenance record index data and generate in-phase load benchmarks, calculate the phase residuals of each mode based on the in-phase load benchmarks, and generate a phase residual sequence of the operating conditions.

[0076] Perform multimodal confidence self-verification and conflict stripping on the phase residual sequence of the operating condition to generate modal confidence tags and conflict mode tags;

[0077] Based on the connection relationship of wind turbine components, a component topology graph is constructed. The operating condition phase residual sequence, modal confidence label and conflict mode label are mapped to the component topology graph. Adjacency response residual, reverse source residual and cross-node consistency residual are calculated to generate a component topology fault propagation residual map.

[0078] An improved liquid time constant network is constructed. The phase residual sequence of operating conditions, modal confidence label, conflict mode label and component topology fault propagation residual map are input into the improved liquid time constant network. Impulse fault feature enhancement, complex liquid state evolution and component phase coupling propagation are performed to generate a wind turbine fault state characterization.

[0079] Based on the characterization of wind turbine fault status, output wind turbine fault diagnosis results.

[0080] In this embodiment, the multimodal operation data of the fan includes fan type marking, vibration data, acoustic data, temperature data, current data, voltage data, power data, speed data, pressure data, air volume data, adjustment opening data, lubricating oil status data, and unit-related operation data. The unit-related operation data includes unit load data and coal mill commissioning status data.

[0081] The historical operating data includes historical data corresponding to the multimodal operating data of the wind turbine;

[0082] The maintenance record data includes fault repair time, downtime alarm time, component replacement time, manual inspection abnormality markers, and repair component markers;

[0083] The fan structure configuration data includes the mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship between the fan body, rotating bearing, drive motor, inlet air duct, outlet air duct, regulating mechanism, lubrication system and foundation support.

[0084] In this embodiment, generating a standardized multimodal state dataset includes:

[0085] Read the acquisition time, sampling frequency, wind turbine type label, component source, and original characteristic values ​​from the multimodal operation data and historical operation data of the wind turbine;

[0086] Perform outlier removal, missing data completion, and dimensional standardization on the original feature values ​​to generate standardized feature values;

[0087] By associating the collection time, sampling frequency, wind turbine type label, component source, and standardized feature values, a standardized multimodal state dataset is generated.

[0088] In this embodiment, generating the time-aligned multimodal dataset includes:

[0089] Read the acquisition time, sampling frequency, wind turbine type label, component source, and standardized feature values ​​from the standardized multimodal state dataset, and establish a unified diagnostic time axis according to the preset diagnostic time window. Specifically, establishing the unified diagnostic time axis according to the preset diagnostic time window involves:

[0090] Set the diagnostic time window length to 10s, the sliding step size to 5s, the starting point of the first diagnostic time window to the earliest acquisition time in the standardized multimodal state dataset, the center time of each diagnostic time window to the window starting point plus 5s, and write the window starting point and window ending point into the diagnostic time window index table.

[0091] Vibration data, acoustic data, and current data are labeled as high-frequency modal data. Window features are extracted from the high-frequency modal data to generate a high-frequency window feature sequence. Specifically, the extraction of window features from the high-frequency modal data involves:

[0092] Read high-frequency modal data from the same component source and the same data mode within each preset diagnostic time window, extract the maximum value of the standardized feature value as the peak value, extract the arithmetic mean of the standardized feature values ​​as the mean, extract the average of the squared differences between each standardized feature value and the mean as the variance, and extract the number of data points exceeding the mean plus twice the standard deviation within the same preset diagnostic time window as the pulse count;

[0093] Multimodal operation data of the wind turbine, excluding high-frequency modal data, is labeled as low-frequency modal data. This data is then mapped to a preset diagnostic time window according to the acquisition time, generating a low-frequency window state sequence. Specifically, the generation of the low-frequency window state sequence is as follows:

[0094] According to the mapping window number, low-frequency modal data with the same component source and the same data modality are aggregated. When there is one low-frequency modal data in the same preset diagnostic time window, the standardized feature value is directly read as the window status value. When there are more than two low-frequency modal data in the same preset diagnostic time window, the median of the standardized feature values ​​is taken as the window status value. When there is no low-frequency modal data in the same preset diagnostic time window, the window status value of the previous preset diagnostic time window is read as the filling value. The mapping window number, window center time, component source, data modality, and window status value are written into the low-frequency window status table to generate a low-frequency window status sequence.

[0095] Modal time drift is generated based on the offset between the acquisition time of each modality and the center time of the preset diagnostic time window. High-frequency window feature sequences and low-frequency window state sequences are then corrected according to the modal time drift to generate a time-aligned multimodal dataset, where:

[0096] Modal time drift is generated based on the offset between the acquisition time of each modality and the center time of the preset diagnostic time window. Specifically:

[0097] Read the original acquisition time set corresponding to each window record in the high-frequency window feature sequence and the low-frequency window state sequence, calculate the arithmetic mean of all acquisition times in the original acquisition time set to obtain the modal actual center time, subtract the corresponding preset diagnostic time window center time from the modal actual center time to obtain the modal time drift. When the modal time drift is positive, it means that the modal data is later than the unified diagnostic time axis, and when the modal time drift is negative, it means that the modal data is earlier than the unified diagnostic time axis.

[0098] The high-frequency window feature sequence is corrected as follows:

[0099] When the absolute value of the modal time drift of the same component source and the same data mode in the high-frequency window feature sequence does not exceed 5s, the high-frequency window feature record is moved to the preset diagnostic time window closest to the time after correction according to the modal time drift. When the absolute value of the modal time drift exceeds 5s, the corresponding high-frequency window feature record is written into the time drift anomaly marker.

[0100] The low-frequency window state sequence is corrected as follows:

[0101] When the absolute value of the modal time drift of the same component source and the same data mode in the low-frequency window state sequence does not exceed 5s, the low-frequency window state record is moved to the preset diagnostic time window closest to the corrected time according to the modal time drift. When the absolute value of the modal time drift exceeds 5s, the corresponding low-frequency window state record is written to the time drift anomaly marker.

[0102] In this embodiment, generating the phase residual sequence of the operating condition includes:

[0103] Read the time-aligned multimodal dataset and generate a sequence of operating condition change features based on the differences in unit load, speed, power, pressure, air volume, regulation opening, and coal mill operation status between adjacent preset diagnostic time windows.

[0104] Based on the characteristic sequence of operating conditions, the starting acceleration phase, load ramp-up phase, stable ventilation phase, air volume regulation phase, coal mill activation / deactivation disturbance phase, and shutdown deceleration phase are divided into operating condition phase slice data, generating the following:

[0105] The system reads the unit load difference, speed difference, power difference, pressure difference, air volume difference, regulation opening difference, and coal mill operation status changes from the operating condition change characteristic sequence. Preset diagnostic time windows with speed differences greater than 0.04 and power differences greater than 0.04 are marked as start-up acceleration phases; preset diagnostic time windows with unit load differences greater than 0.05 and power differences greater than 0.03 are marked as load ramp-up phases; and preset diagnostic time windows with absolute values ​​of unit load differences not greater than 0.03 and absolute values ​​of speed differences not greater than 0.03 are marked as load ramp-up phases. Preset diagnostic time windows with absolute pressure difference not exceeding 0.03 and absolute air volume difference not exceeding 0.03 are marked as stable ventilation phases; preset diagnostic time windows with absolute opening difference exceeding 0.02 and absolute air volume difference exceeding 0.03 are marked as air volume regulation phases; preset diagnostic time windows with changes in the coal mill's operating status not equal to 0 are marked as coal mill start-up / shutdown disturbance phases; and preset diagnostic time windows with speed difference less than -0.04 or power difference less than -0.04 are marked as shutdown and speed reduction phases.

[0106] Read the maintenance record index data and mark the historical operation data segments corresponding to fault repair time, downtime alarm time, component replacement time, manual inspection anomaly mark and repair component mark as historical anomaly segments;

[0107] From the historical operating data after removing historical abnormal segments, select data that have the same fan type, operating condition phase, preset unit load range, preset speed range, preset air volume range, preset adjustment opening range, coal mill commissioning status and component source as the current preset diagnostic time window, and generate the same phase load benchmark according to the median of the standardized characteristic values ​​of each mode;

[0108] Subtract the standardized feature value of the corresponding mode in the same phase load benchmark from the standardized feature value of each mode within the current preset diagnostic time window to obtain the vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, pressure phase residual, air volume phase residual, power phase residual and lubricating oil state phase residual. Then, arrange them according to the acquisition time, fan type label and component source to generate the working condition phase residual sequence.

[0109] In this embodiment, the generation of modal credibility markers and conflict modal markers includes:

[0110] Read the phase residual sequence of the operating condition and establish a modal residual alignment table according to the same preset diagnostic time window, the same fan type label and the same component source;

[0111] Count the number of missing data, the length of consecutive missing data, and the number of valid residuals for each modality, and generate data integrity markers;

[0112] Historical residual fluctuation ranges are generated based on the residual distribution between historical operating data (after removing historical outliers) and the in-phase load benchmark. Short-time stability markers are then generated by calculating the variation amplitude of each modal phase residual based on the modal residual alignment table, where:

[0113] The historical residual fluctuation range is generated based on the residual distribution between historical operating data after removing historical outliers and the in-phase load benchmark, specifically as follows:

[0114] Read the historical operating data and the same-phase load benchmark after removing historical abnormal segments. Group them according to the same operating condition phase, the same component source and the same data mode. Subtract the standardized eigenvalue of the corresponding mode in the same-phase load benchmark from the standardized eigenvalue of the historical operating data to obtain the historical residual set. Sort the historical residual set in ascending order of value, take the 5th percentile as the lower bound of the historical residual and take the 95th percentile as the upper bound of the historical residual to generate the historical residual fluctuation range.

[0115] The short-time stability marker is generated as follows:

[0116] Read the phase residuals of the same mode from two adjacent preset diagnostic time windows in the modal residual alignment table, subtract the phase residual of the previous mode from the current modal phase residual and take the absolute value to obtain the change amplitude. When the change amplitude is within the historical residual fluctuation range, it is marked as stable. When the change amplitude exceeds the upper limit of the historical residual but does not exceed 1.5 times the upper limit of the historical residual, it is marked as a slight jump. When the change amplitude exceeds 1.5 times the upper limit of the historical residual, it is marked as a short-term jump, and a short-term stability label is generated.

[0117] A trend consistency marker is generated based on the increase / decrease direction of the phase residuals of each mode within the same preset diagnostic time window. A response consistency marker for adjacent components is generated based on the order of residual occurrence and the direction of residual increase / decrease of adjacent components within consecutive preset diagnostic time windows. Where:

[0118] A trend consistency marker is generated based on the increase or decrease direction of the phase residuals of each mode within the same preset diagnostic time window, specifically:

[0119] Read the five types of phase residuals under the same preset diagnostic time window and the same component source. Subtract the phase residual of the previous preset diagnostic time window from the phase residual of the current preset diagnostic time window. If the difference is greater than 0.02, it is recorded as the upward direction. If the difference is less than -0.02, it is recorded as the downward direction. If the difference is greater than or equal to -0.02 and less than or equal to 0.02, it is recorded as the stationary direction. Count the direction with the most numbers among the five types of modes as the main trend direction. Divide the number of modes consistent with the main trend direction by the number of effective residuals to obtain the trend consistency ratio. When the trend consistency ratio is not less than 0.7, a consistent trend mark is generated. When the trend consistency ratio is less than 0.7, a conflict trend mark is generated.

[0120] Based on the order of residual occurrence and direction of residual increase / decrease of adjacent components within a continuous preset diagnostic time window, a consistency flag for the response of adjacent components is generated, specifically:

[0121] Read the mechanical transmission relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship in the connection relationship of the wind turbine components. Read the adjacent component source corresponding to the current component source. Compare the phase residual occurrence order of the current component source and the adjacent component source within 3 consecutive preset diagnostic time windows. If the adjacent component source shows the same direction residual change within 2 preset diagnostic time windows after the current component source, it is recorded as having an adjacent response. If the adjacent component source does not show the same direction residual change within 2 preset diagnostic time windows after the current component source, it is recorded as having no adjacent response. Generate an adjacent component response consistency mark.

[0122] Modal credibility labels are generated based on the combination of data integrity labels, short-term stability labels, trend consistency labels, and adjacent component response consistency labels. Modalities that are missing continuity, have short-term jumps, have opposite trend directions, and have no corresponding responses from adjacent components are labeled as conflicting modes, and conflicting mode labels are generated.

[0123] In this embodiment, generating conflict mode markers includes:

[0124] An integrity score is generated based on the ratio of the number of valid residuals to the number of expected residuals for each modality, and an integrity deduction is generated based on the length of consecutive missing values, where:

[0125] An integrity score is generated based on the ratio between the number of effective residuals and the number of expected residuals for each mode.

[0126] Read the residual sequences corresponding to the same component source and the same mode from the modal residual alignment table. Combine the current preset diagnostic time window and the previous four consecutive preset diagnostic time windows to form a confidence statistics window. The confidence statistics window contains five preset diagnostic time windows. Record the number of residuals that should appear in the confidence statistics window as the number of residuals that should be received. The number of residuals that should be received is 5. Records with non-empty residual fields and finite residual values ​​as valid residual records. Count the number of valid residual records to obtain the number of valid residuals. Divide the number of valid residuals by the number of residuals that should be received to obtain the integrity ratio. Use the integrity ratio as the initial integrity score.

[0127] Integrity deduction is generated based on the length of consecutive missing items, specifically as follows:

[0128] Five residual records are arranged from earliest to latest according to the collection time in the confidence statistics window. Records with empty residual fields or residual values ​​that are not finite values ​​are written as missing marker 1. Records with non-empty residual fields and finite residual values ​​are written as missing marker 0. The longest consecutive length of missing marker 1 is calculated to obtain the consecutive missing length. When the consecutive missing length is 0, the integrity deduction is 0. When the consecutive missing length is 1, the integrity deduction is 0.1. When the consecutive missing length is 2, the integrity deduction is 0.3. When the consecutive missing length is greater than or equal to 3, the integrity deduction is 0.6. The integrity score is obtained by subtracting the integrity deduction from the initial integrity score and limiting it to the range of 0 to 1.

[0129] A stability score is generated based on the comparison between the variation amplitude of the phase residuals of each mode and the historical residual fluctuation range. Specifically, the stability score is generated based on the comparison between the variation amplitude of the phase residuals of each mode and the historical residual fluctuation range.

[0130] Read the residual values ​​of the same mode in the confidence statistics window arranged by acquisition time. Subtract the residual values ​​of two adjacent preset diagnostic time windows and take the absolute value to obtain the change in adjacent residuals. Take the maximum value of all adjacent residual changes in the confidence statistics window as the current phase residual change amplitude. Read the historical residual sequence of the same working condition phase, the same preset power range, the same preset speed range, the same component source, and the same mode in the historical operating data after removing historical abnormal segments. Calculate the change in historical adjacent residuals according to 5 consecutive preset diagnostic time windows as a group. Arrange the historical adjacent residual changes in ascending order of value and read the 5th percentile and 95th percentile as the historical residual fluctuation range.

[0131] The stability score is calculated as follows: when the current phase residual change is less than or equal to the 95th percentile of the historical residual fluctuation range, the stability score is 1; when the current phase residual change is greater than the 95th percentile but less than or equal to 1.5 times the 95th percentile, the stability score is 0.5; and when the current phase residual change is greater than 1.5 times the 95th percentile, the stability score is 0.

[0132] A trend score is generated based on the number of times the phase residuals of each mode increase or decrease in the same direction within the same preset diagnostic time window. An adjacency response score is generated based on the number of times adjacent components exhibit residual responses in the same direction within consecutive preset diagnostic time windows.

[0133] A trend score is generated based on the number of modal phase residuals that increase or decrease in the same direction within the same preset diagnostic time window. Specifically:

[0134] Read the vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, and power phase residual within the same preset diagnostic time window. Subtract the modal residual of the previous preset diagnostic time window from the modal residual of the current preset diagnostic time window to obtain the modal residual difference. When the modal residual difference is greater than 0.02, write it as rising direction 1. When the modal residual difference is less than -0.02, write it as falling direction -1. When the modal residual difference is between -0.02 and 0.02, write it as stationary direction 0. Count the direction with the larger number of rising direction 1 and falling direction -1 as the main trend direction. Record the number of modes with the same main trend direction as the number of modes with the same increasing or decreasing direction. Divide the number of modes with the same increasing or decreasing direction by the total number of modes with rising direction 1 and falling direction -1 to obtain the trend score. When all modes are stationary direction 0, the trend score is 1.

[0135] Adjacency response scores are generated based on the number of times adjacent components exhibit residual responses in the same direction within a consecutive preset diagnostic time window. Specifically:

[0136] Read the adjacent component nodes in the component topology diagram that are connected to the current component node through mechanical transmission, vibration transmission, heat conduction, and electrical coupling. Read the main trend direction and absolute value of residuals of the current component node within the current preset diagnostic time window. When the absolute value of residuals of the current component node is greater than the 75th percentile of the residuals of the same phase and same modality of the same component in the historical data, start the adjacency response statistics. Read the difference of residuals of the same mode of adjacent component nodes within the current preset diagnostic time window and the next two consecutive preset diagnostic time windows. If the direction of the difference of residuals of the same mode of adjacent component nodes is the same as the main trend direction of the current component node and the absolute value of residuals of adjacent component nodes is greater than the 75th percentile of the residuals of the same phase and same modality of the same component in the historical data, it is recorded as 1 residual response in the same direction. Divide the number of residual responses in the same direction by the number of adjacent component nodes to obtain the adjacency response score. When the adjacency response score is greater than 1, it is taken as 1.

[0137] Modal confidence scores are generated based on integrity score, integrity deduction, stability score, trend score, and adjacency response score. Modal confidence labels are then generated according to the interval to which the modal confidence score belongs, where:

[0138] Modal credibility scores are generated based on integrity score, integrity deduction, stability score, trend score, and adjacency response score, specifically as follows:

[0139] Read the integrity score, stability score, trend score, and adjacency response score. Multiply the integrity score by 0.3 to get the integrity contribution value, multiply the stability score by 0.25 to get the stability contribution value, multiply the trend score by 0.25 to get the trend contribution value, and multiply the adjacency response score by 0.2 to get the adjacency response contribution value. Add the integrity contribution value, stability contribution value, trend contribution value, and adjacency response contribution value to get the modal reliability score. A weight of 0.3 corresponds to whether the data is usable, a weight of 0.25 corresponds to whether short-term jumps exist, a weight of 0.25 corresponds to whether the multimodal change direction is consistent, and a weight of 0.2 corresponds to whether there is a synchronous response in the connection relationship of wind turbine components.

[0140] Modal confidence labels are generated based on the interval to which the modal confidence score belongs, specifically as follows:

[0141] Modes with a modal confidence score greater than or equal to 0.75 are written into the trusted input tag; modes with a modal confidence score greater than or equal to 0.5 and less than 0.75 are written into the input tag to be verified; modes with a modal confidence score less than 0.5 are written into the suppression input tag; and the trusted input tag, the input tag to be verified, and the suppression input tag are arranged in the order of vibration, acoustics, temperature, current, and power to generate the modal confidence tag.

[0142] Continuously missing modes, short-term abrupt change modes, and modes with opposite trends and no corresponding response from adjacent components are written into conflict mode tags.

[0143] In this embodiment, generating the component topology fault propagation residual map includes:

[0144] Read the connection relationships of the fan components, and set the fan body, rotating bearing, drive motor, inlet duct, outlet duct, regulating mechanism, lubrication system, and foundation support as component nodes. Write a component node marker for each node, and set mechanical connection relationships, aerodynamic flow relationships, vibration transmission relationships, heat conduction relationships, and electrical coupling relationships as connection edges to generate a component topology diagram, where:

[0145] Read the connection relationships of the fan components, specifically:

[0146] Read the component name field, installation location field, mechanical connection field, aerodynamic flow field, vibration transmission field, heat dissipation contact field, and cable connection field from the fan structure configuration data. Limit the component names to fan body, rotating bearing, drive motor, inlet duct, outlet duct, regulating mechanism, lubrication system, and foundation support. Assign a unique component node label to each component name: fan body is numbered N1, rotating bearing is numbered N2, drive motor is numbered N3, inlet duct is numbered N4, outlet duct is numbered N5, regulating mechanism is numbered N6, lubrication system is numbered N7, and foundation support is numbered N8.

[0147] Define the connecting edges, specifically:

[0148] Establish a mechanical connection relationship between two components in the mechanical connection field that have a main shaft connection, coupling connection, bearing support connection, or rigid assembly connection; establish a pneumatic flow relationship between two components in the pneumatic flow field that have a medium inflow, medium outflow, air duct connection, or pressure transmission path; establish a vibration transmission relationship between two components in the vibration transmission field that have a structural contact, common support, or rigid installation relationship; establish a heat conduction relationship between two components in the heat dissipation contact field that have a heat conduction path or lubricating oil heat exchange path; and establish an electrical coupling relationship between two components in the cable connection field that have an electrical power transmission, drive control, or status signal connection.

[0149] Generate a component topology diagram, specifically as follows:

[0150] Establish an 8x8 adjacency matrix, with row and column numbers corresponding to N1 to N8. Write the positions with mechanical connections into M, the positions with pneumatic flow relationships into A, the positions with vibration transmission relationships into V, the positions with heat conduction relationships into T, the positions with electrical coupling relationships into E, and the positions without connections into 0. When there are more than two types of connections at the same matrix position, write the corresponding connection edge types into the connection edge type set in the order of M, A, V, T, E. Generate the component topology diagram based on the adjacency matrix and the connection edge type set.

[0151] The phase residual sequence of operating conditions is mapped to component nodes according to the component source. Modal confidence tags and conflict mode tags are mapped to component nodes and connecting edges to generate component residual topology data, where:

[0152] Mapping the phase residual sequence of operating conditions to component nodes according to component origin is as follows:

[0153] Read the preset diagnostic time window number, window center time, fan type mark, component source, operating condition phase, vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, pressure phase residual, air volume phase residual, power phase residual, and lubricating oil status phase residual from the operating condition phase residual sequence. Match the component source to the corresponding component node mark in N1 to N8. Write the 8 types of phase residuals into the node residual vector of the corresponding component node in the order of vibration, acoustic, temperature, current, pressure, air volume, power, and lubricating oil status.

[0154] Modal confidence tags and conflict mode tags are mapped to component nodes and connection edges, specifically as follows:

[0155] Read the component source, data mode, and confidence flag from the modality confidence flag, write the trusted input flag to 1, write the input flag to be verified to 0.5, write the suppressed input flag to 0, and write the value to the data mode position of the corresponding component node. Read the component source, data mode, and conflict type from the conflict mode flag, write the conflict mode to the corresponding component node, and synchronously write the conflict modes involving the component nodes at both ends of the same connection edge to the corresponding connection edge.

[0156] Generate component residual topology data, specifically as follows:

[0157] Using the preset diagnostic time window number and component node tag as the primary key, the node residual vector, confidence value, conflict type, connection edge type and adjacency matrix position are written into the same topology record to generate component residual topology data.

[0158] Based on the phase residual difference, residual increase / decrease direction, and modal confidence marker of the component nodes at both ends of the connecting edge within the same preset diagnostic time window, the adjacent response residual is calculated, where:

[0159] Read the component nodes at both ends of the connecting edge, specifically:

[0160] Read the non-zero adjacency matrix positions in the component topology diagram, take the component node corresponding to the row number as the upstream component node, take the component node corresponding to the column number as the downstream component node, and read the node residual vectors of the upstream and downstream component nodes within the same preset diagnostic time window.

[0161] The phase residual difference is calculated as follows:

[0162] Calculate the vibration phase residual difference, acoustic phase residual difference, temperature phase residual difference, current phase residual difference, pressure phase residual difference, air volume phase residual difference, power phase residual difference, and lubricating oil state phase residual difference in the upstream and downstream component nodes respectively. Each difference is the absolute value of the downstream component node phase residual minus the upstream component node phase residual.

[0163] The direction of increase or decrease in residuals is calculated as follows:

[0164] Read the node residual vectors of the component nodes at both ends of the connection edge within the current preset diagnostic time window and the previous preset diagnostic time window. Subtract the phase residual of the previous preset diagnostic time window from the phase residual of the current preset diagnostic time window. If the difference is greater than 0.02, it is recorded as the upward direction. If the difference is less than -0.02, it is recorded as the downward direction. If the difference is between -0.02 and 0.02, it is recorded as the stationary direction. Write the residual increase / decrease direction of the corresponding modes of the upstream component node and the downstream component node into the connection edge direction record.

[0165] The adjacent response residuals are calculated as follows:

[0166] Read the modal confidence markers of the component nodes at both ends of the connection edge, retain the difference values ​​of the corresponding confident input markers among the 8 types of phase residual difference values, multiply the difference value of the corresponding input marker to be verified by 0.5, write the difference value of the corresponding suppression input marker to 0, write the modes with the same direction of increase or decrease of residuals of upstream component nodes and downstream component nodes into the same direction response marker, write the modes with opposite directions of increase or decrease of residuals into the opposite direction response marker, sum the processed 8 types of difference values ​​and divide by the number of effective modes to obtain the adjacent response residual;

[0167] Starting from the component node with the highest residual magnitude, backtracking along the connecting edges, the reverse source residual is calculated based on the order of residual occurrence and the difference in residual intensity between upstream and downstream component nodes, where:

[0168] Reverse backtracking, specifically:

[0169] Starting from the reverse backtracking point, upstream component nodes are read in reverse along the connection edges in the component topology diagram. Mechanical connection relationships are read in reverse according to the power input direction and support transmission direction, pneumatic flow relationships are read in reverse according to the medium flow direction, vibration transmission relationships are read in reverse according to the connection direction of adjacent structures, heat conduction relationships are read in reverse according to the direction of proximity to the heat source, and electrical coupling relationships are read in reverse according to the direction of power input and control signal input. The maximum backtracking depth is set to 3 connection edges to obtain the candidate source component node set.

[0170] The reverse source residual is calculated as follows:

[0171] Read the node residual amplitude of the candidate source component node and the reverse backtracking starting point within three consecutive preset diagnostic time windows. The node residual amplitude is the arithmetic mean of the absolute values ​​of the eight types of phase residuals corresponding to the trusted input marker. If the residual of the candidate source component node appears earlier than the reverse backtracking starting point, and the residual amplitude of the candidate source component node is not less than 0.6 of the residual amplitude of the reverse backtracking starting point, then the candidate source component node is marked as the source candidate node. The absolute value of the residual amplitude of the reverse backtracking starting point minus the residual amplitude of the source candidate node is taken as the reverse source residual.

[0172] Based on the order of residual occurrence, direction of residual increase / decrease, and edge type of multiple component nodes within a continuous preset diagnostic time window, cross-node consistency residuals are calculated. These residuals are then combined with adjacency response residuals, reverse source residuals, cross-node consistency residuals, modal confidence markers, and conflict mode markers to generate a component topology fault propagation residual map, where:

[0173] The cross-node consistency residual is calculated as follows:

[0174] Read the node residual amplitude and residual increase / decrease direction of each component node within three consecutive preset diagnostic time windows. Subtract the node residual amplitude of the previous preset diagnostic time window from the node residual amplitude of the current preset diagnostic time window. A difference greater than 0.02 is recorded as an upward direction, a difference less than -0.02 is recorded as a downward direction, and a difference between -0.02 and 0.02 is recorded as a stationary direction. Read the order of occurrence of residuals of upstream and downstream component nodes along the same connection edge. If the upstream component node first shows an upward direction and the downstream component node shows the same direction of change within two preset diagnostic time windows, write a consistent propagation mark. Otherwise, write a non-consistent propagation mark. The difference in residual amplitude between the two nodes of the connection edge corresponding to the non-consistent propagation mark is taken as the cross-node consistent residual.

[0175] The combination of adjacency response residuals, reverse source residuals, cross-node consistency residuals, modal credibility tags, and conflict modal tags is as follows:

[0176] Read the adjacent response residuals, reverse source residuals, cross-node consistency residuals, modal confidence markers and conflict mode markers according to the preset diagnostic time window number, component node markers and connection edge types, and write the data corresponding to the same component node in the same preset diagnostic time window into the same graph record.

[0177] Generate the component topology fault propagation residual map, specifically as follows:

[0178] Write the component node label, connection edge type, node residual vector, adjacency response residual, reverse source residual, cross-node consistency residual, modal confidence label, conflict modal label, source candidate node and consistency propagation label into the graph matrix, arrange the graph matrix in ascending order according to the preset diagnostic time window number, and generate the component topology fault propagation residual graph.

[0179] In this embodiment, the generation of wind turbine fault state characterization includes:

[0180] An improved liquid time constant network is constructed, comprising a reliable mode selection layer, an impact energy amplification layer, a polarization complex liquid evolution layer, a phase coupling state propagation layer, and a conflict mode bypass verification layer. Specifically, the construction of the improved liquid time constant network is as follows:

[0181] This invention constructs an improved liquid time constant network based on the original backbone structure of the liquid time constant network, which includes input mapping, liquid neuron state update, time constant decay, recursive connection, and state readout. It uses the operating condition phase residual sequence as the time-series input, modal confidence labels and conflict mode labels as input selection criteria, and component topology fault propagation residual maps as structural propagation input. The improved liquid time constant network consists of a trusted mode selection layer, an impact energy amplification layer, a polarized complex liquid evolution layer, a phase coupling state propagation layer, a conflict mode bypass verification layer, and a fault state characterization output layer connected sequentially. A trusted mode selection layer is set before input mapping to filter vibration phase residuals, acoustic phase residuals, temperature phase residuals, current phase residuals, pressure phase residuals, airflow phase residuals, power phase residuals, and lubricating oil state phase residuals based on modal confidence labels, generating a trusted mode input sequence. An impact energy amplification layer is set before the liquid neuron state update. The layer reads vibration phase residual burst segments, acoustic phase residual pulse segments, current phase residual period shift segments, pressure phase residual oscillation segments, and air volume phase residual fluctuation segments and performs hierarchical enhancement. In the liquid neuron state update structure, a polarized complex liquid evolution layer is set to write the amplitude residual into the real part of the complex liquid state, write the phase shift between adjacent preset diagnostic time windows into the imaginary part of the complex liquid state, and generate complex liquid evolution states along the liquid time constant update path. In the recursive connection structure, a phase coupling state propagation layer is set to propagate the complex liquid evolution states across components based on mechanical connection relationships, aerodynamic flow relationships, vibration transmission relationships, heat conduction relationships, and electrical coupling relationships. In the state readout bypass, a conflict mode bypass verification layer is set to read the conflict mode bypass sequence and generate conflict mode verification evidence, forming an improved liquid time constant network that includes credible input screening, impact residual enhancement, complex liquid evolution, topological phase propagation, and conflict bypass verification.

[0182] The operating condition phase residual sequence, modal confidence label, and conflict mode label are input into the confidence mode selection layer to extract the confidence residual data and conflict residual data, and generate the confidence mode input sequence and conflict mode bypass sequence.

[0183] The reliable modal input sequence is input into the impact energy amplification layer, and the vibration phase residual sudden increase segment, acoustic phase residual pulse segment, current phase residual period shift segment, pressure phase residual oscillation segment, and air volume phase residual fluctuation segment are extracted to generate the impact-enhanced liquid input sequence, wherein:

[0184] The impact energy amplification layer includes:

[0185] Trusted modal input buffer: stores the trusted modal input sequence and the phase residuals of each mode;

[0186] Vibration Sudden Segment Reader: Stores vibration phase residual difference values, historical vibration residual thresholds, and vibration sudden segment markers;

[0187] Acoustic pulse segment reader: stores acoustic phase residual peak value, mean value, historical acoustic residual threshold, and acoustic pulse segment markers;

[0188] Current cycle offset segment reader: stores the number of current phase residual sign changes, historical current residual thresholds, and current cycle offset segment markers;

[0189] Pressure oscillation segment reader: stores pressure phase residual oscillation amplitude, historical pressure residual threshold, and pressure oscillation segment markers;

[0190] Airflow fluctuation segment reader: stores airflow phase residual difference, historical airflow residual threshold, and airflow fluctuation segment markers;

[0191] Impact Fragment Summarizer: Stores various impact fragment tags and generates impact fragment summary tags;

[0192] Graded amplifier: Stores residual amplification factor and amplifies the phase residual of the triggered impact segment mark;

[0193] Shock-enhanced sequence output unit: stores the amplified phase residual and the unamplified phase residual, and generates the shock-enhanced liquid input sequence;

[0194] Extracting the segment with a sudden increase in vibration phase residual, specifically:

[0195] Read the vibration phase residual in the reliable modal input sequence, calculate the difference between the vibration phase residual of the current preset diagnostic time window and the vibration phase residual of the previous preset diagnostic time window. When the difference is greater than 0.08 and the current vibration phase residual is greater than the 90th percentile of the historical vibration residual under the same working condition phase, mark the corresponding preset diagnostic time window as a vibration phase residual burst segment.

[0196] Extracting acoustic phase residual pulse segments, specifically:

[0197] Read the acoustic phase residual in the trusted modal input sequence, calculate the peak value and mean value of the acoustic phase residual within three consecutive preset diagnostic time windows. When the peak value is greater than 1.5 times the mean value and the peak value is greater than the 90th percentile of the historical acoustic residual under the same working condition phase, mark the preset diagnostic time window where the peak value is located as the acoustic phase residual pulse segment.

[0198] Extracting the current phase residual period offset segment, specifically:

[0199] Read the current phase residual in the trusted modal input sequence, calculate the number of sign changes of adjacent difference values ​​of the current phase residual within 5 consecutive preset diagnostic time windows. When the number of sign changes is not less than 3 and the average absolute value of the current phase residual is greater than the 85th percentile of the historical current residual under the same operating condition phase, mark the corresponding 5 consecutive preset diagnostic time windows as current phase residual period offset segments.

[0200] Extract the pressure phase residual oscillation segment, specifically:

[0201] Read the pressure phase residual in the trusted modal input sequence, calculate the number of sign changes of adjacent differences of the pressure phase residual within 5 consecutive preset diagnostic time windows, and calculate the difference between the maximum and minimum values ​​of the pressure phase residual within 5 consecutive preset diagnostic time windows. When the number of sign changes is not less than 3, the difference between the maximum and minimum values ​​is greater than 0.06, and the average absolute value of the pressure phase residual is greater than the 85th percentile of the historical pressure residual under the same working condition phase, mark the corresponding 5 consecutive preset diagnostic time windows as pressure phase residual oscillation segments.

[0202] Extracting the airflow phase residual fluctuation segment, specifically:

[0203] Read the air volume phase residual in the trusted modal input sequence, calculate the difference between the air volume phase residual of the current preset diagnostic time window and the air volume phase residual of the previous preset diagnostic time window. When the absolute value of the difference is greater than 0.07 and the current air volume phase residual is greater than the 90th percentile of the historical air volume residual under the same operating condition phase, mark the corresponding preset diagnostic time window as an air volume phase residual fluctuation segment.

[0204] The shock-enhanced liquid input sequence is as follows:

[0205] Multiply the phase residuals corresponding to the vibration phase residual sudden increase segment, acoustic phase residual pulse segment, current phase residual period shift segment, pressure phase residual oscillation segment, and air volume phase residual fluctuation segment by 1.5, retain the original value of the phase residuals at non-segment positions, and arrange the enhanced vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, pressure phase residual, air volume phase residual, power phase residual, and lubricating oil state phase residual in ascending order according to the preset diagnostic time window number to generate the impact-enhanced liquid input sequence;

[0206] A phase shift is generated based on the phase residual difference between consecutive preset diagnostic time windows in the impact-enhanced liquid input sequence. The amplitude residual is written into the real part of the complex liquid state, and the phase shift is written into the imaginary part of the complex liquid state. Continuous state evolution is performed along the liquid time constant update path to generate a complex liquid evolution state, where:

[0207] Phase shift is generated based on the phase residual difference between consecutive preset diagnostic time windows in the impact-enhanced liquid input sequence, specifically as follows:

[0208] Read the phase residual of the same mode in two consecutive preset diagnostic time windows in the impact-enhanced liquid input sequence, subtract the phase residual of the previous preset diagnostic time window from the phase residual of the current preset diagnostic time window to obtain the single-mode phase residual difference, arrange the five types of single-mode phase residual differences in modal order to generate a phase offset vector;

[0209] The amplitude residual is written into the real part of the complex liquid state, and the phase shift is written into the imaginary part of the complex liquid state, specifically as follows:

[0210] Read the impact-enhanced liquid input sequence of the current preset diagnostic time window, map the five types of amplitude residuals into a 64-dimensional amplitude driving vector through a 64-dimensional linear mapping matrix, and write it into the real part of the complex liquid state. Map the phase offset vector into a 64-dimensional phase driving vector through a 64-dimensional linear mapping matrix, and write it into the imaginary part of the complex liquid state.

[0211] Continuous state evolution, specifically:

[0212] The preset diagnostic time window length is set to 10s, the liquid time constant is set to 30s, and the ratio of the calculated window length to the liquid time constant is 0.333. The real part of the previous preset diagnostic time window is multiplied by 0.667, the current 64-dimensional amplitude driving vector is multiplied by 0.333, and the two products are added together to obtain the current real part state. The imaginary part of the previous preset diagnostic time window is multiplied by 0.667, the current 64-dimensional phase driving vector is multiplied by 0.333, and the two products are added together to obtain the current imaginary part state. The current real part state and the current imaginary part state are combined to generate the complex liquid evolution state.

[0213] The complex liquid evolution state and the component topology fault propagation residual map are input into the phase-coupled state propagation layer. Component phase-coupled propagation is performed along mechanical connections, aerodynamic flow relationships, vibration transmission relationships, heat conduction relationships, and electrical coupling relationships to generate a topology-coupled liquid state, where:

[0214] The phase-coupled state propagation layer includes:

[0215] Node state mapper: Reads complex liquid evolution states and component node markers, and maps complex liquid evolution states to corresponding component nodes;

[0216] Connection edge reader: Reads mechanical connection relationships, pneumatic flow relationships, vibration transmission relationships, heat conduction relationships, and electrical coupling relationships to determine the propagation path between component nodes;

[0217] Propagation coefficient configurator: Reads the connection edge type and configures the corresponding propagation coefficient for different connection edges;

[0218] Topology residual reader: Reads adjacency response residuals, reverse source residuals, and cross-node consistency residuals to generate propagation correction basis;

[0219] Connecting Edge Propagation Calculator: Calculates the propagation state of connecting edges based on the complex liquid state of nodes, propagation coefficients, and propagation correction criteria;

[0220] Node propagation aggregator: Aggregates the propagation states of connected edges received by nodes of the same component and generates node propagation increments;

[0221] Topology Coupled State Outputter: Merges the node propagation increment with the original node complex liquid state to generate a topology coupled liquid state;

[0222] The topologically coupled liquid state is generated as follows:

[0223] Read the component node markers, connection edge types, adjacency response residuals, reverse source residuals, and cross-node consistency residuals from the component topology fault propagation residual map, and map the complex liquid evolution states to the corresponding component nodes of the fan body, rotating bearing, drive motor, inlet duct, outlet duct, regulating mechanism, lubrication system, and foundation support according to the component source, and generate the node complex liquid states;

[0224] The propagation coefficients for mechanical connection relationships are set to 1, pneumatic flow relationships to 0.9, vibration transmission relationships to 0.8, electrical coupling relationships to 0.7, and heat conduction relationships to 0.5. These coefficients are set based on the following: mechanical connection relationships correspond to the direct structural transmission path between the rotor, bearings, couplings, and foundation support; pneumatic flow relationships correspond to the pressure and airflow transmission path between the inlet duct, fan body, and outlet duct; vibration transmission relationships correspond to the vibration diffusion path generated by rigid installation and structural contact; electrical coupling relationships correspond to the abnormal linkage path of drive motor current, power, and control signals; and heat conduction relationships correspond to the temperature change transmission path between bearings, motors, lubrication systems, and ducts.

[0225] Read the upstream and downstream component nodes of each connection edge, multiply the complex liquid state of the upstream component node by the corresponding connection edge propagation coefficient, multiply by the adjacency response residual, multiply by the normalized value of the negative of the cross-node consistency residual to obtain the connection edge propagation state, sum all connection edge propagation states received by the same component node to obtain the node propagation increment, add the node propagation increment to the original node complex liquid state to generate the topology coupling liquid state;

[0226] The conflict mode bypass sequence is input into the conflict mode bypass verification layer to generate conflict mode verification evidence. Based on the topologically coupled liquid state and the conflict mode verification evidence, a wind turbine fault state characterization is generated, wherein:

[0227] The conflict mode bypass check layer includes:

[0228] Conflict sequence reader: Reads conflict mode bypass sequences and conflict residual data, and filters the data modes that need to be verified;

[0229] Topology response reader: Reads the topologically coupled liquid state and generates the topology response value of the corresponding mode;

[0230] Orientation Comparator: Compares conflicting residual directions with topological response directions to generate supporting evidence, counter-evidence, and neutral evidence;

[0231] Evidence counter: Counts the number of supporting evidence, the number of disproving evidence, and the number of neutral evidence, and generates an evidence quantity vector;

[0232] Bypass verifier: Generates conflict modality verification evidence based on evidence quantity vector;

[0233] Fault characterization splicer: splices together topologically coupled liquid state and conflict mode verification evidence to generate a fault state characterization of the wind turbine;

[0234] The generation of conflict modality verification evidence is as follows:

[0235] Read the conflict residual data of each preset diagnostic time window in the conflict modal bypass sequence, read the topological coupling liquid state of the same preset diagnostic time window, calculate whether the direction of the non-zero position in the conflict residual data is consistent with the modal position corresponding to the topological coupling liquid state, if the direction is consistent, it is recorded as supporting evidence, if the direction is opposite, it is recorded as contradictory evidence, and if the direction is stable, it is recorded as neutral evidence. Count the number of supporting evidence, the number of contradictory evidence and the number of neutral evidence, and generate conflict modal verification evidence.

[0236] The following are the specific steps for generating a fault state representation for the wind turbine:

[0237] Read the 64-dimensional real part state and 64-dimensional imaginary part state of the topologically coupled liquid state, calculate the square root of the sum of the squares of the real part state and the imaginary part state in each dimension to obtain the 64-dimensional amplitude-phase joint state, read the number of supporting evidence, the number of contradictory evidence and the number of neutral evidence in the conflict mode verification evidence, and splice the 64-dimensional amplitude-phase joint state with the 3-dimensional conflict mode verification evidence to generate a 67-dimensional wind turbine fault state characterization.

[0238] In this embodiment, the fault diagnosis results of the output fan include:

[0239] Read the topologically coupled liquid state, conflict mode verification evidence, component node markers, and preset diagnostic time window markers from the wind turbine fault state characterization;

[0240] Perform fault category decoding on the topologically coupled liquid state to generate fault categories;

[0241] The location of the faulty component is generated based on the component node markers and the residual concentration locations in the component topology fault propagation residual map.

[0242] Fault levels are classified based on the intensity of anomalies in the fault state characterization of wind turbines and the number of continuously preset diagnostic time windows.

[0243] Based on modal confidence markers, conflict modal verification evidence, adjacent response residuals, reverse source residuals, and cross-node consistency residuals, diagnostic confidence and diagnostic evidence chains are generated. Then, the fault category, fault component location, fault level, diagnostic confidence and diagnostic evidence chains are combined to output the wind turbine fault diagnosis results.

[0244] Example 1: During a continuous auxiliary operation cycle of a thermal power unit, the diagnostic targets are two forced draft fans, two induced draft fans, and two primary air fans within the same unit. Vibration, acoustic, temperature, current, voltage, power, speed, pressure, airflow, regulating opening, and lubricating oil status measurement points are set up for each fan. The system continuously receives 180 minutes of operating data, with a total of 399,600 raw records. Among these, 259,200 are high-frequency data points (vibration, acoustic, and current), and 140,400 are low-frequency data points (temperature, pressure, airflow, speed, power, regulating opening, lubricating oil status, and unit-related data). In this example, except for the acquisition time, diagnostic delay, and number of data points, the unit load, pressure, airflow, speed, power, regulating opening, and various phase residuals are all dimensionless values ​​standardized from 0 to 1.

[0245] During the operating cycle, the load of the standardized unit increased from 0.56 to 0.82 and then fluctuated. The operation status of the coal mill corresponding to one primary air fan changed from three to four units. The outlet pressure of one induced draft fan experienced short-term oscillations, and the measuring point of the rotating support of one forced draft fan showed slight sampling drift. Traditional fixed threshold methods would misjudge the load increase and coal mill operation disturbances as faults. The method of this invention reads the acquisition time, sampling frequency, fan type label, component source, and original feature values, and performs abnormal data removal and completion. The original data contained 426 duplicate records, 91 records in reverse time order, and 37 records with non-finite values. After processing, the effective record retention rate was 99.86%, and the continuous window coverage rate was improved from 96.1% to 99.6%.

[0246] The system establishes a unified diagnostic time axis with a 10-second window length and a 5-second sliding step, forming 2159 preset diagnostic time windows. Vibration, acoustic, and current data are labeled as high-frequency modal data, and the peak value, peak-to-peak value, mean, variance, standard deviation, and number of pulses are calculated within each window. The remaining data are labeled as low-frequency modal data and mapped to the diagnostic time window closest to the center time of the window. Before correction, the average time drift of each mode was 1.86 seconds, and the maximum drift was 4.7 seconds. After correction, the average time drift decreased to 0.42 seconds, and the number of timing mismatch windows decreased from 326 to 49.

[0247] The system classifies operating phases based on differences in unit load, speed, power, pressure, air volume, regulating opening, and changes in the coal mill's operational status. In all windows, the start-up acceleration phase accounts for 8.4%, the load ramp-up phase for 16.7%, the stable ventilation phase for 51.2%, the air volume regulation phase for 13.5%, the coal mill start-up / shutdown disturbance phase for 6.1%, and the shutdown deceleration phase for 4.1%. The system removed 742 historical abnormal segments from historical operating data, retaining 18,426 historical normal segments. Historical data was then filtered according to the same fan type, operating phase, load range, speed range, air volume range, regulating opening range, coal mill operational status, and component origin, generating a load benchmark for the same phase based on the median.

[0248] During the switching window for the operation of the primary ventilation fan and coal mill, the pressure phase residual in the inlet duct was 0.11, and the air volume phase residual was 0.09, both falling within the allowable fluctuation range of the load benchmark for the same phase. At the same time, the traditional fixed threshold method triggered an alarm because the primary air pressure exceeded the threshold of 0.78. The method of this invention marked this segment as a coal mill operation / disruption disturbance phase and did not output a fault alarm. In another induced draft fan, the outlet duct showed pressure phase residuals of 0.18, 0.22, 0.25, 0.21, 0.24, and 0.23 for six consecutive windows within the stable ventilation phase, while the air volume phase residual simultaneously increased from 0.08 to 0.19. The adjustment opening change did not exceed 0.01, and the system determined that this anomaly did not belong to normal air volume adjustment.

[0249] In the multimodal reliability self-verification, the vibration mode of the blower rotating support, due to a continuous missing length of 2, has an integrity score of 0.58 and is written into the input flag to be verified; the pressure mode of the induced draft fan outlet duct has an integrity score of 1, a stability score of 0.5, a trend score of 1, an adjacency response score of 0.83, and a modal reliability score of 0.84, and is written into the reliable input flag. The system generates a component topology fault propagation residual map for 8 component nodes and 5 types of connection edges. Within the abnormal window of the induced draft fan outlet duct, the residual amplitude of the outlet duct node is 0.23, the residual amplitude of the blower body node is 0.16, the residual amplitude of the regulating mechanism node is 0.05, the adjacency response residual of the pneumatic flow connection edge is 0.19, the cross-node consistency residual is 0.04, and the reverse source residual points to the outlet duct node.

[0250] An improved liquid time constant network receives residual inputs from 12 consecutive diagnostic windows, with an input dimension of 12×8×8. The component topology fault propagation residual map input dimension is 12×8×8×5. An impact energy amplification layer enhances the pressure phase residual oscillation segment and the airflow phase residual fluctuation segment; the average residual response before enhancement is 0.21, and after enhancement it is 0.32. A phase coupling state propagation layer propagates the state along the aerodynamic flow relationship and vibration transmission relationship. A conflict mode bypass verification layer generates 5 supporting pieces of evidence, 1 piece of counter-evidence, and 2 pieces of neutral evidence. The final output fault category is duct blockage, the faulty component location is the outlet duct, the fault level is level 2, and the diagnostic confidence level is 0.91.

[0251] The training phase used historical operating samples of similar wind turbines to construct a dataset, containing 18,600 training samples, 4,200 validation samples, and 4,200 test samples, for a total of 27,000 samples. Each sample contained 12 consecutive diagnostic windows. Sample labels were derived from maintenance record index data, including 20,000 fault samples and 7,000 normal samples. On the same batch of test data, the fault category identification accuracy of the traditional fixed threshold alarm combined with the ordinary LTC method was 81.6%, while that of this invention was 92.4%; the fault component location accuracy improved from 74.3% to 89.1%; the fault level determination accuracy improved from 78.2% to 88.7%; the weak fault identification rate improved from 61.5% to 83.6%; the false alarm rate under varying operating conditions decreased from 18.9% to 6.8%; the number of false alarms due to coal mill start-up / shutdown disturbances decreased from 47 to 11; the average diagnostic delay decreased from 18.5s to 9.7s; and the diagnostic evidence chain completeness rate improved from 58.4% to 86.2%. As can be seen from Example 1, the present invention can reduce the interference of operating condition changes and improve the accuracy of weak fault identification, fault component location, fault level determination and diagnostic evidence chain generation in operating scenarios where load changes, air volume regulation, pressure fluctuations, coal mill start-up and shutdown disturbances and sensor time drift coexist.

[0252] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A wind turbine fault diagnosis method based on multimodal data fusion, characterized in that, include: Collect multimodal operation data, historical operation data, maintenance record data, and wind turbine structural configuration data. Standardize the multimodal operation data and historical operation data to generate a standardized multimodal status dataset. Generate maintenance record index data based on maintenance record data and generate wind turbine component connection relationships based on wind turbine structural configuration data. Asynchronous time drift calibration is performed based on a standardized multimodal state dataset to generate a time-aligned multimodal dataset; Perform phase-consistent slicing of the time-aligned multimodal dataset, remove historical abnormal segments based on maintenance record index data and generate in-phase load benchmarks, calculate the phase residuals of each mode based on the in-phase load benchmarks, and generate a phase residual sequence of the operating conditions. Perform multimodal confidence self-verification and conflict stripping on the phase residual sequence of the operating condition to generate modal confidence tags and conflict mode tags; Based on the connection relationship of wind turbine components, a component topology graph is constructed. The operating condition phase residual sequence, modal confidence label and conflict mode label are mapped to the component topology graph. Adjacency response residual, reverse source residual and cross-node consistency residual are calculated to generate a component topology fault propagation residual map. An improved liquid time constant network is constructed. The phase residual sequence of operating conditions, modal confidence label, conflict mode label and component topology fault propagation residual map are input into the improved liquid time constant network. Impulse fault feature enhancement, complex liquid state evolution and component phase coupling propagation are performed to generate a wind turbine fault state characterization. Based on the characterization of wind turbine fault status, output wind turbine fault diagnosis results.

2. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The multimodal operation data of the fan includes fan type marking, vibration data, acoustic data, temperature data, current data, voltage data, power data, speed data, pressure data, air volume data, regulation opening data, lubricating oil status data, and unit-related operation data. The unit-related operation data includes unit load data and coal mill commissioning status data. The historical operating data includes historical data corresponding to the multimodal operating data of the wind turbine; The maintenance record data includes fault repair time, downtime alarm time, component replacement time, manual inspection abnormality markers, and repair component markers; The fan structure configuration data includes the mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship between the fan body, rotating bearing, drive motor, inlet air duct, outlet air duct, regulating mechanism, lubrication system and foundation support.

3. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generation of the standardized multimodal state dataset includes: Read the acquisition time, sampling frequency, wind turbine type label, component source, and original characteristic values ​​from the multimodal operation data and historical operation data of the wind turbine; Perform outlier removal, missing data completion, and dimensional standardization on the original feature values ​​to generate standardized feature values; By associating the collection time, sampling frequency, wind turbine type label, component source, and standardized feature values, a standardized multimodal state dataset is generated.

4. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generated time-aligned multimodal dataset includes: Read the acquisition time, sampling frequency, wind turbine type label, component source and standardized feature value from the standardized multimodal state dataset, and establish a unified diagnostic time axis according to the preset diagnostic time window; Vibration data, acoustic data, and current data are labeled as high-frequency modal data. Window features are extracted from the high-frequency modal data to generate a high-frequency window feature sequence. The multimodal operation data of the wind turbine, excluding high-frequency modal data, is marked as low-frequency modal data and mapped to a preset diagnostic time window according to the acquisition time to generate a low-frequency window state sequence; Modal time drift is generated based on the offset between the acquisition time of each modality and the center time of the preset diagnostic time window. The high-frequency window feature sequence and the low-frequency window state sequence are corrected according to the modal time drift to generate a time-aligned multimodal dataset.

5. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generated phase residual sequence includes: Read the time-aligned multimodal dataset and generate a sequence of operating condition change features based on the differences in unit load, speed, power, pressure, air volume, regulation opening, and coal mill operation status between adjacent preset diagnostic time windows. Based on the sequence of operating condition changes, the starting speed-up phase, load ramp-up phase, stable ventilation phase, air volume adjustment phase, coal mill start-up / shutdown disturbance phase, and shutdown speed-down phase are divided to generate operating condition phase slice data; Read the maintenance record index data and mark the historical operation data segments corresponding to fault repair time, downtime alarm time, component replacement time, manual inspection anomaly mark and repair component mark as historical anomaly segments; From the historical operating data after removing historical abnormal segments, select data that have the same fan type, operating condition phase, preset unit load range, preset speed range, preset air volume range, preset adjustment opening range, coal mill commissioning status and component source as the current preset diagnostic time window, and generate the same phase load benchmark according to the median of the standardized characteristic values ​​of each mode; Subtract the standardized feature value of the corresponding mode in the same phase load benchmark from the standardized feature value of each mode within the current preset diagnostic time window to obtain the vibration phase residual, acoustic phase residual, temperature phase residual, current phase residual, pressure phase residual, air volume phase residual, power phase residual and lubricating oil state phase residual. Then, arrange them according to the acquisition time, fan type label and component source to generate the working condition phase residual sequence.

6. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generated modal credibility markers and conflict modal markers include: Read the phase residual sequence of the operating condition and establish a modal residual alignment table according to the same preset diagnostic time window, the same fan type label and the same component source; Count the number of missing data, the length of consecutive missing data, and the number of valid residuals for each modality, and generate data integrity markers; The historical residual fluctuation range is generated based on the residual distribution between historical operating data after removing historical abnormal segments and the load benchmark with the same phase, and the amplitude of the phase residual change of each mode is calculated based on the modal residual alignment table to generate short-term stability markers. A trend consistency mark is generated based on the increase or decrease direction of the phase residuals of each mode within the same preset diagnostic time window. A response consistency mark of adjacent components is generated based on the order of residual occurrence and the direction of residual increase or decrease of adjacent components within consecutive preset diagnostic time windows. Modal credibility labels are generated based on the combination of data integrity labels, short-term stability labels, trend consistency labels, and adjacent component response consistency labels. Modalities that are missing continuity, have short-term jumps, have opposite trend directions, and have no corresponding responses from adjacent components are labeled as conflicting modes, and conflicting mode labels are generated.

7. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 6, characterized in that, The generation of conflict mode markers includes: Integrity scores are generated based on the ratio of the number of valid residuals to the number of expected residuals for each modality, and integrity deductions are generated based on the length of consecutive missing values. A stability score is generated based on the comparison between the variation amplitude of phase residuals in each mode and the historical residual fluctuation range. A trend score is generated based on the number of times the phase residuals of each mode increase or decrease in the same direction within the same preset diagnostic time window, and an adjacency response score is generated based on the number of times adjacent components exhibit residual responses in the same direction within consecutive preset diagnostic time windows. Modal confidence scores are generated based on integrity scores, integrity deductions, stability scores, trend scores, and adjacency response scores. Modal confidence labels are then generated according to the intervals to which the modal confidence scores belong. Continuously missing modes, short-term abrupt change modes, and modes with opposite trends and no corresponding response from adjacent components are written into conflict mode tags.

8. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generated component topology fault propagation residual map includes: Read the connection relationship of the fan components, set the fan body, rotating bearing, drive motor, inlet air duct, outlet air duct, regulating mechanism, lubrication system and foundation support as component nodes, write component node mark for each component node, set mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship as connection edge, and generate component topology diagram; The phase residual sequence of the operating condition is mapped to the component node according to the component source, and the modal confidence label and conflict mode label are mapped to the component node and the connection edge to generate component residual topology data. Based on the phase residual difference, residual increase / decrease direction and modal confidence marker of the component nodes at both ends of the connection edge within the same preset diagnostic time window, the adjacent response residual is calculated. Starting from the component node with the highest residual magnitude, backtrack along the connecting edge and calculate the reverse source residual based on the order of residual occurrence and the difference in residual intensity between the upstream and downstream component nodes. Based on the order of residual occurrence, direction of residual increase / decrease, and type of connection edge of multiple component nodes within a continuous preset diagnostic time window, cross-node consistency residuals are calculated. Then, by combining adjacency response residuals, reverse source residuals, cross-node consistency residuals, modal confidence markers, and conflict modal markers, a component topology fault propagation residual map is generated.

9. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The generation of wind turbine fault state characterization includes: An improved liquid time constant network is constructed, which includes a reliable mode selection layer, an impact energy amplification layer, a polarization complex liquid evolution layer, a phase coupling state propagation layer, and a conflict mode bypass verification layer. The operating condition phase residual sequence, modal confidence label, and conflict mode label are input into the confidence mode selection layer to extract the confidence residual data and conflict residual data, and generate the confidence mode input sequence and conflict mode bypass sequence. The reliable modal input sequence is input into the impact energy amplification layer, and the vibration phase residual sudden increase segment, acoustic phase residual pulse segment, current phase residual period shift segment, pressure phase residual oscillation segment and air volume phase residual fluctuation segment are extracted to generate the impact-enhanced liquid input sequence. Phase shift is generated based on the phase residual difference between consecutive preset diagnostic time windows in the impact-enhanced liquid input sequence. The amplitude residual is written into the real part of the complex liquid state, and the phase shift is written into the imaginary part of the complex liquid state. Continuous state evolution is performed along the liquid time constant update path to generate a complex liquid evolution state. The complex liquid evolution state and the component topology fault propagation residual map are input into the phase coupling state propagation layer. The component phase coupling propagation is performed along the mechanical connection relationship, aerodynamic flow relationship, vibration transmission relationship, heat conduction relationship and electrical coupling relationship to generate the topology coupling liquid state. The conflict mode bypass sequence is input into the conflict mode bypass verification layer to generate conflict mode verification evidence. Based on the topologically coupled liquid state and the conflict mode verification evidence, a wind turbine fault state characterization is generated.

10. The wind turbine fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The fault diagnosis results of the output fan include: Read the topologically coupled liquid state, conflict mode verification evidence, component node markers, and preset diagnostic time window markers from the wind turbine fault state characterization; Perform fault category decoding on the topologically coupled liquid state to generate fault categories; The location of the faulty component is generated based on the component node markers and the residual concentration locations in the component topology fault propagation residual map. Fault levels are classified based on the intensity of anomalies in the fault state characterization of wind turbines and the number of continuously preset diagnostic time windows. Based on modal confidence markers, conflict modal verification evidence, adjacent response residuals, reverse source residuals, and cross-node consistency residuals, diagnostic confidence and diagnostic evidence chains are generated. Then, the fault category, fault component location, fault level, diagnostic confidence and diagnostic evidence chains are combined to output the wind turbine fault diagnosis results.

Images