A method and device for intelligent operation and maintenance of a variable pitch system based on edge cloud layered feature mapping and semantic reasoning

Abstract

The application relates to the technical field of intelligent operation and maintenance of wind turbine generators, and discloses a variable pitch system intelligent operation and maintenance method and device based on edge cloud layered feature mapping and semantic reasoning, which comprises the following steps: performing parameter analysis and state gating on collected data of a wind turbine generator and a variable pitch system, and performing high-frequency data cleaning on released analog quantities; performing feature derivation based on a high-frequency physical mechanism and data frequency reduction based on time window statistics to generate a structured frequency-reduced data frame; assembling a high-dimensional input feature tensor and inputting the same into an anomaly detection model to obtain a comprehensive anomaly probability vector, so that an anomaly classification result is determined; performing feature contribution degree attribution on the high-dimensional input feature tensor to determine core contribution features, so that a structured metadata is generated; and generating a device anomaly operation and maintenance report of the variable pitch system based on an analysis result of the structured metadata. The application can improve the early micro-degradation perception capability of the variable pitch system.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for wind turbine generator sets, and more specifically, to an intelligent operation and maintenance method and device for pitch systems based on edge-cloud hierarchical feature mapping and semantic reasoning. Background Technology

[0002] With the continuous increase in the single-unit capacity of wind turbine generators and the ongoing expansion of wind farm cluster operations, the operational reliability of the pitch system has an increasingly prominent impact on the generator's power generation efficiency and operational safety. The pitch system achieves power control by adjusting the blade pitch angle and performs feathering braking under abnormal operating conditions. Internally, it includes a pitch driver, pitch motor, angle encoder, mechanical transmission mechanism, backup power supply, and communication link, characterized by the coupling of multiple electromechanical and thermal processes. The wind turbine's data acquisition and monitoring control system continuously collects operational data such as wind speed, blade angle, motor current, bus voltage, equipment temperature, status words, and fault codes, providing a data foundation for pitch system condition monitoring. Based on this, how to identify equipment anomalies from continuous, high-frequency, and complex operational data, and transform the diagnostic results into maintenance information that field personnel can understand and execute, has become a crucial issue in the operation and maintenance of pitch systems.

[0003] Current operation and maintenance diagnostics for pitch control systems primarily employ either alarm threshold-based monitoring or centralized diagnostic methods based on machine learning models. Alarm threshold-based monitoring typically sets alarm thresholds for individual operating parameters, outputting corresponding fault codes or alarm messages when these limits are exceeded. Centralized diagnostic methods based on machine learning models transmit the operating data collected from the wind turbine to a central server, where a classification model identifies the equipment's operating status. Some operation and maintenance systems also generate fault descriptions and maintenance suggestions based on fault codes, model outputs, or input operating data to assist on-site personnel in troubleshooting.

[0004] However, existing pitch system operation and maintenance technologies still have several problems: First, high-frequency data under different operating conditions often use the same processing procedures. Invalid data generated during feathering shutdowns, system initialization, and manual maintenance still consume computing resources and may interfere with subsequent diagnostic results. Therefore, the equal-interval sampling method used to reduce data transmission volume easily misses short-duration current surges, voltage fluctuations, and angle deviations during pitch operations, causing critical anomaly information to be lost before transmission. Second, existing diagnostic models mostly use raw sensor parameters as input, which is insufficient for identifying early minor degradation of the equipment. When normal samples far outnumber abnormal samples and multiple components may experience concurrent anomalies, conventional classification models tend to favor normal states, resulting in missed detections of high-risk anomalies. Third, existing model outputs usually remain at the level of fault category or anomaly probability, making it difficult to form clear troubleshooting criteria. Furthermore, directly inputting continuous, high-frequency, and unordered physical time-series data into large-scale language models can lead to excessive input data volume, difficulty in extracting effective information, and inference results deviating from the true state of the equipment, making it difficult to support intelligent operation and maintenance and accurate diagnosis of pitch systems.

[0005] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes an intelligent operation and maintenance method and device for pitch systems based on edge-cloud hierarchical feature mapping and semantic reasoning, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] According to one aspect of the present invention, an intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning is provided. The method includes: performing parameter parsing and state gating on the collected data of the wind turbine and pitch system, and performing high-frequency data cleaning on the released analog quantities; using the released analog quantities for feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning to generate structured down-frequency data frames; parsing the structured down-frequency data frames and assembling a high-dimensional input feature tensor, inputting the high-dimensional input feature tensor into an anomaly detection model to obtain a comprehensive anomaly probability vector; using the comprehensive anomaly probability vector to determine the anomaly classification result by combining it with the corresponding adaptive dynamic safety threshold; determining the comprehensive anomaly probability vector where the anomaly probability exceeds the corresponding adaptive dynamic safety threshold, and attributing the feature contribution of the high-dimensional input feature tensor to determine the core contributing features; using the core contributing features to backtrack historical down-frequency physical statistics to generate structured metadata; and generating an equipment anomaly operation and maintenance report for the pitch system based on the parsing results of the structured metadata.

[0009] According to another aspect of the present invention, an intelligent operation and maintenance device for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning is also provided. The device includes: an edge feature module for performing parameter parsing and state gating on the collected data of the wind turbine and pitch system, and performing high-frequency data cleaning on the released analog quantities; the released analog quantities are used for feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning to generate structured down-frequency data frames; a cloud diagnostic module for parsing the structured down-frequency data frames and assembling a high-dimensional input feature tensor, inputting the high-dimensional input feature tensor into an anomaly detection model to obtain a comprehensive anomaly probability vector; the comprehensive anomaly probability vector is used to determine the anomaly classification result by combining it with the corresponding adaptive dynamic safety threshold; a pitch attribution module for determining the comprehensive anomaly probability vector where the anomaly probability exceeds the corresponding adaptive dynamic safety threshold, and performing feature contribution attribution on the high-dimensional input feature tensor to determine the core contributing features; the core contributing features are used to backtrack historical down-frequency physical statistics to generate structured metadata; and a semantic operation and maintenance module for generating an equipment anomaly operation and maintenance report for the pitch system based on the parsing results of the structured metadata.

[0010] The beneficial effects of this invention are as follows:

[0011] (1) This invention addresses the operational characteristics of pitch systems, which include the synchronous operation of three blades, position loop tracking, motor drive, and heat load accumulation. It implements data gating on the edge computing device side based on the operating status, isolates invalid data under conditions of feathering shutdown, system initialization, and manual maintenance, and cleans and derives physical characteristics of blade angle, pitch motor current, DC bus voltage, encoder angle, and equipment temperature under effective operating conditions. This enables the three-blade synchronization deviation, position tracking error, current mutation, and heat load accumulation to enter the anomaly classification process, which is beneficial to improving the ability to perceive early minor degradation of the pitch system.

[0012] (2) The present invention performs time window statistics on the high-frequency analog quantities and physical derivative characteristics of the pitch system on the edge computing device side, and converts the millisecond-level time series data into structured down-frequency data frames containing mean, maximum, minimum and variance. While reducing the amount of data transmitted by the edge cloud and the computing burden of the central server, it retains the current surge, bus voltage fluctuation, blade angle deviation and other short-term abnormal characteristics that occur during the pitch operation, and reduces the loss of key fault information due to sampling down-frequency.

[0013] (3) In view of the strong coupling, possible concurrent anomalies and scarce fault samples among the pitch drive, pitch motor, backup power supply, angle encoder, mechanical transmission mechanism and soft communication link, this invention adopts independent multi-label anomaly classifiers to output the occurrence probability of each anomaly category, and increases the penalty weight corresponding to the high-risk anomaly underreporting in the model training, and adjusts the working condition adaptive dynamic safety threshold in combination with wind condition changes, thereby improving the identification ability of a few types of anomalies and multi-component concurrent anomalies, and taking into account the suppression of false alarms and the control of high-risk fault underreporting under complex wind conditions.

[0014] (4) When the anomaly probability of the pitch system exceeds the adaptive dynamic safety threshold, this invention performs feature contribution attribution on the anomaly classification results and traces back the historical physical value sequence of the core contribution features to determine their changing trends. Thus, the anomaly category, anomaly probability, core contribution features, and physical feature changing trends are converted into structured metadata. The large language model does not need to directly process massive continuous time-series numerical values. It can combine the historical maintenance work orders, fault manuals, and equipment manuals of the pitch system to generate maintenance and troubleshooting guidance reports with physical evidence chains, thereby forming a closed loop of intelligent operation and maintenance of the pitch system from underlying high-frequency signal perception, anomaly identification, fault attribution to operation and maintenance decision output. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0017] Figure 2 This is a flowchart of data acquisition and preprocessing for an edge computing device according to an embodiment of the present invention;

[0018] Figure 3 This is a flowchart of anomaly detection and online classification reasoning according to an embodiment of the present invention;

[0019] Figure 4 This is a flowchart of the structured metadata generation process according to an embodiment of the present invention;

[0020] Figure 5 This is a flowchart of semantic reasoning and equipment anomaly maintenance report generation according to an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0022] According to an embodiment of the present invention, an intelligent operation and maintenance method and apparatus for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning is provided.

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, an intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning is provided. This method includes: Step S1, performing parameter parsing and state gating on the collected data of the wind turbine and pitch system, and performing high-frequency data cleaning on the analog quantities released after state gating; the released analog quantities are used for feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning to generate structured down-frequency data frames; Step S2, parsing the structured down-frequency data frames and assembling high-frequency data... A high-dimensional input feature tensor is input into the anomaly detection model to obtain a comprehensive anomaly probability vector. The comprehensive anomaly probability vector is used to determine the anomaly classification result by combining the corresponding adaptive dynamic safety threshold. Step S3: Determine the comprehensive anomaly probability vector where the anomaly probability exceeds the corresponding adaptive dynamic safety threshold, and attribute the feature contribution of the high-dimensional input feature tensor to determine the core contributing features. The core contributing features are used to backtrack historical down-frequency physical statistics to generate structured metadata. Step S4: Based on the parsing results of the structured metadata, generate an equipment anomaly operation and maintenance report for the pitch system.

[0024] Specifically, this invention adopts a layered architecture of edge-cloud collaboration, which aims to take into account the real-time acquisition and preprocessing calculation of massive high-frequency SCADA time-series physical quantities of the underlying wind turbine units, as well as the high-dimensional feature calculation of the high-level complex model and the high-level semantic reasoning of the large language model (LLM), thereby constructing a complete technical closed loop from data perception to intelligent operation and maintenance suggestions.

[0025] It should be noted that the core of this invention lies in the fact that Generative Large Models (LLMs) excel at discrete text reasoning. Directly inputting massive amounts of continuous time-series data may lead to serious context overruns and model illusions. This invention proposes dynamic temporal attribution and structured semantic mapping. Through LightGBM anomaly classification and SHAP extraction of core contributing features that trigger anomalies, it further extracts the trends of physical quantity changes to ultimately generate high-dimensional structured semantic feature descriptions. SCADA sliding window statistics and physical feature derivation calculations are only used as input preprocessing for the LightGBM anomaly classification step, aiming to reduce data frequency and avoid context overruns. Simultaneously, it incorporates physical priors, elevating the underlying massive data to the level of physical quantity change mechanisms, extracting more information to improve the classification accuracy of the model. The ultimate goal is to address the modal differences in feature forms and data volume between the underlying physical signal perception and the top-level semantic reasoning, extracting high-dimensional semantic structure information for LLM to better perform reasoning and generation.

[0026] It should be added that, unlike existing technologies that input continuous high-frequency SCADA feature data into LLM, which easily leads to context overload and logical illusion, and directly use large language models as numerical classification models, this invention transforms the underlying high-frequency, low-dimensional, massive data into structured high-dimensional semantic features. The final input to the large model is JSON structured text. The core is to solve the problem of mapping high-dimensional, disordered underlying physical time-series signals into structured high-dimensional features across modalities. In addition, the numerical classification work of this invention is completed by the underlying LightGBM, and then further dynamic temporal attribution is performed by the SHAP interpreter.

[0027] Specifically, the multidimensional feature tensor in this invention, on the one hand, calculates feature physical quantities in high-frequency data to provide a more reliable basis for anomaly classification, and on the other hand, performs statistical calculations through a sliding window to reduce the frequency and aggregate feature quantities within the time window. In the prior art, the actual cost of fine-tuning large models is high, requiring computing power and extensive preprocessing of knowledge mechanisms. This invention, however, before the large model generates an answer, retrieves knowledge related to the structured semantic conclusions obtained in the previous process through RAG retrieval, and fully utilizes the semantic reasoning capabilities of LLM to generate operation and maintenance suggestions.

[0028] Furthermore, this invention explicitly embeds prior physical knowledge of wind power, such as kinematics and thermodynamics, into the feature space based on native SCADA data. The aim is to integrate physical priors to propose high-dimensional physical characteristics and increase the amount of information. On this basis, by introducing TreeSHAP, this invention can trace back which features in the input feature set can more strongly prove the existence of a certain anomaly, thereby further integrating this information into structured semantic features, enabling large models to perform semantic reasoning better.

[0029] It should also be noted that in this invention, the edge computing device serves as a data sensing anchor point, deployed inside the wind turbine tower base control cabinet or nacelle. This device directly connects to the wind turbine's main control SCADA (Supervisory Control And Data Acquisition) data source via industrial fieldbuses (such as OPC UA, Modbus TCP, EtherCAT, and other standard communication protocols), possessing the significant characteristics of direct high-frequency data acquisition and front-end computing power. This layer focuses on the characteristic sensing and preprocessing of underlying physical signals. It primarily undertakes real-time cleaning and conversion of high-frequency signals, as well as extreme bandwidth optimization tasks for cloud transmission. Core functional modules include: data acquisition and status gating, frequency reduction aggregation, and structured packaging.

[0030] It should be noted that in this invention, the central server primarily undertakes semantic-level feature mapping and reasoning, deployed in the wind farm's central control room. This node is equipped with abundant heterogeneous computing resources (such as GPU / NPU computing cards) and locally deploys a Large Language Model (LLM) and a wind power historical operation and maintenance knowledge base based on Retrieval Augmentation (RAG). This layer focuses on high-level semantic processing and reasoning generation of data, abstract multi-dimensional feature tensor reasoning computation, and then the complex core computational task of cross-modal natural language semantic reasoning generation. Core functional modules include: real-time anomaly detection, dynamic SHAP feature attribution and trend extraction, LLM semantic reasoning, and intelligent operation and maintenance report output.

[0031] In one embodiment, the data collected by the wind turbine and pitch system is acquired by an edge computing device, including high-frequency real-time SCADA data streams acquired by the wind turbine main control system and the pitch subsystem. Parameter parsing and state gating are performed, including: unpacking the original messages of the high-frequency real-time SCADA data stream, and dividing the unpacked parameters into static attribute metadata, digital state quantities, and high-frequency analog quantities according to the physical attributes and update frequency of the variables; wherein, the static attribute metadata is used as device tag information, the digital state quantities are used as the basis for dynamic gating, and the high-frequency analog quantities are written into a circular buffer; based on the digital state quantities, the wind turbine operating state is identified; when the wind turbine operating state is generator brake engaged, system initialization, feathering shutdown lock, or manual maintenance mode, a lockout command is output, the circular buffer is cleared, and subsequent processing is paused; when the wind turbine operating state is grid-connected power generation, yaw seeking wind, or dynamic pitch control, a conduction command is output to perform high-frequency data cleaning on the analog quantities released by the conduction command.

[0032] In one embodiment, high-frequency data cleaning of the released analog quantity includes: calculating the window mean and window standard deviation of the current parameter within a time sliding window containing a preset number of sampling points; setting the current verification point whose deviation from the window mean is greater than a preset multiple of the window standard deviation to a null value; retrieving the two most recent valid data points before and after the null value or the location of the communication packet loss, and performing linear interpolation according to the time ratio; stopping the linear interpolation when the continuous missing time exceeds a limit threshold, and outputting sensor disconnection or communication abnormality; and performing a first-order exponential moving average on the interpolated high-frequency analog quantity to obtain the cleaned released analog quantity.

[0033] It should be added that the expressions for the window mean, window standard deviation, linear interpolation value, and first-order exponential moving average are as follows: ; ; ; In the formula, The number of sampling points within the time sliding window For the first Values ​​of each sampling point Window mean The standard deviation of the window; for The input value after linear interpolation at time 1; The time to be interpolated and Timestamps of two consecutive valid data points and For the corresponding valid sample values for The first-order exponential moving average at time t; This is the smoothing coefficient.

[0034] In one embodiment, feature derivation based on high-frequency physical mechanisms includes: within a high-frequency time sliding window, obtaining the three-blade angle range based on the actual angles of the three blades, and calculating the pitch angle synchronization deviation rate based on the three-blade angle range and the average actual angle of the three blades; obtaining the position tracking error integral based on the pitch angle setpoint and the actual angle, obtaining the current dynamic change rate based on the pitch motor current at adjacent sampling times, and determining the motor transient thermal load accumulation based on the pitch motor current within the high-frequency time sliding window; solving the bus voltage fluctuation standard deviation using the DC bus voltage within the high-frequency time sliding window, and determining the encoder redundancy dispersion based on the actual angle of the main encoder and the redundant encoder angle; and constructing physically derived features based on the three-blade angle range, pitch angle synchronization deviation rate, position tracking error integral, current dynamic change rate, motor transient thermal load accumulation, bus voltage fluctuation standard deviation, and encoder redundancy dispersion.

[0035] In one embodiment, data down-conversion based on time window statistics includes: within the down-conversion time window, calculating the mean, maximum, minimum, and variance of the analog quantity and physical derived features released after high-frequency data cleaning, respectively, to obtain down-conversion analog quantity statistical features and physical derived statistical features; pressing the down-conversion analog quantity statistical features and physical derived statistical features into a time-series feature vector according to the feature index order; concatenating digital state quantities to the end of the time-series feature vector, and appending static attribute metadata and timestamps to the message header to encapsulate a structured down-conversion data frame.

[0036] Step 1: Data Acquisition and Preprocessing

[0037] The data acquisition and preprocessing process of edge computing devices is as follows: Figure 2 As shown, the specific process includes the following steps.

[0038] 1.1 Explanation of SCADA core parameters:

[0039] The edge computing device in this invention will collect SCADA data from the wind turbine main control system and pitch subsystem in real time, with a sampling period of up to 20ms. To comprehensively represent the dynamic characteristics of the pitch system, the feature parameters collected from the SCADA system include multiple dimensions such as meteorological environment, electrical drive, mechanical feedback, and logic control. The specific parameter list and its physical meaning in the anomaly detection model are shown in Table 1; where, blade 1 / 2 / 3 represents blade 1, blade 2, and blade 3, such as (blade 1 / 2 / 3) redundant encoder angle refers to the redundant encoder angle of blade 1, blade 2, and blade 3.

[0040] Table 1. Specific parameters and their physical meaning in the anomaly detection model.

[0041]

[0042] 1.2 Parameter Analysis and State Gating:

[0043] After receiving the high-frequency real-time SCADA data stream from the wind turbine and pitch system, the edge computing device, in order to avoid invalid operating condition data consuming computing resources and interfering with the classification accuracy of the subsequent LightGBM model, first performs state gating and digital quantity parsing steps. The specific technical implementation logic is as follows:

[0044] 1) Message decomposition and parameter classification:

[0045] The edge-side parsing engine unpacks the collected raw messages and, based on the physical attributes and update frequency of the variables, divides them into three subsets for differentiated processing: ① Static attribute metadata: Data such as version number, motor rated frequency, motor rated speed, and gear ratio are non-time-series features and do not participate in high-frequency dynamic calculations. The edge side uses this as tag information, appending it to the data frame header only when the state changes or when low-frequency data is packaged, to reduce transmission load. ② Discrete digital state quantities: Parameters such as generator braking mode, control words, status words, and fault codes are discretely represented. Their values ​​indicate the current state sequence number, such as: 0-alarm, 10-normal shutdown, 30-rapid shutdown, etc. These discrete parameters are categorized into a digital state quantity set for subsequent dynamic gating. ③ High-frequency analog data: The remaining continuous time-series parameters (such as current wind speed / direction, actual angle, DC bus voltage, motor current, temperature, and grid-side parameters, etc.) are pushed into the edge-side data buffer, awaiting gating instructions to determine whether to execute subsequent feature derivation calculations.

[0046] 2) Dynamic gating based on state machines:

[0047] Utilizing the extracted digital state quantities, this invention constructs a dynamic gating trigger logic to achieve on-demand allocation of edge computing power and isolation of invalid data: ① Sleep and Truncation Logic: Real-time monitoring of unit operating status. If the unit is identified as being in generator brake engaged state, system initialization state, feathering shutdown lockout, or manual maintenance mode, the gating trigger outputs a lockout command. At this time, the edge computing device automatically clears the high-frequency data in the current circular buffer and suspends subsequent processing steps. ② Activation and Conduction Logic: Only when the status word and control word indicate that the unit is in grid-connected power generation, yaw seeking wind, or active dynamic pitch control conditions, the gating trigger outputs a conduction command, entering the next level data cleaning and feature derivation module.

[0048] 1.3 High-frequency data cleaning:

[0049] To eliminate SCADA data anomalies caused by sensor interference and communication packet loss, thereby improving the reliability of subsequent feature derivation calculations and system anomaly detection, this system preprocesses and cleans the high-frequency data stream according to the following three steps:

[0050] 1) Outlier removal based on sliding window: For abrupt outliers caused by electromagnetic interference in the sensor, the Laida criterion is used ( (Criteria) Combined with a time-based sliding window, local detection and elimination are performed. Within a sliding window containing N sampling points, the mean value of each parameter is calculated. with standard deviation : ; In the formula, This represents the value of the sampled points within the window. For the current verification point... If satisfied Then determine If the value is an outlier due to transient distortion, set it to null and proceed to the next step of processing.

[0051] 2) Missing value imputation based on linear interpolation algorithm: For local data sequence breaks caused by communication packet loss or outlier removal, a linear interpolation algorithm is used for numerical imputation to ensure the continuity of the time series. Assume that at time... There are missing values ​​at this location. By retrieving the two most recent valid data points before and after it. and The fill value is calculated based on a linear proportional relationship, and the interpolation formula is as follows: In the formula, , These are the timestamps of the valid points. , This corresponds to the valid sampled value. It should be noted that if the continuous missing time of the sensor exceeds the system's set limit threshold, interpolation will stop and a sensor disconnection or communication abnormality message will be output.

[0052] 3) Noise Smoothing Filtering Based on First-Order Exponential Moving Average: To eliminate the underlying white noise caused by high-frequency mechanical vibrations in the SCADA system and prevent fluctuations in subsequent calculations of derived features such as the first derivative, a first-order exponential moving average (EMA) algorithm is used to smooth the interpolated data stream. Its filtering formula is defined as: In the formula, for The smoothed output value after time-lapse filtering. for The actual input value at time 1. This is the smoothed output value from the previous sampling time. This is a smoothing coefficient, with a range of values. Its numerical value determines the system's attenuation weight for historical data and its sensitivity to new data.

[0053] 1.4 Feature Derivation Based on High-Frequency Physical Mechanisms:

[0054] This invention utilizes the temporal continuity of high-frequency analog quantities and the spatial coupling between different sensor parameters to calculate a series of derived variables based on physical mechanisms for subsequent anomaly detection model judgment. It explicitly integrates prior physical knowledge in the wind power field into data features, providing a more reliable basis for subsequent anomaly detection models to detect minor abnormal faults.

[0055] The system operates within a set high-frequency time sliding window (the window contains W sampling points, and the sampling interval is...). Within the range, the physical derivative features extracted and the calculation formulas obtained through real-time edge computing are shown in Table 2.

[0056] Table 2 Real-time computation at the edge

[0057] Specifically, in Table 2, Formula 1 is: Formula 2 is: Formula 3 is: Formula four is: Formula 5 is: Formula six is: Formula 7 is: .

[0058] 1.5 Data downsampling and structured aggregation:

[0059] 1) Data frequency reduction based on time window statistics:

[0060] The core principle of this step lies in optimizing transmission bandwidth for edge-cloud collaboration by employing a time-window statistical aggregation method on the edge computing devices. To significantly reduce the bandwidth pressure on the original millisecond-level data flow transmitted to the central server while fully preserving high-frequency transient characteristics (such as current and voltage oscillations), the system does not use traditional equal-interval sampling frequency reduction, but instead employs a fixed-time-window feature extraction scheme. The specific implementation plan is as follows:

[0061] Specifically, edge computing devices set sliding time windows in the order of seconds (such as 1 second or 5 seconds). For the cleaned original high-frequency analog quantities and derived variables within the window, a unified statistical aggregation calculation is performed, mapping them to second-level statistics representing the characteristics of that time period. Each high-frequency time series variable is within the window. The data frequency is transformed into the following four statistical dimensions as statistical features of the down-frequency analog signal: ① Mean: representing the overall physical baseline state within a second-level period; ② Maximum and Minimum values: capturing high-frequency glitches, abrupt changes, and transient peak loads; ③ Variance: representing the intensity and degree of dynamic fluctuation of the high-frequency signal within that period. Through the above calculations, the data frequency is reduced from the millisecond level to the second level, which, while conforming to the input format of the LightGBM model on the central side, achieves efficient lightweight data collaboration between the edge and the central server.

[0062] 2) Data Structure Packaging: After completing the frequency reduction feature, on the edge computing device side, discrete statistics, state variables, and static labels are standardized and aggregated to construct a structured frequency reduction feature message. The assembly logic of the packaging protocol is as follows: ① Time-series feature encapsulation: The four statistical features (mean, maximum, minimum, and variance) of all variables involved in each second-level period are pushed into the time-series feature vector in index order; ② State metadata appending: Real-time digital state variables (such as current operating status fault codes, control words, and generator braking modes) parsed in the concatenation at the end of the time-series feature vector (section 1.2 above) are appended; and static attribute metadata and timestamps are appended to the message header; ③ Structured frame packaging: The above combined data is encapsulated into standardized structured data frames and uploaded to the central server at regular intervals through the network channel as the input benchmark for the subsequent LightGBM multi-label anomaly detection model.

[0063] In one embodiment, the structured down-frequency data frame is parsed by the central server. Parsing the structured down-frequency data frame and assembling a high-dimensional input feature tensor includes: unpacking the structured down-frequency data frame according to a preset communication protocol to obtain static attribute metadata and digital state variables; the static attribute metadata is used as a model routing index to match the corresponding pre-trained model instance; the static attribute metadata is located in the header of the unpacked data, including the version number, the rated frequency of the pitch motor, and the gear ratio; the unpacked digital state variables are used as query keys to input the global mapping dictionary; the global mapping dictionary outputs the corresponding continuous integer index when the query is successful, and outputs the default fault-tolerant index and records the corresponding unknown status code when the query is unsuccessful; wherein, the continuous integer index is used as the category feature of the pre-trained model instance; within a unified time step, the category feature, the down-frequency analog quantity statistical feature, and the physical derived statistical feature are horizontally concatenated to obtain the high-dimensional input feature tensor.

[0064] In one embodiment, the anomaly detection model consists of several independent binary classifiers; the binary classifiers are gradient boosting decision tree-based binary classifiers; the training method of the anomaly detection model includes: extracting historical data from the historical SCADA data of the wind turbine and pitch system backward from the time of the fault occurrence as positive samples, and extracting historical data of equal length from the stable operating period as negative samples; the positive and negative samples are used to perform high-frequency data cleaning, feature derivation based on high-frequency physical mechanisms, and data down-frequency reduction based on time window statistics to obtain a historical feature matrix; based on pitch drive anomalies, pitch motor anomalies, backup power system anomalies, angle encoder anomalies, mechanical transmission anomalies, and soft communication anomalies, the historical feature matrix is ​​non-interchangeably processed. Multiple labels are used for labeling, and a binary classifier is trained for each anomaly label. The false positive penalty weight is normalized to a baseline value, and each binary classifier is trained using a cost-sensitive loss function where the false negative penalty weight is greater than the false positive penalty weight. This yields binary classifiers corresponding to each anomaly label, and the trained binary classifiers constitute the anomaly detection model. The cost-sensitive loss function is a loss function obtained by introducing an asymmetric cost matrix into the binary cross-entropy loss function. The asymmetric cost matrix is ​​used to apply a false negative penalty weight to the false negative loss caused by anomaly samples being predicted as normal states, and a false positive penalty weight to the false positive loss caused by normal samples being predicted as anomaly states, so that each binary classifier preferentially fits the feature boundaries of high-risk anomalies in the pitch system.

[0065] It should be added that: pitch drive malfunction indicates a problem with the inverter or DC bus drive layer equipment in the pitch control cabinet; pitch motor malfunction indicates a problem with the electrical components of the pitch motor body or internal windings; backup power system malfunction indicates a problem with the supercapacitor bank or backup battery bank; angle encoder malfunction indicates a problem with the absolute or relative encoder that measures the blade position; mechanical transmission malfunction indicates a problem with the pitch reducer, gearbox, or pitch bearing; and soft communication malfunction indicates a problem with the data transmission link of the slip ring or control bus.

[0066] It should be added that the penalty weight for underreporting is determined based on the severity coefficient of the corresponding anomaly category, the converted value of spare parts replacement cost, and the converted value of power generation loss during the shutdown. The penalty weight for false reporting is determined based on the standardized average cost of a single unplanned shutdown tower inspection, and the standardized average cost is normalized to a benchmark value.

[0067] It should also be noted that the expressions for the cost-sensitive loss function and the missed detection penalty weights are as follows:

[0068] ;

[0069] ;

[0070] In the formula, For the first The loss value of the anomaly. For the first The true label of the class exception, For the first Predicted probability of anomalies Weighting for penalties for underreporting Weighting for false alarm penalties For the first The severity coefficient of the anomaly. This is the discounted value for spare parts replacement costs. This is the converted value of power generation loss during the shutdown period.

[0071] In one embodiment, the method for obtaining the comprehensive anomaly probability vector includes: inputting the high-dimensional input feature tensor into each binary classifier in parallel, accumulating the weights of the leaf nodes of each decision tree in each binary classifier to obtain the original log odds of the corresponding anomaly; mapping the original log odds to anomaly probabilities using a logistic function, and obtaining the comprehensive anomaly probability vector by concatenating the anomaly probabilities; the anomaly probability value is between zero and one; determining the anomaly classification result includes: determining a baseline safety threshold based on the missed detection penalty weight and the false alarm penalty weight, and performing environmental stress compensation on the baseline safety threshold according to the wind speed variance in the current reduced-frequency analog quantity statistical features to obtain the operating condition adaptive dynamic safety threshold; if all anomaly probabilities are lower than the corresponding operating condition adaptive dynamic safety threshold, the anomaly classification result is determined to be a normal state; otherwise, the anomaly category with anomaly probability exceeding the corresponding operating condition adaptive dynamic safety threshold is determined as the anomaly classification result.

[0072] It should be added that the expressions for the baseline safety threshold and the adaptive dynamic safety threshold under operating conditions are as follows: ; In the formula, For the first The baseline safety threshold for class anomalies, For the first False alarm penalty weights for exception types For the first The penalty weight for missed reports of anomalies, The normalized wind speed variance For the first Compensation coefficient for abnormal operating conditions, For the first The absolute safety upper limit threshold for class exceptions. For the first The corresponding working condition adaptive dynamic safety threshold for each type of anomaly.

[0073] Step 2: Multi-label anomaly detection based on LightGBM:

[0074] This module is deployed at the system's central server (AI server), playing a crucial role in the overall edge-cloud collaborative architecture. Its main purpose is to receive high-frequency down-frequency statistics and derived features from edge computing devices, and use the LightGBM algorithm based on a gradient boosting tree framework to perform nonlinear mapping of high-dimensional feature domains, thereby achieving accurate identification of various concurrent anomalies in the pitch system.

[0075] 2.1 Model Routing and Model Input:

[0076] After receiving structured down-frequency data frames periodically uploaded by the edge computing devices, the central server performs message unpacking, model routing allocation, and high-dimensional feature tensor reassembly according to the following sequential steps to finally generate model input conforming to the LightGBM engine standard format:

[0077] 1) Data frame reception and unpacking: The central server (AI inference server) receives structured down-frequency data frames transmitted from the edge side in real time, performs unpacking and stripping operations according to the preset communication protocol, and restores the binary message or standard communication object into a structured multi-dimensional data dictionary in preparation for splitting and parsing.

[0078] 2) Dynamic model routing based on static metadata: Extract static attribute metadata (such as version number, pitch motor rated frequency, gear ratio, etc.) from the header of the unpacked data and use it as a model routing index. The system uses this index to match the corresponding LightGBM pre-trained model instance, thereby ensuring the consistency of physical baselines for different models.

[0079] 3) Category Feature Mapping of Digital Status Quantities: Extract real-time digital status quantities from the unpacked data, including generator braking mode, status words, control words, and various fault codes at the unit level. Since the original status codes and fault codes at the wind turbine's underlying layer are usually non-contiguous hexadecimal instructions or extremely large jump values ​​(e.g., 0x00, 0x2A, 9999, etc.), to avoid the exponential explosion of the feature space and the generation of a high-dimensional sparse matrix caused by traditional one-hot encoding, this invention constructs a continuous integer mapping mechanism based on a global hash dictionary. The specific execution steps are as follows: Step A (Global Mapping Dictionary Maintenance): Maintain a global status mapping dictionary in the central server's memory. This dictionary uses the original discrete status codes that the wind turbine's underlying layer may output as keys, and increments by consecutive non-negative integers (e.g., ...). Step B (Real-time Retrieval and Dynamic Mapping): When the system parses the real-time raw digital status code, it uses it as the query key to input into the hash dictionary for comparison and retrieval. If a match is found, the corresponding continuous integer index is extracted as the output; if a match is not found (i.e., the system encounters an unknown status code or a new type of fault code for the first time), the system executes boundary protection logic, mapping the unknown code to a preset default fault-tolerant index (such as -1 or a dedicated Unknown category index), and asynchronously records the unknown code for subsequent dictionary updates. Step C (Category Feature Declaration): After the above transformation, the original discrete digital status quantities are uniformly regularized into a compact continuous integer sequence. When assembling the feature tensor, the system explicitly declares these continuous integer indices as LightGBM's native categorical features, so that when the tree model splits nodes, it can directly find the optimal discrete split point through the histogram algorithm, greatly improving inference efficiency.

[0080] 4) High-dimensional feature tensor assembly and caching: Within a unified time step, the system performs feature dimension concatenation. The categorical features generated in step (3) are horizontally concatenated with the SCADA down-frequency analog quantity features (mean, maximum, minimum, and variance of each parameter) and the physical derived features (see Table 2). The concatenated data combination constitutes a high-dimensional input feature tensor containing all dimensional features. Input the high-dimensional feature tensor The feature tensor is used as input to an anomaly detection model to perform multi-label anomaly classification inference. Further, the feature tensor... Store it in the cache queue for use in subsequent steps.

[0081] 2.2 Training Set Construction and Model Training:

[0082] This system utilizes historical SCADA data from wind farms at the central server to construct, train, and validate the LightGBM model. To address the complex concurrent fault mechanisms of wind turbines and the industry pain point of extremely imbalanced samples, this embodiment adopts the following technical solution:

[0083] 1) Training Set Construction: To build a high-quality training sample library, the historical SCADA time-series database of the target wind farm is retrieved. ① Time-series alignment and sample truncation: Based on the recorded fault occurrence time, historical high-frequency data of T hours (e.g., 2 hours) are trunculated backward as positive samples (fault samples); during the stable operation period of the unit, data segments of equal length are randomly truncated as negative samples (normal samples). ② Feature engineering processing: The truncated historical samples undergo the same data preprocessing operations as the edge computing device data preprocessing process described in step 1 to generate a standardized historical feature matrix.

[0084] 2) Model classification label definition:

[0085] To achieve rapid delimitation of faults in the core equipment of the pitch system, thereby providing clear physical entity pointers for subsequent RAG knowledge retrieval and semantic generation of large language models, and considering that a single underlying physical fault in the wind turbine pitch system may simultaneously trigger multiple apparent anomalies, this invention, based on the core hardware composition of the pitch system, comprehensively evaluates and subjectively defines six anomaly modes. The system constructs a non-mutually exclusive multi-label vector containing seven dimensions. .

[0086] Specific tags are defined as : Pitch drive malfunction: refers to malfunctions in drive layer equipment such as frequency converters and DC buses within the pitch control cabinet; Pitch motor malfunction: Malfunction of the pitch motor body and its internal windings and other electrical components; Backup power system malfunction: Supercapacitor bank or backup battery bank malfunction. Angle encoder malfunction: The absolute / relative encoder sensor responsible for measuring the blade position is malfunctioning. Abnormal mechanical transmission: Highly coupled mechanical components such as pitch reducers, gearboxes, or pitch bearings; : Soft communication anomaly: Data transmission link such as slip ring or control bus (e.g., CAN / EtherCAT); Normal state: Mutually exclusive label: if and only if to When the model output probabilities are all below the dynamic safety threshold Set to 1, otherwise set to 0.

[0087] During the model training phase, the ground truth labels in the historical training samples are manually annotated by wind power professionals. These professionals extract SCADA time-series feature data within historical time windows and combine it with synchronously recorded unit digital status codes, fault codes, and fault records for comprehensive analysis. They manually confirm which (or which pieces) of core equipment experienced a problem during that period, thus labeling the sample with the corresponding anomaly dimension (0 or 1). During the model inference phase, six independent LightGBM binary classifiers deployed on the central server correspond to... Anomaly patterns. For the input real-time feature tensor, each classifier computes in parallel and outputs the probability of the corresponding anomaly occurring. The anomaly label is determined based on probability; the anomaly diagnosis model outputs an anomaly category to the next stage, which may not only have one anomaly label, but may have multiple anomalies in the current state.

[0088] 3) Model building, training, and inference:

[0089] Specifically, model construction: An ensemble of multiple weak learners (i.e., decision trees) built based on the Gradient Boosting Decision Tree (GBDT) architecture is termed an ensemble tree. Because the boundaries of the high-dimensional physical feature matrix of the wind turbine pitch system are extremely complex, a single decision tree cannot achieve high-precision fitting. Therefore, this invention employs an ensemble tree strategy, continuously iteratively training the model using a forward distribution algorithm. Each newly added decision tree is specifically used to fit the residual (i.e., the negative gradient between the true label and the current predicted value) of the prediction results of all tree combinations in the previous round, thereby constructing a strong classification model through the serial combination of numerous weak classifiers. Specifically, this invention constructs six independent LightGBM binary classifiers for the non-mutually exclusive anomaly label space of various core devices in the pitch system, corresponding to... to These are the six anomaly modes. The specific input-output mapping mechanism during the model inference and output phases is as follows:

[0090] Specifically, the model input: for the current time step The system will generate the high-dimensional feature tensor from section 2.1 above. (Including statistical features of SCADA down-frequency analog signals, physical derived features, and category features mapped to continuous integers) are input synchronously and in parallel into these 6 independent LightGBM model instances. In each independent LightGBM binary classifier, for the k-th anomaly ( When the feature tensor After input, the binary classifier Each decision tree will calculate the weight of each leaf node (i.e., the score of an anomaly judgment for a single tree), and the final output of this binary classifier is... The weights of the tree nodes in the decision tree are directly summed and denoted as the original log-odds. The calculation formula is as follows: In the formula, Indicates the first The decision trees assign scoring weights to the current feature tensor. This is the original log-odds ratio. The range of values ​​is Furthermore, the logarithmic probability is mapped to... through the sigmoid function of the output layer. The probability of an interval is calculated using the following formula: After the system executes inference for six sub-models in parallel, it concatenates the results to output a comprehensive anomaly probability vector. Among them, the statistical characteristics of down-frequency analog quantities.

[0091] Specifically, in model training: In wind turbine pitch systems, fault samples are extremely rare, and the cost of misjudgment is highly uneven. False negatives (FN) lead to missed faults, resulting in physical and economic costs far exceeding the cost of a single unplanned downtime for inspection caused by false positives (FP). Therefore, this invention introduces an asymmetric cost matrix into the default binary cross-entropy loss function of LightGBM to increase the loss for missed faults (FN), thereby improving the model's sensitivity to abnormal trends. For the first... Each anomalous sub-model employs a cost-sensitive loss function with weighted penalties. Its calculation formula is defined as: In the formula, For this sample in the 1st The true label on the class exception, For the model to predict the occurrence of the first Confidence probability of an anomaly class Weighting for underreporting penalties (cost of false negatives). This refers to the false positive penalty weight (false positive cost). False positive penalty weight. Set the standardized average cost of a single unplanned downtime tower inspection, normalize it to a baseline value of 1, and assign a penalty weight for underreporting. The calculation formula is .in, For the first The severity coefficient of an anomaly jeopardizing the overall safety of the unit is mapped to an integer from 1 to 9 based on the severity of the anomaly (e.g., a missed report could lead to a mechanical transmission anomaly causing tower collapse or runaway). Take the highest value 9). The cost of replacing the standard spare part after the component is damaged (relatively converted to) (multiples of) The equivalent power generation loss corresponding to the average downtime required to replace this component (relatively converted to) (multiples of) It is an important parameter in the binary cross-entropy loss function. Setting the false positive penalty weight to 1 is to determine a relative reference. When calculating the false positive penalty weight, the relative reference value of 1 is... , Value; in other words, if the standardized average cost of a single unplanned shutdown tower inspection is x, and the cost of replacing the core component for anomaly i is 1.5x, then The value is 1.5. This mechanism transforms engineering economic indicators into gradient amplification multipliers, forcing LightGBM to prioritize capturing the feature boundaries of high-risk faults.

[0092] Specifically, model inference: After the system executes the inference of 6 sub-models in parallel, it concatenates and outputs a comprehensive anomaly probability vector. To avoid the defects of conventional fixed-threshold pitch systems, which are prone to high-frequency false alarms or missed alarms under complex and variable wind conditions, this invention introduces a condition-adaptive dynamic safety threshold mechanism. The specific implementation steps of this mechanism are as follows: ① Calculate the theoretical baseline threshold: Based on the cost weights during model training, the system first calculates the baseline safety threshold that minimizes the overall expected loss. ② Real-time environmental stress compensation: Extract the wind speed variance from the current SCADA frequency reduction characteristics as a real-time environmental stress index. (Normalization to) The final dynamic security threshold is calculated as follows: In the formula, The working condition compensation coefficient is set based on historical experience. The absolute safety upper limit threshold is set at 0.6. When the unit faces extreme gusts or high turbulence conditions, As the threshold increases, the safety threshold will rise accordingly to absorb reasonable physical oscillations. Under stable operating conditions, it will be relatively close to the baseline value. ③ Comprehensive mutual exclusion judgment: The safety threshold is determined if and only if the output probability of all elements in the comprehensive probability vector is strictly lower than the dynamic safety threshold corresponding to the current moment (i.e., ...). The system infers mutual exclusion labels. The probability of the (normal state) occurring is 1; otherwise, it is determined that an event has occurred. An exception occurs and triggers the subsequent SHAP dynamic attribution process.

[0093] 2.3 Online Anomaly Classification Reasoning:

[0094] The online anomaly classification and inference module is primarily deployed on the central server. This process aims to construct a multi-dimensional feature matrix by acquiring structured data frames transmitted from edge computing devices in real time. Further, an inference engine composed of multiple LightGBM classifiers performs multi-label anomaly classification, and finally outputs the anomaly determination through confidence gating. The specific process is as follows: Figure 3 As shown, the process includes: Step ①, data frame reception and unpacking; Step ②, model dynamic routing; Step ③, category feature mapping; Step ④, feature vector assembly and caching; Step ⑤, parallel model inference; Step ⑥, joint probability vector calculation. Steps ① to ④ are as described in Section 2.1 above. Step ⑤ involves assembling the structured feature vectors... enter A parallel binary anomaly classification model For the first Each classifier, the system internally The weights of the leaf nodes output by each decision tree are directly summed using scalar methods to obtain the current feature tensor. Triggering the Step 6: Aggregate the log odds of each binary anomaly classification model, perform exponential smoothing and normalization using the Sigmoid function, and generate a multidimensional joint anomaly probability vector for the current time window. Each dimension represents the confidence level that the current system has a corresponding anomaly.

[0095] In one embodiment, structured metadata determines the physical feature change trend through historical down-frequency physical statistics, and maps the anomaly classification results, corresponding confidence levels, core contributing features, and physical feature change trends. Determining the core contributing features includes: when the anomaly probability of any anomaly category exceeds the corresponding adaptive dynamic safety threshold, intercepting the current high-dimensional input feature tensor from the feature tensor cache queue, and activating the tree model interpreter corresponding to any anomaly category; calculating the feature contribution value (SHAP value used in this embodiment) of each feature in the high-dimensional input feature tensor to any anomaly category using the tree model interpreter, and taking the value of each feature contribution. The absolute values ​​are sorted in descending order of absolute contribution, and the top-ranked features (a predetermined number) are identified as core contributing features. Structured metadata is generated, including: extracting the physical value sequences of each core contributing feature from the feature tensor cache queue within past continuous time windows, and fitting the slope of each physical value sequence using the least squares method; determining the corresponding physical feature change trend based on the relationship between the change slope and preset gradual change and abrupt change thresholds; and serializing the timestamp, anomaly classification result, corresponding confidence level, feature name of each core contributing feature, SHAP value, current physical value, change slope, and physical feature change trend to obtain structured metadata.

[0096] It should be added that, based on the relationship between the slope of change and the preset gradual change threshold and abrupt change threshold, the corresponding physical characteristic change trend is determined, including: when the absolute value of the slope of change is less than or equal to the gradual change threshold, the corresponding physical characteristic change trend is determined to be steady state; when the absolute value of the slope of change is greater than the gradual change threshold and less than or equal to the abrupt change threshold, the corresponding physical characteristic change trend is determined to be gradual degradation; when the absolute value of the slope of change is greater than the abrupt change threshold, the corresponding physical characteristic change trend is determined to be a step abrupt change; the gradual change threshold and the abrupt change threshold are obtained by statistical analysis based on the historical SCADA data of the corresponding wind field, and the gradual change threshold is less than the abrupt change threshold.

[0097] It should also be noted that the expression for the slope of change is: In the formula, For the first The slope of change of each core contribution feature The number of consecutive time windows. For the first Time index of a continuous time window The average of time indices. For the first Physical values ​​for a continuous time window. This is the mean of the physical value sequence.

[0098] Step 3: Structured metadata mapping based on dynamic SHAP:

[0099] 3.1 Overview: In high-safety-level energy and industrial control fields such as wind power pitch control systems, the direct application of machine learning algorithms such as integrated tree models or deep learning often faces challenges in terms of engineering reliability and scientific traceability due to their black-box nature and lack of interpretability. This section aims to solve this inherent problem of uninterpretability and provide a quantitative and traceable physical evidence chain for downstream Large Language Models (LLM). Its core principle is: when the anomaly probability output by the multi-label classification model exceeds the calculated dynamic safety threshold, the system selectively activates the TreeSHAP interpreter. Unlike traditional schemes that only perform static attribution at the moment of anomaly occurrence, this invention proposes a dynamic temporal attribution mechanism: first, extracting the Top-N core features that contribute the most to the anomaly at the current moment; then, backtracking the original physical numerical sequences of these N features through a historical sliding window, and using first-order derivative fitting to calculate their true physical evolution trend (such as step abrupt change or gradual degradation). Ultimately, the anomaly labels, confidence levels, core features, and their physical evolution trends in multidimensional numerical form are encapsulated into standardized JSON data. This enables the mapping from the underlying high-dimensional continuous numerical feature matrix to the high-level discrete structured form, transforming pure mathematical continuous time-series physical quantities into natural language / structured text forms that can be directly understood and reasoned about by generative large models (LLM). This bridges the gap between underlying heterogeneous data and high-level semantic large models through cross-modal protocols.

[0100] 3.2 Implementation steps: such as Figure 4 As shown in the diagram, the system performs dynamic attribution and semantic mapping on the central server according to the steps outlined in the diagram above, as follows:

[0101] Step ①, Confidence-gated activation: Monitor the output probability of each binary classifier in the LightGBM multi-label anomaly detection framework. If the first classifier is determined to be a classifier with a confidence level of 100%, the second classifier will be determined to be a classifier with a confidence level of 100%. Output probability of class exceptions Exceeding the dynamic security threshold The system then intercepts the current time point. Input feature tensor And activate the model for that anomaly category. The TreeSHAP interpreter.

[0102] Step ②: Top-N Contribution Feature Extraction Based on SHAP: By calculating the expected marginal contribution of feature combinations, the black-box nature of the ensemble tree model is broken. High-confidence classification results are traced back to calculate the specific contribution of each feature in the current input tensor to the anomaly determination result, thereby accurately locating the physical feature quantity that caused the anomaly. Specific implementation sub-steps include: 1) Marginal contribution calculation: Inputting the current time-to-time model... Feature tensor Input Model The corresponding TreeSHAP interpreter. Based on the additive interpretation model, the model... Predicted output It can be decomposed into the sum of the baseline value and the contribution of each feature. Calculate the current tensor. The Middle SHAP value of each feature 1) Marginal contribution: This refers to the marginal contribution of all features. 2) Marginal contribution ranking and TOP-N feature extraction: Extract the marginal contribution of all features and take their absolute values. For feature tensors Each feature is sorted in descending order of its absolute contribution. The system extracts the top-ranked features. The characteristic variables are those that are highly correlated with the current anomaly. One characteristic.

[0103] Step 3: Extraction of physical feature change trends based on historical window backtracking: After selecting the TOP-N contribution features in Step 2, the original down-frequency physical statistics (such as mean and maximum) of these features are backtracked through historical time series windows, and their change slopes are calculated to determine the qualitative label of the change trend. Specific steps include: 1) Historical window data extraction: Based on the statistical physical quantities (such as motor current - maximum) corresponding to the TOP-N features determined in Step 2, the feature dimension is extracted from the feature tensor cache queue in the past... A sequence of physical values ​​over a continuous time window 2) Slope (first derivative) fitting: for the sequence Introducing time series indexes The feature was fitted using the least squares method. Slope of linear change on the time axis The calculation formula is: In the formula, For time window indexing, The average of the time index; For the characteristic physical values ​​of the corresponding window, 3) Trend mapping: The calculated absolute value of the slope With each eigenvalue Preset engineering experience threshold (gradually varying threshold) With mutation threshold Both are obtained through statistical analysis of SCADA data for each wind turbine in each wind farm. The threshold is calculated by taking the mean from a large amount of statistical data. This threshold is used to qualitatively determine the magnitude of changes in the changing characteristics, assisting the large model in understanding the output (it does not directly affect the model's judgment). A qualitative trend label is output by comparing the results. For example, if step ② determines that the feature with the highest contribution is the mean value of the inverter temperature, the slope of the fitted time window is... ,like If the feature has changed drastically, it is labeled as a step change; if The characteristic change trend is determined to be gradual deterioration; otherwise, it is determined to be in a steady state.

[0104] Step 4, Structured Mapping and Output: The anomaly classification results, confidence scores, and quantitative and qualitative calculations of TOP-N physical features generated above are serialized into JSON structured metadata. This structured data will serve as the core input source for the prompt word project and will be pushed to the large language model. A partial code example is shown below:

[0105] "{"Timestamp": "2026-05-25T14:32:01Z",

[0106] "Detected_Anomalies": [

[0107] {"Anomaly_Type": "Class_2 pitch motor anomaly",

[0108] "Confidence": 0.94},

[0109] {"Anomaly_Type": "Class 5 mechanical transmission anomaly",

[0110] "Confidence": 0.88}

[0111] ],

[0112] "Top_N_Features": [

[0113] {"Feature_Name": "Maximum value of dynamic rate of change of current",

[0114] "SHAP_Contribution": 3.12,

[0115] "Current_Value": "45.5 A / s",

[0116] "Fitted_Slope": 12.5,

[0117] "Trend_Label": "Step mutation"},

[0118] {"Feature_Name": "Average value of cumulative transient thermal load on motor",

[0119] "SHAP_Contribution": 2.45,

[0120] "Current_Value": "8500 A 2 ·s",

[0121] "Fitted_Slope": 150.2,

[0122] "Trend_Label": "Gradual degradation"},

[0123] {"Feature_Name": "Maximum value of pitch angle synchronization deviation rate",

[0124] "SHAP_Contribution": 1.98,

[0125] "Current_Value": "2.8%",

[0126] "Fitted_Slope": 0.05,

[0127] "Trend_Label": "High-level stable state"} ]

[0129] }".

[0130] In one embodiment, generating an equipment anomaly maintenance report for the pitch system includes: parsing structured metadata to extract anomaly categories, feature names of each core contributing feature, and corresponding physical feature change trends, and concatenating the extraction results into a structured natural language query vector; inputting the structured natural language query vector into a locally deployed vector database, and retrieving associated document fragments matching the anomaly categories and physical feature change trends based on vector similarity; the vector database stores historical maintenance work orders for the wind farm, pitch system fault manuals, and equipment manuals; formatting and assembling the structured metadata and associated document fragments by calling maintenance diagnosis prompt word templates; the maintenance diagnosis prompt word templates are used to constrain the output content; the output content includes at least anomaly cause analysis and maintenance inspection suggestions; and inputting the formatted and assembled prompt words into a locally deployed large language model for semantic reasoning to output the equipment anomaly maintenance report for the pitch system.

[0131] Step 4: Cross-modal semantic reasoning:

[0132] This module is deployed on a central server and utilizes locally deployed Large Language Model (LLM) and Retrieval Enhancement Generation (RAG) technologies to transform the structured data output from the previous stage into maintenance and troubleshooting reports that field engineers can directly execute, thus completing a technical closed loop from underlying data perception to maintenance business decisions. Figure 5 As shown, the specific process is as follows:

[0133] Step 1, Structured JSON Parsing and Search Term Assembly: The system receives the JSON structured metadata output in Step 3, extracts the anomaly categories and feature names and qualitative trend labels from the Top_N features, and concatenates them into a structured natural language query vector.

[0134] Step 2: Domain Knowledge Retrieval Based on RAG: Input the above query vectors into the locally deployed vector database (covering historical wind farm maintenance work orders, pitch system fault manuals, equipment manuals, etc.). Through vector similarity calculation, retrieve the Top-K historical handling solutions or standard operating procedures that best match the current anomaly type and feature evolution trend as the external knowledge base context.

[0135] Step 3: Dynamic Injection of Prompt Templates: The system calls a preset operation and maintenance diagnosis prompt template, and formats and assembles the parsed JSON data with the related document fragments retrieved by RAG. The template strictly constrains the LLM output, requiring it to be output according to the structure of anomaly cause analysis and operation and maintenance inspection suggestions.

[0136] Step 4, Local Large Model Inference and Result Output: The concatenated prompt words are used for semantic inference via the LLM large language model API. In this invention, the LLM uses an open-source domestic large model, and the basic model structure and pre-training process of this large language model are not fine-tuned. The LLM combines the multi-dimensional information of the input to eliminate the understanding barrier of technical terms and generate and output the final troubleshooting guidance report for the pitch system equipment anomaly.

[0137] According to another embodiment of the present invention, an intelligent operation and maintenance device for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning is also provided. This device includes: an edge feature module, used for parameter parsing and state gating of the acquired data from the wind turbine and pitch system, and for high-frequency data cleaning of the released analog quantities; the released analog quantities are used for feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning to generate structured down-frequency data frames; and a cloud diagnostic module, used for parsing the structured down-frequency data frames and assembling high-dimensional inputs. The feature tensor is input into the anomaly detection model to obtain a comprehensive anomaly probability vector. This comprehensive anomaly probability vector is used to determine the anomaly classification result by combining it with the corresponding adaptive dynamic safety threshold for the operating condition. The pitch attribution module is used to determine the comprehensive anomaly probability vector when the anomaly probability exceeds the corresponding adaptive dynamic safety threshold for the operating condition, and to attribute the feature contribution of the high-dimensional input feature tensor to determine the core contributing features. The core contributing features are used to backtrack historical down-frequency physical statistics to generate structured metadata. The semantic operation and maintenance module is used to generate an equipment anomaly operation and maintenance report for the pitch system based on the parsing results of the structured metadata.

[0138] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent operation and maintenance of a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning, characterized in that, include: The collected data from the wind turbine and pitch system are subjected to parameter analysis and state gating, and the released analog quantities are cleaned using high-frequency data cleaning. The released analog quantity is used to perform feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning, so as to generate structured down-frequency data frames. The structured down-frequency data frame is parsed and a high-dimensional input feature tensor is assembled. The high-dimensional input feature tensor is then input into the anomaly detection model to obtain a comprehensive anomaly probability vector. The comprehensive anomaly probability vector is used to determine the anomaly classification result by combining it with the corresponding adaptive dynamic safety threshold for the working condition. A comprehensive anomaly probability vector is determined when the anomaly probability exceeds the corresponding adaptive dynamic safety threshold for the working condition. The feature contribution of the high-dimensional input feature tensor is attributed to determine the core contribution feature. The core contribution feature is used to backtrack historical down-frequency physical statistics to generate structured metadata. Based on the parsing results of the structured metadata, an equipment anomaly maintenance report for the pitch system is generated.

2. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 1, characterized in that, The data collected by the wind turbine and pitch system is collected by edge computing devices, including high-frequency real-time SCADA data streams collected by the wind turbine main control system and the pitch subsystem. The parameter parsing and state gating process includes: The original message of the high-frequency real-time SCADA data stream is unpacked, and the parameters obtained from unpacking are divided into static attribute metadata, digital status quantities and high-frequency analog quantities according to the physical attributes and update frequency of the variables. The static attribute metadata is used as device tag information, the digital status quantities are used as dynamic gating basis, and the high-frequency analog quantities are written into a circular buffer. Based on the digital status variables, the operating status of the wind turbine is identified; When the wind turbine is in generator brake engaged, system initialization, feathering shutdown locked, or manual maintenance mode, a lockout command is output to clear the annular buffer and suspend subsequent processing. When the wind turbine is in grid-connected power generation, yaw-seeking wind, or dynamic pitch control mode, a conduction command is output to perform high-frequency data cleaning on the analog quantity released by the conduction command.

3. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 2, characterized in that, The high-frequency data cleaning of the released analog quantity includes: Within a time sliding window containing a preset number of sampling points, calculate the window mean and window standard deviation of the current parameter, and set the current verification point whose deviation from the window mean is greater than a preset multiple of the window standard deviation to a null value. Retrieve the two most recent valid data points before and after the location of the null value or communication packet loss, and perform linear interpolation according to the time ratio; the linear interpolation stops when the continuous missing time exceeds the limit threshold, and outputs sensor disconnection or communication abnormality; The high-frequency analog quantity after interpolation is subjected to a first-order exponential moving average to obtain the released analog quantity after cleaning.

4. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 1, characterized in that, The feature derivation based on high-frequency physical mechanisms includes: Within a high-frequency time sliding window, the angle range of the three blades is obtained based on the actual angles of the three blades, and the pitch angle synchronization deviation rate is calculated based on the angle range of the three blades and the average actual angle of the three blades. The position tracking error integral is obtained by comparing the pitch angle setpoint with the actual angle. The dynamic rate of change of the current is obtained by comparing the pitch motor current at adjacent sampling times. The transient thermal load accumulation of the motor is determined based on the pitch motor current within the high-frequency time sliding window. The standard deviation of bus voltage fluctuation is calculated using the DC bus voltage within a high-frequency time sliding window, and the encoder redundancy dispersion is determined based on the actual angle of the main encoder and the angle of the redundant encoder. Physically derived features are constructed based on the three-blade angle range, pitch angle synchronization deviation rate, position tracking error integral, current dynamic change rate, motor transient thermal load accumulation, bus voltage fluctuation standard deviation, and encoder redundancy dispersion.

5. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 4, characterized in that, The data down-frequency reduction based on time window statistics includes: Within the frequency reduction time window, the mean, maximum, minimum and variance of the released analog quantity and the physical derived characteristics after high-frequency data cleaning are calculated respectively to obtain the statistical characteristics of the frequency reduction analog quantity and the statistical characteristics of the physical derived characteristics. The down-frequency analog quantity statistical features and the physical derived statistical features are pushed into the time-series feature vector according to the feature index order; The digital state quantity is concatenated to the end of the temporal feature vector, and the static attribute metadata and timestamp are appended to the beginning to encapsulate the structured down-frequency data frame.

6. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 1, characterized in that, The structured down-frequency data frames are parsed by the central server; The process of parsing the structured down-frequency data frames and assembling the high-dimensional input feature tensor includes: The structured down-frequency data frame is unpacked according to a preset communication protocol to obtain static attribute metadata and digital state variables; the static attribute metadata is used as a model routing index to match the corresponding pre-trained model instance. The unpacked digital state quantity is used as the query key and input into the global mapping dictionary; the global mapping dictionary outputs the corresponding consecutive integer index when the query is hit, and outputs the default fault-tolerant index and records the corresponding unknown state code when the query is not hit; wherein, the consecutive integer index is used as the category feature of the pre-trained model instance; Within a unified time step, the category features, down-frequency analog quantity statistical features, and physical derived statistical features are horizontally concatenated to obtain the high-dimensional input feature tensor.

7. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 1, characterized in that, The anomaly detection model consists of several independent binary classifiers; the training methods for the anomaly detection model include: From the historical SCADA data of the wind turbine and pitch system, historical data is extracted backward from the time of the fault as positive samples, and historical data of equal length is extracted from the stable operation period as negative samples. The positive and negative samples are used to perform high-frequency data cleaning, feature derivation based on high-frequency physical mechanisms, and data down-frequency reduction based on time window statistics to obtain a historical feature matrix. Based on the abnormalities of the pitch driver, pitch motor, backup power system, angle encoder, mechanical transmission, and soft communication, the historical feature matrix is ​​labeled with non-mutually exclusive multi-labels, and a binary classifier is trained for each abnormal label. The false positive penalty weight is normalized to a baseline value, and each binary classifier is trained using a cost-sensitive loss function where the false negative penalty weight is greater than the false positive penalty weight. The trained binary classifiers constitute the anomaly detection model.

8. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 7, characterized in that, The methods for obtaining the comprehensive anomaly probability vector include: The high-dimensional input feature tensor is input into each of the binary classifiers in parallel, and the weights of the leaf nodes of each decision tree in each binary classifier are accumulated to obtain the original log odds of the corresponding anomaly. The original logarithmic odds are mapped to anomaly probabilities using a logical function, and the comprehensive anomaly probability vector is obtained by concatenating the various anomaly probabilities. The determination of the anomaly classification result includes: The baseline safety threshold is determined based on the missed detection penalty weight and the false alarm penalty weight, and environmental stress compensation is performed on the baseline safety threshold according to the wind speed variance in the current down-frequency analog quantity statistical characteristics to obtain the operating condition adaptive dynamic safety threshold. If all anomaly probabilities are lower than the corresponding adaptive dynamic safety threshold for operating conditions, then the anomaly classification result is determined to be a normal state; otherwise, the anomaly category with an anomaly probability exceeding the corresponding adaptive dynamic safety threshold for operating conditions is determined as the anomaly classification result.

9. The intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning according to claim 1, characterized in that, The structured metadata determines the trend of physical feature changes through historical down-frequency physical statistics, and maps the anomaly classification results, corresponding confidence levels, core contribution features, and physical feature change trends. The determination of core contribution features includes: When the probability of any anomaly category exceeds the corresponding working condition adaptive dynamic safety threshold, the high-dimensional input feature tensor at the current moment is intercepted from the feature tensor cache queue, and the interpreter corresponding to any anomaly category is activated. The interpreter calculates the feature contribution value of each feature in the high-dimensional input feature tensor to any anomaly category, takes the absolute value of each feature contribution value and arranges them in descending order of absolute contribution, and determines the first preset number of features as the core contributing features. The generation of structured metadata includes: Extract the physical value sequence of each core contribution feature within the past continuous time window from the feature tensor cache queue, and fit the slope of the change of each physical value sequence using the least squares method; Based on the relationship between the slope of change and the preset gradual change threshold and abrupt change threshold, the corresponding physical characteristic change trend is determined; The timestamp, the anomaly classification result, the corresponding confidence level, the feature name of each core contributing feature, the SHAP value, the current physical value, the slope of change, and the trend of physical feature change are serialized to obtain the structured metadata.

10. An intelligent operation and maintenance device for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning, used to implement the intelligent operation and maintenance method for a pitch system based on edge-cloud hierarchical feature mapping and semantic reasoning as described in any one of claims 1-9, characterized in that, The device includes: The edge feature module is used to perform parameter parsing and state gating on the acquired data of the wind turbine and pitch system, and to perform high-frequency data cleaning on the released analog quantities. The released analog quantities are used to perform feature derivation based on high-frequency physical mechanisms and data down-frequency reduction based on time window statistics after cleaning, so as to generate structured down-frequency data frames. The cloud-based diagnostic module is used to parse the structured down-frequency data frames and assemble a high-dimensional input feature tensor. The high-dimensional input feature tensor is then input into the anomaly detection model to obtain a comprehensive anomaly probability vector. The comprehensive anomaly probability vector is used to determine the anomaly classification result by combining it with the corresponding adaptive dynamic safety threshold for the working condition. The pitch attribution module is used to determine the comprehensive anomaly probability vector when the anomaly probability exceeds the corresponding adaptive dynamic safety threshold of the working condition, and to perform feature contribution attribution on the high-dimensional input feature tensor to determine the core contribution feature; the core contribution feature is used to backtrack the historical down-frequency physical statistics to generate structured metadata. The semantic operation and maintenance module is used to generate an equipment anomaly operation and maintenance report for the pitch system based on the parsing results of the structured metadata.

Images