A multi-dimensional time sequence data anomaly feature identification method for a wind turbine

Abstract

The application relates to the technical field of electric digital data processing and pattern recognition, and discloses a multi-dimensional time sequence data abnormal feature identification method for a wind turbine, which comprises the following steps: acquiring a multi-source time sequence data stream of a controlled object state, constructing a feature flow track in a low-dimensional space based on a manifold learning algorithm; solving a topological acceleration component of the track relative to a logical time axis; checking the track evolution legality by using a physical inertia feature; and judging whether the curvature distortion is a non-physical continuous logical instantaneous jump or an abnormal feature. The application realizes logical smooth correction of pulse interference by using a physical motion continuity constraint, and suppresses early warning oscillation; the application captures an implicit degradation trend by extracting an orthogonal residual feature flow, establishes an early warning trigger before curvature fluctuation, and realizes logical decoupling of ontology abnormality and environmental disturbance in combination with space correlation.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing and pattern recognition technology, and in particular relates to a method for identifying anomaly features in multidimensional time-series data of wind turbine generators. Background Technology

[0002] Currently, using monitoring and data acquisition systems to collect multi-source time-series signals during unit operation and constructing feature flow maps is a common method for identifying the characteristics of high-dimensional nonlinear signals. High-dimensional data is projected onto a low-dimensional topological space through manifold learning algorithms, and the evolution logic of digital feature information is determined based on the fluctuation of geometric curvature in the space. However, the actual operating environment of the unit is subject to instantaneous switching on the grid side, lightning strike induction, and broadband pulse interference caused by non-stationary gusts. When processing such excitation signals, existing digital processing procedures often lead to non-physical logical jumps in the topological space due to the lack of deep coupling with the continuous laws of physical motion of the unit. In addition, traditional dimensionality reduction algorithms usually discard the residual components generated during the projection process as noise when pursuing global topological stability, which makes it impossible to effectively extract the early weak fatigue features contained in the data base.

[0003] Besides the physical limitations of the sensor sampling end caused by environmental interference, there are also shortcomings at the algorithm logic level. For example, Chinese invention patent CN120086759A discloses a wind turbine abnormal alarm method and system based on feature parameter recognition. It uses the SVDD model to identify abnormal points of temperature, power and vibration characteristics, and combines blade image similarity verification to expand the monitoring dimensions. However, the core is a pure data-driven probabilistic statistical model. The algorithm architecture does not establish a deep coupling of the physical essential attributes of the mechanical rotational inertia of the unit's transmission chain. The logic calculation procedure lacks the ability to verify the continuity of physical motion. When faced with signal jumps caused by electromagnetic pulses, it produces non-physical logical jumps, causing false alarms. The projection process treats the residual components as meaningless noise and removes them, masking the early weak fatigue characteristics in the data base, which cannot be extracted.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a topology evolution arbitration standard with mechanical and physical constraints, and how to collaboratively analyze the logical coherence between the main topology space and the orthogonal residual flow. Summary of the Invention

[0005] This invention provides a method for identifying anomaly features in multidimensional time-series data of wind turbine generators, comprising the following steps:

[0006] Step S1: Obtain multi-source time-series data streams representing the operating state of the controlled object;

[0007] Step S2: Based on the manifold learning algorithm, the multi-source time-series data stream is mapped to a low-dimensional topological manifold space to construct a feature flow trajectory that represents the state evolution trajectory of the controlled object.

[0008] Step S3: In the low-dimensional topological manifold space, the motion displacement vector of the feature flow trajectory on the logical time axis is calculated in real time, and the second derivative operation of the motion displacement vector with respect to the logical time axis is performed to generate the topological acceleration component characterizing the evolution trajectory change of the feature flow trajectory.

[0009] Step S4: Establish a logical evolution arbitration mechanism based on physical motion continuity constraints, take the topological acceleration component as logical input, and use the preset inertial constraint quantity that reflects the mechanical rotational inertia of the controlled object to verify the evolution legality of the feature flow trajectory.

[0010] Step S5: If the topological acceleration component exceeds the topological evolution boundary value corresponding to the physical motion law, it is determined that the local curvature distortion in the characteristic flow trajectory is a non-physical continuous logical instantaneous jump caused by environmental pulse interference. The topological inertial filtering process is performed on the characteristic flow trajectory, and the evolution coordinates of the characteristic flow trajectory are corrected by using the topological acceleration component to smooth the correction of environmental pulse interference.

[0011] Step S6: If the topological acceleration component does not exceed the topological evolution boundary value, the abnormal characteristics of the controlled object are identified by solving the local topological curvature of the characteristic flow trajectory in the low-dimensional topological manifold space, and an early warning signal containing the risk index and state drift component is output.

[0012] Preferably, step S2 includes: performing spatial dimension reconstruction processing on the multi-source time-series data stream, establishing motion consistency constraints of the feature stream in the low-dimensional topological manifold space using preset geometric measurement criteria; while monitoring the local geometric curvature change state in the low-dimensional topological manifold space, simultaneously calculating the topological deflection angular velocity of the feature stream trajectory relative to the preceding continuous trajectory, and using the topological deflection angular velocity as an auxiliary verification parameter when executing the logical evolution arbitration mechanism; wherein, the topological deflection angular velocity characterizes the degree of drastic directional evolution of the feature stream trajectory in the low-dimensional topological manifold space.

[0013] Preferably, step S2 further includes: during the mapping of multi-source time-series data to a low-dimensional topological manifold space, simultaneously extracting high-dimensional residual feature flows orthogonal to the main manifold space; and performing logical order coherence analysis on the high-dimensional residual feature flows in conjunction with a preset failure evolution model to extract residual coherence energy characterizing the latent degradation trend. Residual coherent energy satisfy: ,in, Let be the total number of orthogonal dimensions of the high-dimensional residual feature flow. For the high-dimensional residual characteristic flow in the first... The residual energy amplitude in each orthogonal dimension For the first The coherence weighting factor for the ... The feature contribution rate is preset for each dimension; based on the distribution of residual coherent energy, a warning trigger logic is established before the local topological curvature of the feature flow trajectory fluctuates.

[0014] Preferably, step S2 includes: establishing a logical mapping relationship between the neighborhood search radius and the local sample distribution density in the multidimensional feature space; adjusting the neighborhood search radius in real time according to the sparsity of the multidimensional feature space; and establishing topological associations between data nodes through a logical elastic resolution mechanism.

[0015] Preferably, the controlled object belongs to any controlled subject in the data source cluster. The method further includes: based on the field group topology logic arbitration mechanism, using spatial correlation to realize the logical decoupling of physical ontology anomalies and synchronous disturbances of multi-source objects; extracting the feature flow trajectories of multiple controlled objects in the data source cluster, and calculating the curvature distortion coordination factor of multiple controlled objects in the same time window; if the curvature distortion coordination factor exceeds the preset systemic environmental disturbance threshold, the current abnormal feature is judged as field group-level system noise, and the issuance of early warning instructions for individual controlled objects is suppressed.

[0016] Preferably, the topological inertial filtering process includes: establishing a topological inertial prediction model based on the evolution path of the feature flow trajectory in the preceding continuous period; and performing a deviation compensation operation on the prediction output of the topological inertial prediction model using the topological acceleration component to generate a logically smoothed synthetic feature flow.

[0017] Preferably, the multi-source time-series data stream includes physical quantity sampling signals with a frequency range of 0.1Hz to 50Hz, which are acquired from sensor arrays deployed on each monitoring node of the controlled object through a specific data interface.

[0018] Preferably, step S6 further includes: calculating the evolution velocity vector of the feature flow trajectory in the low-dimensional topological manifold space; when a local curvature distortion is detected, performing a logical continuity check by extracting the topological deflection angular velocity of the feature flow trajectory relative to the preceding continuous trajectory, and determining whether the curvature distortion has the continuous evolution attribute driven by physical motion.

[0019] Preferably, the topology evolution boundary value is jointly limited by the inertial constraint of the physical properties of the controlled object and the maximum nonlinear response rate under the current operating conditions.

[0020] Preferably, the risk index is generated by normalizing the instantaneous amplitude of the deviation of the feature flow trajectory from the physical motion continuity constraint; the state drift component represents the deviation weight distribution of the abnormal features in each dimension of the data subspace.

[0021] Compared with existing technologies, the multi-dimensional time-series data anomaly feature identification method for wind turbine generators of this invention has the following advantages:

[0022] 1. In the identification of anomaly features in multidimensional time-series data, this invention establishes a digital logic calculation and arbitration mechanism based on physical motion continuity constraints by real-time calculation of the motion displacement vector of feature points in low-dimensional topological space and calculation of the second derivative of the motion displacement vector with respect to the logical time axis to generate topological acceleration components. By using the rotational inertia constraint parameters mapped to the digital space to perform legality calculation and verification on the topological evolution trajectory in the virtual space, the high-frequency discontinuous pulse noise generated by instantaneous switching of the power grid or lightning strike is identified as non-physically continuous logical jumps during the calculation process, thus realizing logical smoothing correction of environmental pulse interference.

[0023] 2. By simultaneously extracting high-dimensional residual feature streams orthogonal to the mainstream shape space during the mapping process of multi-source time-series data to low-dimensional topological space, and performing logical order coherence analysis on the orthogonal residual feature streams in conjunction with a digital failure model, secondary feature extraction of information loss components generated by data dimensionality reduction algorithms is realized; the orthogonal complement set, which was originally regarded as data waste, is transformed into a signal source representing the implicit degradation trend, so that the weak fatigue characteristics in the early stage of evolution can be manifested through the coherence enhancement of residual energy distribution, thereby establishing the logic for early warning triggering before the curvature of the mainstream shape fluctuates.

[0024] 3. By establishing a negative correlation mapping relationship between the neighborhood search radius and the local sample density in the multidimensional feature space, and combining it with the logical coherence analysis of the evolution trajectory of the field-group topological features, this invention constructs a composite discrimination system with both operating condition adaptive capability and causal decoupling capability. The density compensation mechanism eliminates the topological collapse problem caused by uneven sample distribution, ensuring consistent logical resolution in different operating condition areas. At the same time, through differential arbitration of cross-station common mode features, the separation of field-group environmental disturbances from individual unit-specific degradation is achieved. Attached Figure Description

[0025] Figure 1 This is a flowchart of the feature flow evolution and anomaly identification process that introduces physical inertial constraints in this invention.

[0026] Figure 2 This is a diagram of the multi-source data manifold mapping and physical gating collaborative processing architecture of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0028] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal communication between two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0029] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0030] This invention provides a method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators. By constructing a feature flow trajectory within a low-dimensional topological manifold space and introducing physical motion continuity constraints based on mechanical rotational inertia, it achieves smooth correction of environmental pulse interference and monitoring of implicit degradation trends. The execution process includes: acquiring multi-source time-series data streams representing the operating state of the controlled object; mapping the data to a low-dimensional topological manifold space using a manifold learning algorithm to construct a feature flow trajectory; calculating the topological acceleration component of the feature flow trajectory relative to the logical time axis; verifying the evolution legitimacy of the feature flow trajectory using a preset inertial constraint; and performing feature flow trajectory correction or outputting a warning signal containing a risk index based on the verification result. During the data acquisition phase, the system acquires multi-source time-series data streams, including frequencies ranging from [frequency range missing in original text]. to The physical quantity sampling signals are converted into discrete digital sequences through a specific data interface. Due to the high-dimensional nonlinear characteristics of the wind turbine operating environment, the system uses a manifold learning algorithm to reconstruct the spatial dimension of the data. To address the problem of topological connection breaks caused by non-uniform data distribution, a logical mapping relationship is established between the neighborhood search radius and the local sample distribution density in the multidimensional feature space. The kernel density estimation method is used to calculate the local sample density of the current data packet in real time. And based on local sample density Dynamically adjust the neighborhood search radius Neighborhood search radius With local sample density A negative correlation exists, reducing the neighborhood search radius in areas of high data density. To improve local resolution and expand the neighborhood search radius in sparse data regions. To maintain trajectory continuity.

[0031] In a low-dimensional topological manifold space, the system establishes motion consistency constraints using preset geometric measurement criteria, generates characteristic flow trajectories representing the state evolution of the controlled object, and calculates the topological deflection angular velocity of the characteristic flow trajectory relative to the preceding continuous trajectory in parallel, using it as an auxiliary parameter for logical evolution arbitration. To suppress high-frequency discontinuous pulse interference caused by instantaneous grid switching or lightning strikes, the system calculates the motion displacement vector of the characteristic flow trajectory on the logical time axis in real time, performs second-order derivative operations with respect to the logical time axis, thereby generating topological acceleration components, and establishes a scaling operator. This scaling operator is used to map the angular acceleration limit of the transmission chain to the evolutionary boundary value of the topological space. The scaling factor is obtained by calculating the average projection ratio between the instantaneous acceleration vector and the corresponding characteristic flow trajectory displacement vector of the physical sensor reading sequence within a time window, using a scaling operator. This ensures that the mechanical rotational inertia constraints followed by the controlled object in physical space are transformed into geometrically meaningful motion displacement constraints within the topological manifold space. The scaling operator is used; a reference window is selected based on the rated operating state of the controlled object and the period without alarm triggering, and the physical angular acceleration of the main shaft of the transmission chain within the reference window is acquired synchronously. The displacement vector of the characteristic flow trajectory in the low-dimensional topological manifold space, The measured angular acceleration values ​​from the physical sensor are used to determine the measured topological acceleration components by taking the second derivative of the displacement vector with respect to the logical time axis. Fitting multiple groups and The slope obtained from the sampled data determines the scaling operator. The specific calibration procedure is as follows: During the initial no-load rotation phase after the unit is powered on, the control system collects the measured physical angular acceleration of the main shaft of the transmission chain every 20ms, and simultaneously calculates the second-order difference value of the displacement vector of the corresponding characteristic flow trajectory in the low-dimensional topological manifold space in the background; the sum of the absolute values ​​of 50 consecutive sets of second-order difference values ​​is divided by the sum of the absolute values ​​of the corresponding 50 sets of measured physical angular acceleration values, and the resulting arithmetic mean ratio is defined as the scaling operator, which is used as the unique conversion gain for mapping physical dimensions to digital topological space; the scaling operator is used to scale the ratio of the rated maximum torque to the moment of inertia in the mechanical design parameters, thereby calculating the topological evolution boundary value based on the topological displacement unit, and establishing a quantitative mapping of the mechanical rotational inertia constraint in physical space to low-dimensional topological space.

[0032] The system establishes a logical evolution arbitration mechanism based on physical motion continuity constraints. It uses topological acceleration components as logical input and verifies the legality of the evolution of the characteristic flow trajectory using preset inertial constraints reflecting the mechanical rotational inertia of the controlled object. Here, the topological evolution boundary value is jointly limited by the inertial constraints of the controlled object's physical properties and the maximum nonlinear response rate under the current operating conditions. If the topological acceleration component exceeds this topological evolution boundary value, it is determined that the local curvature distortion in the characteristic flow trajectory is a non-physically continuous logical jump. At this time, the system activates the topological inertial filtering program, establishes a topological inertial prediction model based on the evolution path in the preceding continuous cycle, and corrects the evolution coordinates using the topological acceleration components to achieve smooth correction of environmental pulse interference. To address the noise interference generated by the second-order derivative operation in digital sampling, a Savitzky-Golay filter is used to smooth the motion displacement vector of the characteristic flow trajectory before executing step S3. The filter window length is selected to match the sampling frequency, and the polynomial fitting order is set to [value missing]. to The smoothed displacement sequence is input into the second-order derivative calculation unit, and the topology evolution boundary values ​​are based on the rated torque of the transmission chain. With moment of inertia Ratio combined with scaling operator Calculations show that For the rated maximum torque, For the moment of inertia, the topological acceleration components are calculated in real time in continuous motion. If the boundary value is exceeded within a sampling period, topological inertial filtering is performed.

[0033] To address the early fatigue characteristics of wind turbine components, the system simultaneously extracts high-dimensional residual feature flows orthogonal to the main manifold space during the manifold mapping process. Combined with a pre-defined failure evolution model, logical order coherence analysis is performed on the high-dimensional residual feature flows to extract residual coherent energy characterizing latent degradation trends. Specifically, residual coherent energy The calculation formula is as follows: ,in, The residual coherent energy; The total number of orthogonal dimensions of the high-dimensional residual feature flow; For the high-dimensional residual characteristic flow in the first... Residual energy amplitude in each orthogonal dimension; For the first Coherence weighting factors for each dimension, coherence weighting factors According to the The feature contribution rate corresponding to each dimension is preset, and the system is based on the residual coherence energy. The distribution state is determined, and an early warning trigger is established before fluctuations occur in the local topological curvature of the characteristic flow trajectory; the failure evolution model consists of the state transition matrix representing the degradation trajectory of the component, and after extracting the high-dimensional residual characteristic flow, the energy amplitude of each dimension of the residual is determined. The vector and the state transition matrix are multiplied together to obtain the projection component of the residual signal in the preset fault evolution direction. For the first Residual energy amplitude in each orthogonal dimension, coherence weighting factor Based on the sensitivity of each orthogonal dimension to the fatigue damage characteristics of the component, the projected components of each dimension and their corresponding coherence weighting factors are summed. Product generates residual coherent energy ,when If the value deviates from the preset threshold of the healthy baseline distribution, it is determined that the controlled object has a hidden degradation trend and an early warning signal is output.

[0034] In field-level monitoring scenarios, the system achieves causal decoupling between physical anomalies and synchronous disturbances of multi-source objects through cross-site logical collaboration. The system acquires the characteristic flow trajectories of multiple controlled objects within the data source cluster and calculates the curvature distortion coordination factor of each controlled object within the same time window. If the curvature distortion coordination factor exceeds the preset systemic environmental disturbance threshold, the current abnormal feature is identified as field-level system noise, and warning commands for individual controlled objects are suppressed. If the topological acceleration component does not exceed the topological evolution boundary value, the system identifies the abnormal features of the controlled objects by solving the local topological curvature in the low-dimensional topological manifold space and outputs a warning signal containing a risk index and a state drift component. The risk index is generated by normalizing the instantaneous amplitude of the characteristic flow trajectory deviating from the physical motion continuity constraint, and the state drift component is used to characterize the deviation weight distribution of the abnormal features in each dimension of the data subspace, providing a quantitative basis for equipment maintenance.

[0035] Example 1: In the operation scenario of large-scale wind turbine units located in coastal areas with high thunderstorm risk, when a single controlled object faces high-frequency pulse interference caused by instantaneous switching on the grid side or lightning strike induction, and there are weak vibration signals caused by initial fatigue spalling inside its drive train main shaft, the multi-source time-series data stream collected by the sensor array exhibits a nonlinear superposition of high-amplitude random spikes and low signal-to-noise ratio evolution characteristics. To address this complex operating environment, the system acquires multi-source time-series data streams characterizing the operating state of the controlled object and uses the kernel density estimation method to calculate the local sample density in real time. By dynamically adjusting the neighborhood search radius Establish topological associations between data nodes, then use manifold learning algorithms to construct feature flow trajectories in a low-dimensional topological manifold space, and perform second-order derivative operations on the displacement vectors of the feature flow trajectories on the logical time axis to generate topological acceleration components.

[0036] During the execution of the identification logic, the signal jumps caused by electromagnetic pulses exhibit large changes in the second derivative mathematically. The system uses a preset inertial constraint reflecting the mechanical rotational inertia of the controlled object to perform evolutionary legality verification on the topological acceleration components. Trajectory abrupt changes exceeding the topological evolution boundary value are identified as non-physically continuous logical jumps. The system also uses a topological inertial prediction model to perform deviation compensation calculations on the evolutionary coordinates affected by environmental pulse interference, thereby transforming the originally mutually interfering noise filtering process into a logical compliance determination of physical inertial constraints. Simultaneously, the system extracts high-dimensional residual feature flows orthogonal to the mainstream shape space and performs logical order coherence analysis in conjunction with the failure evolution model, calculating the residual energy amplitude of each orthogonal dimension. Coherence weighting factor The sum of the products yields the residual coherent energy. ,in Let be the total number of orthogonal dimensions of the high-dimensional residual feature flow. For orthogonal dimension indexing, utilize the residual coherence energy The intensity distribution captures the latent degradation trend ignored by the mainstream projection process. After the above processing procedure, the system can suppress the alarm oscillation caused by the lightning pulse, extract the residual coherent energy signal that characterizes the early damage of the spindle, and generate an early warning command containing risk index and state drift component.

[0037] Example 2: In a digital verification scenario for monitoring the operating status of wind turbine units, the system uses a multi-source time-series data stream generated by a simulation platform with multi-physics coupling computation capabilities as the input source to verify the technical effectiveness of this method in suppressing environmental pulse interference and capturing weak degradation features. The data stream used in the experiment includes shaft vibration, generator speed, and grid-side voltage sampling signals, with a sampling frequency set to 50Hz. Gaussian white noise with a signal-to-noise ratio of 20dB and discontinuous pulse interference with an amplitude five times the signal reference value generated by simulated electromagnetic shock are injected into the original signal. To ensure the reproducibility of the experiment, the experimental platform performs nonlinear dynamic calculations driven by physical information, maintaining the calculation accuracy at [insert accuracy here]. Order of magnitude; during the experimental parameter calibration process, the system searches for the neighborhood radius. Dynamic optimization is performed, and its decision logic chain identifies local sample density. As the core influencing factor, the neighborhood search radius The setting is used to balance the analytical accuracy of local topology and the stability of global geometric connectivity. When the signal-to-noise ratio of the monitored signal decreases, leading to a decrease in local sample density... When the calculation result is in the low value range, in order to ensure that the geometric topology of the manifold space does not experience logical breaks due to sample sparsity, the neighborhood search radius is... Adjusting towards the upper limit of its value range increases the redundancy of topological associations, while adjusting towards the lower limit increases the local sample density. When the value is in the high range, the system reduces the neighborhood search radius. To avoid overly smoothed feature representations, under this decision rule, for a wind speed of 12 m / s, the local sample density is... The corresponding neighborhood search radius is 152 units. It was set to 0.25 and used as the initial condition for constructing the feature flow trajectory.

[0038] During the test run, the system maps the acquired multi-source time-series data stream containing noise to a low-dimensional topological manifold space. It calculates the topological deflection angular velocity of the feature flow trajectory relative to the preceding continuous trajectory and combines the topological acceleration component to determine the legality of the trajectory evolution. Table 1 records the processing results of the test group and the comparison sample group using only the conventional manifold algorithm under different noise intensities. Among them, the test group introduces a preset inertial constraint quantity reflecting the mechanical rotational inertia of the controlled object. When the instantaneous value of the topological acceleration component exceeds the evolution boundary value of 5.0, the system determines that the current local curvature distortion is a non-physical continuous logical instantaneous jump caused by environmental pulse interference and starts the topological inertial filtering program. It uses the evolution path in the preceding continuous period to establish a prediction model to correct the evolution coordinates. The data in Table 1 show that in the case of electromagnetic pulse interference 2, the recognition accuracy of the comparison sample group drops to 65.2%, while the test group maintains the recognition accuracy at 96.8% through logical smoothing correction.

[0039] Table 1: Comparison of Data Processing Results under Different Working Conditions

[0040]

[0041] Analyzing the data variation patterns in Table 1, within the signal-to-noise ratio (SNR) range above 20 dB, the recognition accuracy of the experimental group remained above 94.0%. However, when the experiment entered out-of-range conditions such as conditions 4 and 5, where the SNR was below 15 dB, the physical features in the original input signal were overwhelmed by noise, leading to a decrease in local sample density. The computational failure and sharp deterioration in recognition accuracy reflect that the geometric measure of the manifold space is limited by the physical boundary of the signal-to-noise ratio at the underlying signal level. To address the quantization of the state drift component in the warning signal, the system uses the Mahalanobis distance algorithm to determine the deviation projection of abnormal feature points relative to the local healthy cluster center within the multidimensional data subspace. By calculating the sensitivity weights of each orthogonal dimension to the current topological curvature distortion, the system obtains the state drift component representing the deterioration direction of the controlled object. Utilizing state drift components A backtracking mapping is achieved from the geometric features of the topological space to the anomaly weights of the physical monitoring nodes.

[0042] Example 3: This example combines Figures 1 to 2 This paper describes a method for identifying anomaly features in multidimensional time-series data of wind turbine generators. Figure 1As shown, step S1 executes multi-source time-series data stream acquisition to collect physical quantity signals of the controlled object from 0.1Hz to 50Hz. Step S2 constructs the feature flow trajectory and maps the manifold to a low-dimensional topological manifold space. Step S3 performs topological acceleration component calculation, that is, the second derivative of the displacement vector with respect to the logical time axis. Step S4 establishes logical evolution arbitration and introduces a preset inertial constraint as the logical input. Based on the physical motion continuity constraint, it is determined whether the topological evolution boundary value is exceeded. If it is, that is, the boundary is exceeded, then step S5 performs topological inertial filtering processing to determine that it is a non-physical continuous logical instantaneous jump and corrects the feature flow trajectory evolution coordinates to smooth and correct environmental pulse interference. If not, that is, it is not exceeded, then step S6 performs abnormal feature identification and calculates the local topological curvature. Finally, a multi-dimensional early warning signal containing risk index and state drift component is output.

[0043] like Figure 2 As shown, the input includes vibration sequence, rotational speed sequence, and electrical sequence. The vibration sequence is connected to the kernel density estimation operator, and the electrical sequence is connected to the variable radius search operator. The processed data from both are combined with the rotational speed sequence and fed into the low-dimensional mapping operator. After mapping, the data is split into the main path and enters the second derivative calculation unit and the physical inertia gating unit. The mechanical rotational inertia J is received as the constraint parameter input. When an over-limit trigger occurs, the trajectory smoothing correction operator is activated and a risk index is generated as output. The branch path is processed by the orthogonal spatial projection operator and the coherent energy aggregation operator, and finally outputs the state drift component. In addition, the trajectory smoothing correction operator and the state drift component are logically connected by dashed lines, indicating that there is data interaction or logical correlation between the two.

[0044] Example 4: In a wind farm cluster monitoring scenario located in a mountainous area with complex terrain and severe wind shear, the controlled object is affected by terrain-induced local rotating turbulence. Its multi-source time-series data stream often contains non-stationary fluctuations highly similar to mechanical damage characteristics. To identify the hidden degradation trend under this complex background, the system uses a central processing unit with double-precision floating-point arithmetic capabilities as the core computing unit, and sets the discretized sliding window length of the data sampling sequence to 1024 sampling points. This is to determine the parameters used to calculate the residual coherent energy. coherence weighting factor In the initial stage, the system executes a standardized parameter calibration procedure, which involves acquiring the high-dimensional residual characteristic flow of the controlled object under known normal operating conditions, and then using principal component analysis (PCA) to solve for the high-dimensional residual characteristic flow. Energy eigenvalues ​​in each orthogonal dimension Calculate the first Feature contribution rate of each orthogonal dimension The contribution rate of this feature For the first Energy eigenvalues The ratio of the coherence weighting factor to the sum of all energy eigenvalues ​​will be used for coherence weighting. Set as the feature contribution rate The proportional quantification values ​​are set as follows for the first to eighth orthogonal dimensions in this embodiment: the weight factor sequence is set to 0.25, 0.20, 0.15, 0.12, 0.10, 0.08, 0.06, and 0.04 based on the measured contribution rate distribution.

[0045] During the logical sequence coherence analysis, the failure evolution model retrieved by the system is a state transition matrix representing the component degradation trajectory. It defines the probability of the topological evolution direction from the normal operating state to the initial fatigue state. The system performs a dot product operation between the real-time extracted high-dimensional residual feature stream and the state transition matrix to obtain the projection amplitude of the residual signal in the failure evolution direction, and uses this as the residual energy amplitude. The input variables are determined using the aforementioned coherence weighting factors. The residual coherence energy is calculated by combining the formula. When the main bearing of the controlled object experiences slight spalling, the residual coherent energy The measured value jumped from the background level of 0.05 to 4.52. At this time, the displacement vector of the mainstream shape space did not show abrupt change due to mechanical inertia constraints. Early damage characteristics were extracted by enhancing the coherence of residual features. For the decoupling logic of field-level environmental disturbances, when the system identifies the curvature distortion of the characteristic flow trajectory of a single controlled object, it extracts the characteristic flow trajectories of five adjacent units within the same data source cluster and calculates the curvature distortion coordination factor. This is obtained by calculating the average Pearson correlation coefficient of the topological acceleration components of each unit within the same time window. If the calculated result of this coordination factor reaches 0.85 or above, it is determined that the current fluctuation is a systemic noise caused by large-scale shear wind. The system then outputs an environmental disturbance flag and suppresses the triggering of the warning. If the coordination factor is lower than 0.30 and the residual coherence energy of the target unit is low, the system will not trigger the warning. If the threshold is continuously exceeded, a warning signal containing state drift components is output. The entire recognition process uses a coherence weighting factor. Standardized calibration and quantitative arbitration of field group coordination factors reduce the uncertainty of parameter setting. Data processing results show that, under the condition of a sliding window length of 1024 points and coherent weight alignment feature contribution rate, the system improves the sensitivity of mechanical degradation features identification under complex mountain conditions by 12.5%, realizing the extraction of latent anomaly features within the framework of physical inertial constraints.

[0046] Example 5: In a unit deployment scenario with a specific transmission chain structure, to establish a physical inertial constraint benchmark for verifying the compliance of the execution trajectory, the system collects the mechanical design parameters of the controlled object, including the transmission chain moment of inertia. and rated maximum torque The maximum angular acceleration limit of the controlled object under rated operating conditions is calculated based on the dynamic control equation. Multi-source time-series data streams are obtained by running the controlled object under load-cutting conditions, and the initial trajectory in the low-dimensional topological manifold space is constructed. The measured values ​​of the topological acceleration components of the initial trajectory under load-cutting conditions are calculated. The benchmark quantization index of the preset inertial constraint quantity is determined by the peak envelope of the measured values, and the topological evolution boundary value is filled by the benchmark quantization index.

[0047] When handling the task of identifying abnormal features for a specific type of generator unit, the state transition matrix in the failure evolution model is configured with parameters through a data program. Discrete sampling sequences of the generator unit within its historical operating cycle are obtained, and a manifold learning algorithm is used to extract feature flow trajectories under different states. The state transition probability from normal operating state to initial fatigue state is then calculated. State transition probability The numerical distribution was obtained by statistically analyzing the migration frequency of feature points across topological subspaces within a time window, combined with the measured residual coherence energy. The weight allocation coefficients in the state transition matrix are corrected, and the normalized probability sequence is written into the storage address space to generate a reference model representing the degradation trajectory of the component.

[0048] Example 6: In the winter operation scenario of a high-altitude mountain wind farm, the controlled object experiences a decrease in the rotational inertia of the transmission chain due to non-uniform ice accumulation on the blade surface. An asymmetric offset occurs, a change in physical state that causes a deviation between the preset inertial constraints and the actual mechanical evolution. The system executes a pre-deployment calibration procedure to correct the topology manifold mapping. This procedure collects raw signals from acceleration and torque sensors under no-load rotation of the unit to establish a reference sample set. Using the pseudo nearest neighbor method (FNN algorithm), the embedding dimension of the multi-source time-series data stream in the reconstruction process is calculated. The minimum integer value of 4, corresponding to a feature loss rate of less than 5%, is set as the mapping dimension of the target topological space.

[0049] After determining the mapping dimension, the system determines the neighborhood search radius through an offline iterative optimization program. The program selects a gradient step size of 0.1 to 0.5 within the parameter search space and calculates the connectivity index of the topological manifold structure under different radii. It uses the second-order variance of the local geodesic distance as the convergence criterion. When the variance value is below 0.01 for three consecutive iterations, the corresponding radius value of 0.28 is determined as the neighborhood search radius of the controlled object under icing conditions. Simultaneously, the calibrated radius is used... The topological acceleration components are recalculated and projected onto the state transition matrix of the failure evolution model, utilizing the measured residual coherence energy. The weight distribution was verified, and the topology deviation caused by the initial operating condition offset was corrected, which improved the accuracy of identifying abnormal blade mass distribution under icing conditions from 72.5% before calibration to 95.8%.

[0050] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for identifying anomaly features in multidimensional time-series data of wind turbine generators, characterized in that, Includes the following steps: Step S1: Acquire a multi-source time-series data stream characterizing the operating state of the controlled object; the multi-source time-series data stream includes physical quantity sampling signals with a frequency range of 0.1Hz to 50Hz, and the physical quantity sampling signals include shaft vibration, generator speed and grid-side voltage sampling signals. The physical quantity sampling signals are acquired from the sensor array deployed on each monitoring node of the controlled object through the data interface. Step S2: Based on the manifold learning algorithm, the multi-source time-series data stream is mapped to a low-dimensional topological manifold space to construct a feature flow trajectory that represents the state evolution trajectory of the controlled object. Step S3: In the low-dimensional topological manifold space, the motion displacement vector of the feature flow trajectory on the logical time axis is calculated in real time, and the second derivative operation of the motion displacement vector with respect to the logical time axis is performed to generate the topological acceleration component characterizing the evolution trajectory change of the feature flow trajectory. Step S4: Establish a logical evolution arbitration mechanism based on physical motion continuity constraints, take the topological acceleration component as logical input, and use the preset inertial constraint quantity that reflects the mechanical rotational inertia of the controlled object to verify the evolution legality of the feature flow trajectory. Step S5: If the topological acceleration component exceeds the topological evolution boundary value corresponding to the physical motion law, then the local curvature distortion in the characteristic flow trajectory is determined to be a non-physical continuous logical transient caused by environmental pulse interference. Topological inertial filtering is then performed on the characteristic flow trajectory. This topological inertial filtering includes: establishing a topological inertial prediction model based on the evolution path of the characteristic flow trajectory within the preceding continuous period; performing deviation compensation calculations on the prediction output of the topological inertial prediction model using the topological acceleration component to generate a logically smoothed synthetic characteristic flow; and correcting the evolution coordinates of the characteristic flow trajectory using the topological acceleration component to smooth and correct environmental pulse interference. The topological evolution boundary value is jointly limited by the inertial constraint of the controlled object's physical properties and the maximum nonlinear response rate under the current operating condition, and the topological evolution boundary value is based on the rated torque of the transmission chain. With moment of inertia Ratio combined with scaling operator Calculations show that For the rated maximum torque, For rotational inertia, the scaling operator Used to map the angular acceleration limit of the transmission chain to the evolution boundary value of the topological space; Step S6: If the topological acceleration component does not exceed the topological evolution boundary value, the abnormal characteristics of the controlled object are identified by solving the local topological curvature of the characteristic flow trajectory in the low-dimensional topological manifold space, and an early warning signal containing the risk index and state drift component is output.

2. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, Step S2 includes: performing spatial dimension reconstruction processing on the multi-source time-series data stream, establishing motion consistency constraints of the feature stream in the low-dimensional topological manifold space using preset geometric measurement criteria; while monitoring the local geometric curvature change state in the low-dimensional topological manifold space, simultaneously calculating the topological deflection angular velocity of the feature stream trajectory relative to the preceding continuous trajectory, and using the topological deflection angular velocity as an auxiliary verification parameter when executing the logical evolution arbitration mechanism; wherein, the topological deflection angular velocity characterizes the degree of drastic directional evolution of the feature stream trajectory in the low-dimensional topological manifold space.

3. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, Step S2 further includes: during the mapping of multi-source time-series data streams to a low-dimensional topological manifold space, simultaneously extracting high-dimensional residual feature streams orthogonal to the main manifold space; and performing logical order coherence analysis on the high-dimensional residual feature streams in conjunction with a preset failure evolution model, wherein the logical order coherence analysis includes analyzing the residual energy amplitude of each dimension. The vector and state transition matrix are multiplied by a dot product to obtain the projection component of the residual signal along the preset fault evolution direction; this is used to extract the residual coherent energy that characterizes the latent degradation trend. Residual coherent energy satisfy: ,in, Let be the total number of orthogonal dimensions of the high-dimensional residual feature flow. For the high-dimensional residual characteristic flow in the first... The residual energy amplitude in each orthogonal dimension For the first The coherence weighting factor for the ... The feature contribution rate is preset for each dimension; based on the distribution of residual coherent energy, a warning trigger logic is established before the local topological curvature of the feature flow trajectory fluctuates.

4. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, Step S2 includes: establishing a logical mapping relationship between the neighborhood search radius and the local sample distribution density in the multidimensional feature space; adjusting the neighborhood search radius in real time based on the sparsity of the multidimensional feature space; and establishing topological relationships between data nodes through a logical elastic resolution mechanism. The logical elastic resolution mechanism includes using a kernel density estimation method to calculate the local sample density of the current data packet in real time. And based on local sample density Dynamically adjust the neighborhood search radius Neighborhood search radius With local sample density There is a negative correlation, which reduces the neighborhood search radius in areas with high data density. To improve local resolution and expand the neighborhood search radius in sparse data regions. To maintain trajectory continuity.

5. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, The controlled object belongs to any controlled subject in the data source cluster. The method also includes: based on the field group topology logic arbitration mechanism, using spatial correlation to realize the logical decoupling of physical ontology anomalies and synchronous disturbances of multi-source objects; extracting the feature flow trajectories of multiple controlled objects in the data source cluster, and calculating the curvature distortion coordination factor of multiple controlled objects in the same time window; if the curvature distortion coordination factor exceeds the preset systemic environmental disturbance threshold, the current abnormal feature is judged as field group-level system noise, and the issuance of early warning instructions for individual controlled objects is suppressed.

6. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, Step S6 further includes: calculating the evolution velocity vector of the feature flow trajectory in the low-dimensional topological manifold space; when a local curvature distortion is detected, performing a logical continuity check by extracting the topological deflection angular velocity of the feature flow trajectory relative to the preceding continuous trajectory, and determining whether the curvature distortion has the continuous evolution attribute driven by physical motion.

7. The method for identifying anomaly features in multi-dimensional time-series data of wind turbine generators according to claim 1, characterized in that, The risk index is generated by normalizing the instantaneous amplitude of the deviation of the feature flow trajectory from the physical motion continuity constraint; the state drift component represents the deviation weight distribution of the abnormal features in each dimension of the data subspace.

Images