A method for determining a fault information based on a wind turbine generator system

Abstract

The present application relates to the technical field of wind turbine fault diagnosis, in particular to a kind of determination method of wind turbine fault information based on, comprising: obtaining multidimensional monitoring data of structure dimension data and process dimension data;Based on process dimension data, generate dynamic operation fingerprint sequence, based on structure dimension data, generate static structure feature map;Extract time variation trend related fingerprint features in fingerprint sequence, identify component inherent state related structure features in atlas, associatedly generate multi-dimensional feature pairs;Check correlation degree and exclude invalid pairs, retain effective feature pair set;Introduce historical operation log to trace state evolution, generate enhanced feature unit with time context annotation;Build feature evolution track, combined with geographical environment correction;Potential fault mode and confidence level are calculated using the corrected trajectory.The method improves the accuracy of fault information determination by dimension-specific feature modeling and historical evolution tracing.

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator fault diagnosis technology, and in particular to a method for determining fault information of wind turbine generators. Background Technology

[0002] Current fault diagnosis methods for wind turbine generators often rely on feature analysis by integrating multi-source monitoring data such as vibration, temperature, and power. However, they fail to distinguish between the dynamic temporal attributes of the data and the inherent structural attributes of the equipment. They mix time-series data reflecting changes in operating status with static data characterizing component design parameters, making it difficult to establish a clear correlation between dynamic trend features and static attribute features. Furthermore, the diagnostic process depends on real-time or short-term monitoring data, with few systems incorporating historical maintenance logs to trace the evolution of feature states over time. This lack of temporal context makes fault precursor identification susceptible to short-term fluctuations.

[0003] Under this technical approach, the blindness of feature fusion reduces the accuracy of fault correlation judgment, while ignoring historical operation and maintenance information makes it difficult to capture the gradual process of faults from latent to manifest. To address this, it is necessary to establish a dynamic operational fingerprint sequence that independently reflects the operational trend for process-dimensional data, generate a static structural feature map representing the inherent state of components for structural-dimensional data, extract time trend and inherent state features, and associate them to generate multi-dimensional feature pairs; at the same time, for effective feature pairs, historical operation and maintenance logs are introduced to trace the state evolution and generate enhanced feature units with time context annotations to solve the problems of ambiguous feature association and insufficient time dimension information in existing methods. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for determining fault information of wind turbine generator sets.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for determining fault information of wind turbine generator sets, comprising:

[0006] Acquire monitoring data from multiple dimensions, including structural and process dimensions;

[0007] Based on process dimension data, a dynamic operation fingerprint sequence for each wind turbine is established;

[0008] Based on structural dimension data, a static structural feature map is generated for each wind turbine.

[0009] Fingerprint features related to time-varying trends are extracted from the dynamic running fingerprint sequence, and structural features related to the inherent state of the component are identified from the static structural feature map. The extracted fingerprint features and the identified structural features are preliminarily associated to generate multidimensional feature pairs.

[0010] The multidimensional feature pairs are subjected to correlation degree verification, and invalid feature pairs with correlation degree lower than a preset correlation degree threshold are filtered out, while the set of valid feature pairs with correlation degree that meets the preset correlation degree threshold is retained.

[0011] For the set of effective feature pairs, historical operation and maintenance logs of wind turbines are introduced, and the state evolution of each effective feature pair in the set is traced to generate an enhanced feature unit with time context annotation.

[0012] Based on the temporal context annotation of the enhanced feature units, feature evolution trajectories spanning multiple historical monitoring cycles are constructed;

[0013] The environmental factors are corrected for the feature evolution trajectory by combining the actual geographical environment information of the target wind turbine group;

[0014] By using the corrected feature evolution trajectory, the potential fault modes of wind turbines and the corresponding confidence levels are calculated.

[0015] As a further aspect of the present invention, establishing the dynamic operation fingerprint sequence for each wind turbine includes:

[0016] Structural dimension data refers to measurements related to the inherent physical properties of wind turbines;

[0017] The structural dimension data includes blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information;

[0018] Define a standardized sampling time window and corresponding sampling frequency;

[0019] According to the sampling frequency, within the sampling time window, the timing data of gearbox lubricating oil temperature, generator winding temperature, tower oscillation amplitude, and active power output are collected synchronously.

[0020] Timestamp alignment is performed on the time-series data acquired synchronously.

[0021] After the timestamp alignment process is completed, the time-series data are spliced ​​and encoded according to a preset arrangement order to form the dynamic running fingerprint sequence.

[0022] As a further aspect of the present invention, generating a static structural feature map for each wind turbine includes:

[0023] Process dimension data refers to the dynamic parameter measurements of a wind turbine that change over time during operation.

[0024] The process dimension data includes gearbox lubricating oil temperature time series data, generator winding temperature time series data, tower sway amplitude time series data, active power output time series data, yaw system action time series data, and wind speed and direction sensing time series data.

[0025] Create a topology graph with the names of key components of the wind turbine as index nodes;

[0026] The blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information are used as attribute labels to fill the corresponding key component name index nodes in the topology diagram.

[0027] In the topology diagram, directed edges connecting the index nodes of the names of each key component are drawn based on the physical connection relationship of the wind turbine and the energy transfer path.

[0028] Each directed edge is assigned a weight value, which is determined based on the design dependency strength or historical failure propagation frequency between the two critical components connected.

[0029] The topological relationship graph with completed attribute filling and weight assignment is output as the static structural feature map.

[0030] As a further aspect of the present invention, extracting fingerprint features related to the time-varying trend from the dynamically running fingerprint sequence includes:

[0031] The dynamic fingerprint sequence is decomposed by wavelet packet transform to obtain wavelet packet coefficients in different frequency bands;

[0032] Calculate the energy value of the wavelet packet coefficients in each frequency band, and select the top few frequency bands with the largest energy values ​​as the dominant frequency bands;

[0033] Extract the statistical features of the wavelet packet coefficients corresponding to the dominant frequency band, including mean, variance, skewness, and kurtosis;

[0034] The identification of structural features related to the inherent state of components from static structural feature maps includes:

[0035] Traverse all key component name index nodes in the static structural feature map and read their attribute labels;

[0036] Calculate the centrality index of each key component name index node;

[0037] Key component name index nodes with centrality indices higher than a preset centrality threshold are selected, and their attribute labels and centrality indices are combined as the structural features.

[0038] As a further aspect of the present invention, the step of initially associating the extracted fingerprint features and the identified structural features to generate multidimensional feature pairs includes:

[0039] A feature descriptor is created for each fingerprint feature, the feature descriptor containing feature type, source time series data identifier, and dominant frequency band information;

[0040] A feature descriptor is created for each structural feature, and the feature descriptor includes the corresponding key component name, attribute type, and centrality index;

[0041] Construct an association rule knowledge base, which defines the physical or logical association relationships between different types of fingerprint features and different types of structural features;

[0042] The feature descriptors of fingerprint features and structural features are input into the association rule knowledge base for matching;

[0043] If a match is successful, the matching fingerprint features and structural features are combined into a multidimensional feature pair, and the association rule on which the match was based is recorded.

[0044] As a further aspect of the present invention, the correlation verification of the multidimensional feature pairs includes:

[0045] Obtain actual fault records of the target wind turbine within a preset historical verification period;

[0046] Examine the changes in each multidimensional feature to its associated fingerprint and structural features before and after the actual fault record occurred;

[0047] If both fingerprint features and structural features exhibit abnormal changes beyond the normal range before the fault occurs, the correlation between the multidimensional feature pair is determined to be high.

[0048] If only one feature shows an abnormal change, or if the changes in the fingerprint feature and structural feature are not significantly correlated with the time of the fault occurrence, then the correlation of the multidimensional feature pair is determined to be low.

[0049] The multidimensional feature pairs with high correlation are retained to form the effective feature pair set.

[0050] As a further aspect of the present invention, the state evolution tracing for each valid feature pair in the set includes:

[0051] Read the historical operation and maintenance log records, which include the time of each maintenance operation, the component to be operated on, the operation type, and the post-operation status remarks;

[0052] For the key components corresponding to the structural features in the effective feature pairs, search the historical operation and maintenance logs for all maintenance operation records involving the key components;

[0053] The maintenance operation records found are arranged in chronological order to form the maintenance history timeline of the key components;

[0054] The maintenance history timeline of the key components is superimposed and compared with the time change curve of the fingerprint features in the effective feature pair;

[0055] On the overlay comparison image, the state of the fingerprint features at each maintenance operation is marked, thereby generating the enhanced feature unit with time context annotation, which includes the maintenance event time point and the corresponding maintenance type.

[0056] As a further aspect of the present invention, the construction of the feature evolution trajectory spanning multiple historical monitoring periods includes:

[0057] Select a target historical monitoring period range;

[0058] Extract all maintenance event time points that occurred within the target's historical monitoring period from the temporal context annotations of the enhanced feature units;

[0059] Using the maintenance event time point as the dividing point, the target historical monitoring period is divided into multiple continuous evolution stages;

[0060] Within each evolutionary stage, the stage statistics of fingerprint features in the enhanced feature units are calculated, and the key attribute values ​​of the structural features at the end of each evolutionary stage are recorded.

[0061] The stage statistics and key attribute values ​​of all evolutionary stages are linked together in chronological order to form the feature evolution trajectory, which is presented in the form of a stage sequence.

[0062] As a further aspect of the present invention, the step of modifying the feature evolution trajectory by incorporating the actual geographical environment information of the target wind turbine cluster includes:

[0063] The actual geographic environment information is obtained, including the altitude of the wind turbine location, the average annual temperature, the air density, the turbulence intensity level, and the salt spray corrosion level.

[0064] An environmental factor impact comparison table is established, which defines the quantitative impact factors of different environmental factor values ​​on the aging rate and performance degradation coefficient of different components of wind turbines.

[0065] From each evolutionary stage of the feature evolution trajectory, feature parameters related to the current environmental factors are extracted;

[0066] Based on the environmental factor impact comparison table, find the quantitative impact factors corresponding to the extracted feature parameters;

[0067] Using the found quantitative influencing factors, numerical compensation or conversion is performed on the stage statistics or key attribute values ​​of the corresponding evolution stages in the feature evolution trajectory to obtain the corrected feature evolution trajectory.

[0068] As a further aspect of the present invention, the step of using the modified feature evolution trajectory to deduce the potential fault modes of the wind turbine and the corresponding confidence levels includes:

[0069] The stage statistics sequence and key attribute value sequence of each evolution stage are extracted from the corrected feature evolution trajectory to form the fault prediction input feature set;

[0070] The fault prediction input feature set is input into a pre-trained long short-term memory neural network model, which is trained using historical fault data and corresponding feature evolution trajectories as samples.

[0071] The time dependence of the input feature set is modeled by multiple memory units of the long short-term memory neural network model, and the hidden state sequence corresponding to each evolution stage is output.

[0072] The final hidden state sequence is input into the fully connected layer for classification calculation to obtain the occurrence probability of various potential fault modes, including gearbox bearing wear, generator insulation deterioration, blade fatigue damage, tower foundation settlement, and yaw system jamming.

[0073] The fault mode with the highest probability of occurrence is selected as the main prediction result, and the probability value is used as the confidence level corresponding to this fault mode.

[0074] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0075] This study employs a technique that uses process-dimensional data to establish dynamic operational fingerprint sequences and structural-dimensional data to generate static structural feature maps. Fingerprint features related to time-varying trends are extracted from the dynamic sequences, while structural features related to the inherent state of components are identified from the static maps. These two are then preliminarily correlated to generate multi-dimensional feature pairs, and the correlation degree is verified. This multi-dimensional modeling establishes a clear correspondence between dynamic evolution features and static attribute features. The correlation degree verification eliminates invalid feature pairs, and the retained valid feature pairs accurately reflect the coupling relationship between changes in operational state and the inherent properties of components. This enhances the ability of feature combinations to characterize fault precursors and lays a more reliable feature foundation for subsequent fault inference.

[0076] Historical operation and maintenance logs are introduced for the set of effective feature pairs. The state evolution of each feature pair is traced to generate enhanced feature units with time context annotations. This constructs a feature evolution trajectory spanning historical monitoring cycles and incorporates geographical environment corrections. The introduction of historical operation and maintenance logs restores the changes of features throughout the entire component lifecycle. The time context enhances the explanatory power of features for fault evolution stages, and the geographical environment correction makes the trajectory more closely match the actual operating conditions of the unit. As a result, the potential fault modes and confidence levels calculated based on the trajectory are closer to real risk scenarios, improving the accuracy of fault information determination. Attached Figure Description

[0077] Figure 1 This is a flowchart of a method for determining fault information of a wind turbine generator set according to the present invention;

[0078] Figure 2 A flowchart for establishing a dynamically running fingerprint sequence;

[0079] Figure 3 A flowchart for extracting fingerprint features and structural features;

[0080] Figure 4 A time-series change diagram of bearing fingerprint characteristics and operation and maintenance intervention response for wind turbine-A generator;

[0081] Figure 5 Thermographic diagram of aging rate of multiple fans and components. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0083] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0084] See Figure 1 This process acquires multi-dimensional monitoring data, including structural and process dimensions. Structural data refers to measurements related to the inherent physical properties of the wind turbine, while process data refers to measurements of dynamic parameters that change over time during operation. Based on the process data, a dynamic operational fingerprint sequence is established for each wind turbine. Based on the structural data, a static structural feature map is generated for each wind turbine. Fingerprint features related to time-varying trends are extracted from the dynamic operational fingerprint sequence, and structural features related to the inherent state of components are identified from the static structural feature map. The extracted fingerprint features and the identified structural features are initially correlated to generate multi-dimensional feature pairs. The correlation degree of these multi-dimensional feature pairs is verified, and invalid feature pairs with correlation degrees below a preset correlation degree threshold are removed, retaining the set of valid feature pairs that meet the preset correlation degree threshold. For the set of valid feature pairs, historical operation and maintenance logs of the wind turbine are introduced, and the state evolution of each valid feature pair in the set is traced to generate enhanced feature units with time context annotations. Based on the time context annotations of the enhanced feature units, a feature evolution trajectory spanning multiple historical monitoring cycles is constructed. By incorporating the actual geographical environment information of the target wind turbine cluster, environmental factors are corrected for the feature evolution trajectory. Using the corrected feature evolution trajectory, the potential fault modes of the wind turbines and the corresponding confidence levels are calculated.

[0085] See Figure 2In one embodiment of the present invention, a standardized sampling time window is set to 3600 seconds and the corresponding sampling frequency is 1 Hz. Within the 3600-second sampling time window at a sampling frequency of 1 Hz, time-series data of gearbox lubricating oil temperature, generator winding temperature, tower oscillation amplitude, and active power output are simultaneously acquired. The synchronously acquired time-series data includes raw measurements from the sensor array. For example, the gearbox lubricating oil temperature time-series data is recorded as a sequence containing 3600 data points within the sampling window, and the generator winding temperature time-series data... The time-series data of the tower oscillation amplitude and active power output are recorded with the same length. The time-series data acquired synchronously are timestamped and aligned using interpolation to ensure all data points correspond to the same time reference. The time-series data after timestamping are then spliced ​​and encoded according to the order of gearbox lubricating oil temperature, generator winding temperature, tower oscillation amplitude, and active power output, forming a dynamic operating fingerprint sequence. The encoding process of the dynamic operating fingerprint sequence can be expressed by the following formula:

[0086]

[0087] in: This indicates that the fingerprint sequence is being run dynamically. This represents a vector of time-series data for gearbox lubricating oil temperature. This represents a time-series data vector of generator winding temperature. This represents a time-series data vector of the tower's oscillation amplitude. Represents the active power output timing data vector, symbol This indicates a vector concatenation operation. In practice, data comparison is reflected in the differences in the dynamic operating fingerprint sequences of different wind turbines. For example, the gearbox lubricating oil temperature time series data of wind turbine A located in a wind farm shows periodic fluctuations, while the same data of another wind turbine B shows a stable trend. This comparison helps to identify individual operating characteristics.

[0088] In some embodiments, generating a static structural feature map for each wind turbine includes the following process: process dimension data includes gearbox lubricating oil temperature time-series data, generator winding temperature time-series data, tower oscillation amplitude time-series data, active power output time-series data, yaw system action time-series records, and wind speed and direction sensing time-series data. A topology map is created with the names of key components of the wind turbine as index nodes. Key component names include blades, hubs, towers, transmission systems, and foundation anchoring structures. Blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information are used as attribute labels to fill the corresponding key component name index nodes in the topology map. For example, the blade length attribute label is assigned "50 meters," the hub height attribute label is assigned "80 meters," and the tower material attribute label is assigned "steel." The transmission system design parameter is labeled "gear ratio 1:100", and the foundation anchorage structure information is labeled "concrete base". In the topology diagram, directed edges connecting the index nodes of each key component name are drawn based on the physical connection relationship and energy transfer path of the wind turbine. These directed edges include those from the blade to the hub, from the hub to the transmission system, from the transmission system to the tower, and from the tower to the foundation anchorage structure. Each directed edge is assigned a weight value, determined based on the design dependency strength or historical fault conduction frequency between the two connected key components. For example, the weight of the directed edge from the blade to the hub is set to 0.9 based on the design dependency strength, and the weight of the directed edge from the hub to the transmission system is set to 0.7 based on the historical fault conduction frequency. The topology diagram with completed attribute filling and weight assignment is output as a static structural feature map. In some embodiments, data comparison is reflected through the static structural feature maps of different wind turbines. For example, the tower material attribute of wind turbine A is "steel", while the tower material attribute of wind turbine B is "composite material". This difference affects the attribute labels and weight assignments in the topology diagram.

[0089] Optionally, when establishing a dynamic running fingerprint sequence, the sampling time window and sampling frequency can be adjusted according to monitoring needs. For example, the sampling time window can be set to 7200 seconds, and the sampling frequency can be set to 0.5 Hz. The range of synchronously acquired time-series data can be expanded to include vibration sensing time-series data. The timestamp alignment process uses a linear interpolation method to ensure data consistency. Optionally, when generating a static structural feature map, the key component name index nodes of the topology graph can be expanded to include nodes such as generators and yaw systems. The weight values ​​of directed edges are calculated based on the dependency table in the design document provided by the wind turbine manufacturer. The weight assignment formula uses normalized frequency statistics, but this will not be elaborated here to avoid divergence.

[0090] It is understandable that the establishment of dynamic operational fingerprint sequences relies on the synchronous acquisition and encoding of process-dimensional data. Alignment processing of time-series data on gearbox lubricating oil temperature, generator winding temperature, tower oscillation amplitude, and active power output ensures consistency in the time dimension. The splicing operation forms a unified vector, facilitating subsequent feature extraction. Similarly, the generation of static structural feature maps integrates structural-dimensional data with topological relationships. Blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information serve as attribute labels to provide static descriptions of components. Directed edges and weight values ​​model the physical connections and fault propagation paths between components.

[0091] See Figure 3 In one embodiment of the present invention, wavelet packet transform decomposition is performed on the dynamic running fingerprint sequence. The dynamic running fingerprint sequence is a vector containing multiple time-series data concatenations. The wavelet packet transform decomposition uses the db4 wavelet basis function for a 3-level decomposition to obtain wavelet packet coefficients of 8 different frequency bands from low frequency to high frequency. The energy value of the wavelet packet coefficients of each frequency band is calculated. The energy value is the sum of the squares of all wavelet packet coefficients in the same frequency band. The top 3 frequency bands with the largest energy values ​​are selected as the dominant frequency bands. The statistical features of the wavelet packet coefficients corresponding to the dominant frequency bands are extracted. The statistical features include mean, variance, skewness, and kurtosis. The mean is calculated as the arithmetic mean of the wavelet packet coefficients of the dominant frequency band. The variance is calculated as the average of the squares of the differences between the wavelet packet coefficients of the dominant frequency band and their mean. Skewness is used to measure the asymmetry of the distribution of wavelet packet coefficients. Kurtosis is used to measure the steepness of the distribution of wavelet packet coefficients. In practice, data comparison is reflected in the differences in dominant frequency bands decomposed from the dynamic operation fingerprint sequences of different wind turbines or different time periods. For example, under normal operation, the wavelet packet energy of the dynamic operation fingerprint sequence may be concentrated in the low frequency band, while when early anomalies occur, the energy may shift to the high frequency band. This shift in dominant frequency band can be used as a basis for comparison.

[0092] In some embodiments, identifying structural features related to the inherent state of components from a static structural feature map includes the following process: traversing all key component name index nodes in the static structural feature map, where key component name index nodes include blades, hubs, towers, transmission systems, and foundation anchoring structures; reading the attribute labels of the key component name index nodes, where attribute labels include blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information; and calculating the centrality index of each key component name index node, where the centrality index quantifies the importance of the node in the topology graph. The calculation formula is:

[0093]

[0094] in: Represents a node The central indicator, This represents the set of all nodes in the static structural feature map. Represents a node To the node The total number of shortest paths, Represents a node To the node Passing Nodes The number of shortest paths and the centrality index calculation depend on the connectivity and weight values ​​of directed edges in the static structural feature map. Key component name index nodes with centrality indices higher than a preset centrality threshold are selected, for example, a threshold of 0.5. These key component name index nodes with centrality indices higher than 0.5, along with their attribute labels and centrality indices, are combined as structural features. In some embodiments, data comparison is reflected through the centrality indices of nodes in the static structural feature maps of different wind turbines. For example, in the map of wind turbine A, the centrality index of the "transmission system" node is 0.8, while in the map of wind turbine B, the centrality index of the same node is 0.4. This difference reflects the different importance of key components in different structural designs.

[0095] Optionally, the number of wavelet packet transform decomposition levels and the type of wavelet basis functions can be adjusted. For example, a 4-level decomposition can be performed using the sym6 wavelet basis function. The number of dominant frequency bands can be determined based on the cumulative energy contribution rate; for example, frequency bands with a cumulative energy contribution rate exceeding 85% can be selected as dominant frequency bands. Optionally, the centrality index can be calculated using metrics other than betweenness centrality, such as degree centrality or proximity centrality. The preset centrality threshold can be determined based on the statistical distribution percentiles of historical graph nodes. It can be understood that the process of extracting fingerprint features from a dynamically running fingerprint sequence relies on signal processing methods. Wavelet packet transform decomposition maps the time-series signal to the time-frequency domain, energy calculation identifies the main components in the signal, and statistical feature extraction obtains distribution morphology information from the dominant frequency band coefficients. It can also be understood that the process of identifying structural features from a static structural feature graph relies on graph theory analysis methods. Traversing nodes and reading attribute labels obtains static descriptions. Centrality index calculation is based on the node's path control capability in the network, and the screening operation identifies key components based on a quantization threshold.

[0096] In one embodiment of the present invention, a feature descriptor is established for each fingerprint feature. The feature descriptor includes feature type, source time-series data identifier, and dominant frequency band information. For example, for a fingerprint feature extracted from wavelet packet decomposition, its feature type is recorded as "wavelet packet coefficient kurtosis", its source time-series data identifier is recorded as "gearbox lubricating oil temperature time-series data", and its dominant frequency band information is recorded as "frequency band number 3". A feature descriptor is also established for each structural feature. The feature descriptor includes the corresponding key component name, attribute type, and centrality index. For example, for a structural feature identified from a static structural feature map, its corresponding key component name is recorded as "transmission system", its attribute type is recorded as "gear ratio", and its centrality index is recorded as "0.85". A knowledge base of association rules is constructed, defining the physical or logical relationships between different types of fingerprint features and different types of structural features. This knowledge base is stored in rule tables. For example, a rule might specify that when a fingerprint feature's feature type is "wavelet packet coefficient kurtosis" and its source time-series data is identified as "gearbox lubricating oil temperature time-series data," it can be associated with a structural feature whose corresponding key component name is "transmission system" and its attribute type is "gear ratio." The feature descriptors of the fingerprint and structural features are input into the knowledge base for matching. The matching process compares the fields in the feature descriptors with the rule conditions in the knowledge base. If a match is successful, the matched fingerprint and structural features are combined into a multi-dimensional feature pair, and the association rule upon which the match was based is recorded. In practice, data comparison reveals differences in the initial multidimensional feature pairs generated by different wind turbines or under different operating conditions. For example, for wind turbine A, its fingerprint feature of "gearbox oil temperature high frequency energy" is successfully matched with the structural feature of "transmission system gear ratio" to generate a multidimensional feature pair. However, due to different structural features, the same fingerprint feature of wind turbine B may be matched with the structural feature of "tower material" to generate different multidimensional feature pairs.

[0097] In some embodiments, the correlation verification of multidimensional feature pairs includes the following process: obtaining actual fault records of the target wind turbine within a preset historical verification period, such as the past 12 months; the actual fault records include the fault occurrence time, the faulty component, and the fault type; checking the changes of the fingerprint features and structural features associated with each multidimensional feature pair before and after the actual fault record; the checking operation involves retrieving the numerical sequence of the fingerprint features within the time window before and after the fault occurrence, and the status records of the component attributes corresponding to the structural features before and after the fault occurrence. If both the fingerprint features and structural features show abnormal changes exceeding the normal range before the fault occurrence, the correlation of the multidimensional feature pair is determined to be high. For example, if the value of the fingerprint feature continuously exceeds the upper limit of the threshold for 72 hours before the fault occurrence, and the status record of the component attribute corresponding to the structural feature shows "increased wear," then the correlation is determined to be high. If only one feature shows abnormal changes, or if the changes of the fingerprint features and structural features are not significantly correlated with the fault occurrence time, then the correlation of the multidimensional feature pair is determined to be low. For example, if the fingerprint features show abnormalities but the structural features have no related records, or if the changes of both occur one week after the fault occurrence. Multidimensional feature pairs with high correlation are retained to form a valid feature pair set. In some embodiments, data comparison is reflected by the change in the feature pair set before and after verification. For example, 15 multidimensional feature pairs are generated before verification. After correlation verification, 8 of them are determined to be highly correlated and retained, while the other 7 are screened out due to low correlation. This reflects the method's ability to filter out random or spurious correlations.

[0098] Optionally, the construction of the association rule knowledge base can be based on domain knowledge graphs or historical fault association analysis, and the rule conditions can be extended to include combined constraints containing dominant frequency band information and attribute type values. Optionally, the logic for determining the degree of association can be formalized as a scoring function based on temporal consistency and the magnitude of change. The calculation formula is:

[0099]

[0100] in: This represents the correlation score between multidimensional feature pairs. Is it the time of fingerprint feature anomaly? Time of failure Proximity indicator function It refers to the magnitude of abnormal changes in fingerprint features. It is the magnitude of change in the structural characteristic state. , , These are preset weighting coefficients, when If the correlation exceeds a preset threshold, it is considered to be high.

[0101] Understandably, the initial association step links dynamic fingerprints with static structural information through a predefined association rule knowledge base. Feature descriptors provide a standardized matching interface, and a successful match means the identification of a potential dynamic-static association pattern. Similarly, the association verification step uses historical fault records as ground-level data for validation. By examining the collaborative changes in features before and after the actual fault, the reliability of the association is assessed, and filtering out low-association pairs improves the accuracy of subsequent analyses.

[0102] In one embodiment of the present invention, historical maintenance log records are read. These log records include the time, target component, operation type, and post-operation status notes for each maintenance operation. For key components corresponding to structural features in valid feature pairs (e.g., a key component "generator bearing"), all maintenance operation records involving "generator bearing" are searched in the historical maintenance log records. These records are arranged chronologically to form a maintenance history timeline for "generator bearing." The maintenance history timeline for "generator bearing" is then overlaid and compared with the time variation curves of fingerprint features in the valid feature pairs. These fingerprint feature time variation curves are curves representing the changes in feature values ​​potentially related to the generator bearing state over time, derived from dynamically running fingerprint sequences. The state of the fingerprint features at the time of each maintenance operation is marked on the overlay comparison image, thereby generating enhanced feature units with time context annotations. These time context annotations include the maintenance event time point and the corresponding maintenance type. In practice, data comparison can be achieved by comparing the enhanced feature units of two wind turbines with different maintenance histories. For example, the "generator bearing" of wind turbine A has two lubrication maintenance records in the past year. On its superimposed comparison map, the fingerprint features show a specific change pattern near the two maintenance time points. On the other hand, the same component of wind turbine B has no maintenance records, and its fingerprint feature curve shows a different evolution trend. This difference reflects the impact of maintenance activities on the feature state.

[0103] In some embodiments, constructing a feature evolution trajectory spanning multiple historical monitoring periods includes the following steps: First, a target historical monitoring period range is selected, for example, from January 1, 2025 to December 31, 2025. Second, all maintenance event time points occurring within the target historical monitoring period range are extracted from the temporal context annotations of the enhanced feature units. Third, the target historical monitoring period range is divided into multiple consecutive evolutionary stages, using the maintenance event time points as dividing points. For example, if there are two maintenance events within the target historical monitoring period, with time points of April 1, 2025 and September 1, 2025, the entire period is divided into three stages: Stage 1 (January 1 to March 31), Stage 2 (April 2 to August 31), and Stage 3 (September 2 to December 31). Fourth, within each evolutionary stage, the stage statistics of the fingerprint features in the enhanced feature units are calculated, and the key attribute values ​​of the structural features at the end of each evolutionary stage are recorded. The calculation of stage statistics may include mean, standard deviation, slope, etc., such as calculating the mean of the fingerprint feature time series within each stage. The calculation formula is:

[0104]

[0105] in: This represents the stage mean of fingerprint features at the current evolutionary stage. This indicates the number of fingerprint feature data points in the current evolutionary stage. Indicates the first The values ​​of fingerprint feature data points. Key attribute values ​​of structural features may be read directly from static structural feature maps, or attribute values ​​updated by detection at the end of a stage. The stage statistics and key attribute values ​​of all evolutionary stages are concatenated in chronological order to form a feature evolution trajectory, which is presented as a stage sequence. See Table 1 for an exemplary comparison of feature evolution trajectory data.

[0106] Table 1: Comparison of Data on the Evolution Trajectory Stages of Different Wind Turbines

[0107]

[0108] Optionally, when generating enhanced feature units, the overlay and comparison operation can be completed in specialized time series analysis software, and the labeling operation can automatically record the statistics of fingerprint features within a specific time window before and after the maintenance event as part of the annotation. Optionally, when dividing evolutionary stages, if two maintenance event time points are too close, they can be merged and treated as a single maintenance interval; the selection of stage statistics can be adjusted according to the type of fingerprint features. For example, for vibration-related fingerprint features, the root mean square value can be selected as the stage statistics.

[0109] It is understandable that the state evolution tracing step links discrete maintenance records with continuous feature changes on a timeline. The maintenance history timeline provides annotations of potential causes of state changes, and overlay comparison reveals the relationship between maintenance interventions and feature responses. Similarly, the feature evolution trajectory construction step segments continuous monitoring cycles by maintenance events. Stage statistics quantify the behavior of features within stable operating ranges, and key attribute values ​​record the stage nodes of component states, forming a structured historical sequence of equipment state evolution when strung together over time.

[0110] See Figure 4 In the process of tracing the state evolution and constructing the characteristic trajectory of the wind turbine-A generator bearing, a clear causal relationship was formed between the temporal changes of fingerprint feature values ​​(temperature / vibration comprehensive index) and maintenance intervention events. In specific operations, the dynamic operation fingerprint sequence was extracted as a feature curve that changes continuously over time, and maintenance events in historical maintenance logs were superimposed on this curve as discrete time anchor points. The two key maintenance interventions (lubrication maintenance on 2025-04-01 and lubrication maintenance + clearance adjustment on 2025-09-01) are marked with red dashed lines. The significant changes in fingerprint feature values ​​before and after these interventions reveal the direct relationship between maintenance interventions and characteristic responses: after the first maintenance, the fingerprint feature value dropped from a stage one average of 72.5 to a stage two average of 68.2, indicating that the single lubrication maintenance effectively suppressed the early abnormality of the bearing; while after the second maintenance (including clearance adjustment), although the feature value fluctuated in the short term, it rose overall to a stage three average of 75.1, reflecting that the bearing entered a new operating state range after the clearance adjustment. The time-series graph, serving as a visualization carrier for enhanced feature units, overlays and compares the maintenance history timeline with the fingerprint feature change curve, enabling temporal context annotation of effective feature pairs (generator bearing structural features - vibration fingerprint features). Using maintenance events as dividing points, the entire year of 2025 was divided into three evolutionary stages. Within each stage, the statistical value (mean) of the fingerprint features and the key structural attribute (bearing clearance) together constitute the feature evolution trajectory of the wind turbine-A generator bearing, providing structured time-series input for subsequent environmental correction and fault mode estimation.

[0111] In one embodiment of the present invention, actual geographical environment information is obtained, including the altitude, average annual temperature, air density, turbulence intensity level, and salt spray corrosion level of the wind turbine's location. For example, for wind turbine C installed in a coastal area, its actual geographical environment information is recorded as follows: altitude 20 meters, average annual temperature 15°C, air density 1.225 kg / m³, turbulence intensity level B, and high salt spray corrosion level. An environmental factor influence comparison table is established, which defines the quantitative influence factors of different environmental factor values ​​on the aging rate and performance degradation coefficient of different components of the wind turbine. The environmental factor influence comparison table exists in the form of a database table, where one row of records may indicate that when the salt spray corrosion level is "high," the quantitative influence factor for the "surface aging rate" of the "blade" component is 1.5, and the quantitative influence factor for the "corrosion rate" of the "tower" component is 1.8. From each evolutionary stage of the feature evolution trajectory, feature parameters related to the current environmental factors are extracted. For example, from an evolutionary stage where "leading edge corrosion depth" is the key attribute value, this attribute value is extracted as a feature parameter related to the environmental factor "salt spray corrosion level". Based on the environmental factor influence comparison table, the quantitative influence factor corresponding to the extracted feature parameter is found. For example, based on "salt spray corrosion level: high" and the component "blade", the quantitative influence factor is found to be 1.5. Using the found quantitative influence factor, the stage statistics or key attribute values ​​of the corresponding evolutionary stages in the feature evolution trajectory are numerically compensated or converted to obtain the corrected feature evolution trajectory. The numerical compensation can apply a correction formula; for example, the correction calculation for stage statistics or key attribute values ​​is as follows:

[0112]

[0113] in: This represents the corrected stage statistics or key attribute values. This represents the original stage statistics or key attribute values ​​before correction. This represents the coefficient that maps the j-th environmental factor (such as salt spray corrosion level). This indicates the corresponding quantitative influencing factors found in the environmental factors impact comparison table. In practice, data comparison reflects the differences in the correction results of the same feature evolution trajectory under different actual geographical environmental information. For example, for a feature evolution trajectory showing the evolution of "blade leading edge corrosion depth" from 0.1 mm to 0.2 mm, the corrected trajectory value of wind turbine C located in a high salt spray corrosion environment may be converted to an evolution from 0.15 mm to 0.3 mm, while the corrected trajectory value of wind turbine D located in a low salt spray corrosion environment may still be 0.1 mm to 0.2 mm. This comparison highlights the moderating role of environmental factors in equipment degradation assessment.

[0114] In some embodiments, inferring the potential fault modes of a wind turbine and the corresponding confidence levels using the modified feature evolution trajectory includes the following process: Extracting the stage statistics sequence and key attribute value sequence for each evolution stage from the modified feature evolution trajectory to form a fault prediction input feature set. The stage statistics sequence is, for example, a vector of "vibration characteristic stage mean" values ​​arranged in chronological order for multiple consecutive evolution stages, and the key attribute value sequence is a vector of "bearing clearance measurement values" for the corresponding stage. The fault prediction input feature set is then input into a pre-trained long short-term memory (LSTM) neural network model, which is trained using historical fault data and corresponding feature evolution trajectories as samples. The time dependency of the input feature set is modeled using multiple memory units of the LTM neural network model, outputting the hidden state sequence corresponding to each evolution stage. The final hidden state sequence is then input into a fully connected layer for classification calculation to obtain the probability of occurrence of various potential fault modes. Potential fault modes include gearbox bearing wear, generator insulation degradation, blade fatigue damage, tower foundation settlement, and yaw system jamming. The fully connected layer uses a softmax activation function to output the probability value of each fault mode. The fault mode with the highest probability of occurrence is selected as the primary prediction result, and the probability value is used as the confidence level corresponding to this fault mode. In some embodiments, data comparison can be reflected by the prediction results of the same wind turbine at different time points or under different environmental correction conditions. For example, without environmental correction, the input feature set leads the long short-term memory neural network model to predict a 65% probability of "gearbox bearing wear"; after applying environmental correction, due to the feature value being recalculated, the same model may predict a 78% probability of "gearbox bearing wear", and the confidence level changes accordingly. This demonstrates the impact of environmental correction on fault prediction results.

[0115] Optionally, in the environmental factor correction, if multiple environmental factors simultaneously affect the same characteristic parameter, the combined effect of the quantified influencing factors can be reflected in the correction formula through weighted summation or multiplication. Numerical compensation can also use an additive model instead of a multiplicative model. Optionally, in fault mode estimation, the Long Short-Term Memory (LSTM) neural network model can be replaced with other recurrent neural network variants with temporal modeling capabilities, such as gated recurrent unit networks, where the output dimension of the fully connected layers is strictly consistent with the preset number of potential fault mode types. It can be understood that the step of correction based on actual geographical environmental information incorporates the external conditions of equipment operation into the state assessment framework. The environmental factor influence comparison table provides the mapping relationship between environmental stress and equipment degradation effects. The correction formula uses quantified influencing factors to calibrate the observed characteristic trajectory, making the corrected characteristic evolution trajectory more reflective of the inherent degradation process of the equipment under standard or specific environments. It is understandable that the step of using the corrected trajectory to infer the fault mode realizes the mapping from the temporal state sequence to the fault probability. The memory unit of the long short-term memory neural network model captures the long-term dependency pattern of the state evolution. The fully connected layer maps the learned time pattern to the specific fault category probability. Selecting the highest probability value as the prediction result and its confidence provides a quantifiable fault risk estimate.

[0116] See Figure 5The figure illustrates the differences in aging rates among different wind turbines and key components. The graph is organized by component name (blade, tower, gearbox, generator) as rows and wind turbine number (Wind Turbine C, Wind Turbine D, Wind Turbine E, Wind Turbine F, Wind Turbine G) as columns. The numerical value and color depth of each cell collectively represent the aging rate coefficient, with a color gradient from light yellow (low value) to dark red (high value), corresponding to an aging rate quantification range from 0.4 to 0.9. The data distribution shows that the aging rates of all components in Wind Turbine C are generally high, with the tower reaching an aging rate coefficient of 0.92, the highest value in the entire graph. This is highly correlated with the stress effects of the high-salt-spray corrosion environment along the coast where the wind turbine is located. In contrast, the aging rates of all components in Wind Turbine D are at a lower level, with the gearbox having an aging rate coefficient of only 0.40, reflecting the mitigating effect of its environment on equipment degradation. At the component level, the tower is most sensitive to environmental factors, followed by the blades. The aging rates of the gearbox and generator are relatively stable, which is consistent with the differences in material properties, stress states, and environmental exposure levels of different components. The core value of the heat map lies in the fact that, through environmentally corrected aging rate coefficients, it eliminates the interference of geographical environmental differences on equipment degradation assessment, making the aging states of different wind turbines and different components comparable. For example, after correction for the high salt spray environmental factor, the original aging rate coefficient of wind turbine C blade is reduced from the original observed value to 0.85, while the coefficient of wind turbine D blade, due to its low salt spray environment, is only 0.42. This comparison clearly reveals the moderating effect of environmental stress on the equipment degradation process, providing an accurate input basis for subsequent fault mode estimation based on the corrected feature evolution trajectory.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for determining fault information based on wind turbine generator sets, characterized in that, include: Acquire monitoring data from multiple dimensions, including structural and process dimensions; Based on process dimension data, a dynamic operation fingerprint sequence for each wind turbine is established; Based on structural dimension data, a static structural feature map is generated for each wind turbine. Fingerprint features related to time-varying trends are extracted from the dynamic running fingerprint sequence, and structural features related to the inherent state of the component are identified from the static structural feature map. The extracted fingerprint features and the identified structural features are preliminarily associated to generate multidimensional feature pairs. The multidimensional feature pairs are subjected to correlation degree verification, and invalid feature pairs with correlation degree lower than a preset correlation degree threshold are filtered out, while the set of valid feature pairs with correlation degree that meets the preset correlation degree threshold is retained. For the set of effective feature pairs, historical operation and maintenance logs of wind turbines are introduced, and the state evolution of each effective feature pair in the set is traced to generate an enhanced feature unit with time context annotation. Based on the temporal context annotation of the enhanced feature units, feature evolution trajectories spanning multiple historical monitoring cycles are constructed; The environmental factors are corrected for the feature evolution trajectory by combining the actual geographical environment information of the target wind turbine group; By using the corrected feature evolution trajectory, the potential fault modes of wind turbines and the corresponding confidence levels are calculated.

2. The method for determining fault information of a wind turbine generator set according to claim 1, characterized in that, The establishment of the dynamic operational fingerprint sequence for each wind turbine includes: Structural dimension data refers to the measured values ​​related to the inherent physical properties of wind turbines; The structural dimension data includes blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information; Define a standardized sampling time window and corresponding sampling frequency; According to the sampling frequency, within the sampling time window, the timing data of gearbox lubricating oil temperature, generator winding temperature, tower oscillation amplitude, and active power output are collected synchronously. Timestamp alignment is performed on the time-series data acquired synchronously. After the timestamp alignment process is completed, the time-series data are spliced ​​and encoded according to a preset arrangement order to form the dynamic running fingerprint sequence.

3. The method for determining fault information of a wind turbine generator set according to claim 2, characterized in that, The process of generating a static structural feature map for each wind turbine includes: Process dimension data refers to the dynamic parameter measurements of a wind turbine that change over time during operation. The process dimension data includes gearbox lubricating oil temperature time series data, generator winding temperature time series data, tower sway amplitude time series data, active power output time series data, yaw system action time series data, and wind speed and direction sensing time series data. Create a topology graph with the names of key components of the wind turbine as index nodes; The blade length, hub height, tower material properties, transmission system design parameters, and foundation anchoring structure information are used as attribute labels to fill the corresponding key component name index nodes in the topology diagram. In the topology diagram, directed edges connecting the index nodes of the names of each key component are drawn based on the physical connection relationship of the wind turbine and the energy transfer path. Each directed edge is assigned a weight value, which is determined based on the design dependency strength or historical failure propagation frequency between the two critical components connected. The topological relationship graph with completed attribute filling and weight assignment is output as the static structural feature map.

4. The method for determining fault information of a wind turbine generator set according to claim 3, characterized in that, Extracting fingerprint features related to time-varying trends from the dynamically running fingerprint sequence includes: The dynamic fingerprint sequence is decomposed by wavelet packet transform to obtain wavelet packet coefficients in different frequency bands; Calculate the energy value of the wavelet packet coefficients in each frequency band, and select the top few frequency bands with the largest energy values ​​as the dominant frequency bands; Extract the statistical features of the wavelet packet coefficients corresponding to the dominant frequency band, including mean, variance, skewness, and kurtosis; The identification of structural features related to the inherent state of components from static structural feature maps includes: Traverse all key component name index nodes in the static structural feature map and read their attribute labels; Calculate the centrality index of each key component name index node; Key component name index nodes with centrality indices higher than a preset centrality threshold are selected, and their attribute labels and centrality indices are combined as the structural features.

5. The method for determining fault information of a wind turbine generator set according to claim 4, characterized in that, The preliminary association between the extracted fingerprint features and the identified structural features to generate multidimensional feature pairs includes: A feature descriptor is created for each fingerprint feature, the feature descriptor containing feature type, source time series data identifier, and dominant frequency band information; A feature descriptor is created for each structural feature, and the feature descriptor includes the corresponding key component name, attribute type, and centrality index; Construct an association rule knowledge base, which defines the physical or logical association relationships between different types of fingerprint features and different types of structural features; The feature descriptors of fingerprint features and structural features are input into the association rule knowledge base for matching; If a match is successful, the matching fingerprint features and structural features are combined into a multidimensional feature pair, and the association rule on which the match was based is recorded.

6. The method for determining fault information of a wind turbine generator set according to claim 5, characterized in that, The correlation verification of the multidimensional feature pairs includes: Obtain actual fault records of the target wind turbine within a preset historical verification period; Examine the changes in each multidimensional feature to its associated fingerprint and structural features before and after the actual fault record occurred; If both fingerprint features and structural features exhibit abnormal changes beyond the normal range before the fault occurs, the correlation between the multidimensional feature pair is determined to be high. If only one feature shows an abnormal change, or if the changes in the fingerprint feature and structural feature are not significantly correlated with the time of the fault occurrence, then the correlation of the multidimensional feature pair is determined to be low. The multidimensional feature pairs with high correlation are retained to form the effective feature pair set.

7. The method for determining fault information of a wind turbine generator set according to claim 6, characterized in that, The process of tracing the state evolution of each valid feature pair in the set includes: Read the historical operation and maintenance log records, which include the time of each maintenance operation, the component to be operated, the operation type, and the post-operation status remarks; For the key components corresponding to the structural features in the effective feature pairs, search the historical operation and maintenance logs for all maintenance operation records involving the key components; The maintenance operation records found are arranged in chronological order to form the maintenance history timeline of the key components; The maintenance history timeline of the key components is superimposed and compared with the time change curve of the fingerprint features in the effective feature pair; On the overlay comparison image, the state of the fingerprint features at each maintenance operation is marked, thereby generating the enhanced feature unit with time context annotation, which includes the maintenance event time point and the corresponding maintenance type.

8. The method for determining fault information of a wind turbine generator set according to claim 7, characterized in that, The constructed feature evolution trajectory spanning multiple historical monitoring periods includes: Select a target historical monitoring period range; Extract all maintenance event time points that occurred within the target's historical monitoring period from the temporal context annotations of the enhanced feature units; Using the maintenance event time point as the dividing point, the target's historical monitoring period is divided into multiple continuous evolutionary stages; Within each evolutionary stage, the stage statistics of fingerprint features in the enhanced feature units are calculated, and the key attribute values ​​of the structural features at the end of each evolutionary stage are recorded. The feature evolution trajectory is formed by sequentially linking the stage statistics and key attribute values ​​of all evolution stages in chronological order. The feature evolution trajectory is presented in the form of a stage sequence.

9. The method for determining fault information of a wind turbine generator set according to claim 8, characterized in that, The step of modifying the feature evolution trajectory based on environmental factors, incorporating actual geographical information about the target wind turbine cluster, includes: The actual geographic environment information is obtained, including the altitude of the wind turbine location, the average annual temperature, the air density, the turbulence intensity level, and the salt spray corrosion level. An environmental factor impact comparison table is established, which defines the quantitative impact factors of different environmental factor values ​​on the aging rate and performance degradation coefficient of different components of wind turbines. From each evolutionary stage of the feature evolution trajectory, feature parameters related to the current environmental factors are extracted; Based on the environmental factor impact comparison table, find the quantitative impact factors corresponding to the extracted feature parameters; Using the found quantitative influencing factors, numerical compensation or conversion is performed on the stage statistics or key attribute values ​​of the corresponding evolution stages in the feature evolution trajectory to obtain the corrected feature evolution trajectory.

10. The method for determining fault information of a wind turbine generator set according to claim 9, characterized in that, The process of using the corrected feature evolution trajectory to deduce the potential fault modes of the wind turbine and the corresponding confidence levels includes: The stage statistics sequence and key attribute value sequence of each evolution stage are extracted from the corrected feature evolution trajectory to form the fault prediction input feature set; The fault prediction input feature set is input into a pre-trained long short-term memory neural network model, which is trained using historical fault data and corresponding feature evolution trajectories as samples. The time dependence of the input feature set is modeled by multiple memory units of the long short-term memory neural network model, and the hidden state sequence corresponding to each evolution stage is output. The final hidden state sequence is input into the fully connected layer for classification calculation to obtain the occurrence probability of various potential fault modes, including gearbox bearing wear, generator insulation deterioration, blade fatigue damage, tower foundation settlement, and yaw system jamming. The fault mode with the highest probability of occurrence is selected as the main prediction result, and the probability value is used as the confidence level corresponding to this fault mode.

Images