Wind power equipment fault early warning method and system based on edge calculation

Abstract

The invention belongs to the technical field of wind power equipment operation and maintenance, and discloses a wind power equipment fault early warning method based on edge calculation. The method comprises the following steps: acquiring vibration, temperature and current parameter operation data through a wind power equipment sensor, generating an original data sequence, and preprocessing to obtain a preliminary feature set; the time sequence gradient of each parameter is analyzed and calculated through a time sequence, and a parameter change rate sequence is obtained through optimization of a gradient descent algorithm. When the sequence exceeds a threshold value, a stable change trend is obtained through sliding window smoothing, and anomaly classification and risk grading are completed in combination with a historical working condition database and a support vector machine. And starting deep analysis according to an edge node resource state, constructing a fine prediction model to generate an early warning signal containing a fault probability and occurrence time, matching a risk level to activate a local response, and outputting an alarm. According to the invention, through edge localization processing and multi-algorithm cooperation, accurate pre-judgment of faults is realized, and the operation and maintenance response efficiency and early warning accuracy of wind power equipment are improved.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment operation and maintenance technology, and in particular relates to a wind power equipment fault early warning method and system based on edge computing. Background Technology

[0002] As a core pillar of clean energy, wind power plays a crucial role in the global energy transition, and the stable operation of its equipment directly affects the reliability and economic benefits of energy supply. Especially in remote or offshore wind farms, equipment failures can lead to high maintenance costs and long-term downtime losses, making the development of efficient fault early warning systems an urgent need for the wind power industry. This need has driven the adoption of edge computing technology—reducing reliance on the cloud and improving response speed and system reliability through localized data processing on the device side.

[0003] However, existing early warning methods have significant limitations: most rely on periodic checks or simple threshold alarms, making them difficult to adapt to dynamic operating environments. Natural fluctuations in parameters caused by external factors such as wind speed and temperature can easily lead to false alarms and missed alarms in threshold judgments; moreover, traditional methods require the transmission of massive amounts of raw data to the cloud for analysis, which increases the communication burden and may cause early warning delays due to network problems, a problem that is even more prominent in remote wind farms.

[0004] The core technological challenge in edge computing environments lies in the ability to dynamically analyze parameters. Parameters such as vibration and temperature in wind turbines fluctuate complexly with environmental and load changes, and their rate of change is crucial for assessing their health. However, edge nodes have limited computing resources, making high-precision rate of change analysis under low power consumption a primary challenge. More importantly, the parameter characteristics of equipment vary significantly under different operating conditions, and a single threshold or fixed rule cannot distinguish between normal fluctuations and abnormal trends—for example, increased vibration amplitude at high wind speeds is normal, but a continuously abnormally high rate of change may indicate a fault. Existing systems struggle to dynamically adapt to such complex conditions, leading to misjudgments and missed diagnoses. Therefore, how to accurately analyze parameter rate of change in real time and dynamically adapt to complex operating conditions with limited edge node resources is key to overcoming the bottleneck of wind power fault early warning technology. Summary of the Invention

[0005] This invention provides a method and system for early warning of wind power equipment faults based on edge computing, so as to achieve accurate fault prediction and improve the efficiency of wind power equipment operation and maintenance response and the accuracy of early warning.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a wind power equipment fault early warning method based on edge computing, comprising: The operation data of vibration, temperature and current parameters are collected by sensors configured on the wind power equipment to generate raw data sequences, and the raw data sequences are preprocessed to obtain a preliminary feature set. Based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and a gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence. When the parameter change rate sequence exceeds a preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. The system calls a preset historical operating condition database, performs feature matching between the stable change trend and the historical operating condition data, and uses a support vector machine to classify the anomaly types of the matched features to obtain the risk level under the current operating condition. The real-time resource status of edge nodes is obtained, and the risk level is matched and evaluated with the resource status. If it is determined that the edge node has sufficient resources, a deep analysis process is initiated to build a refined anomaly prediction model. Based on the refined anomaly prediction model, the evolution trend of potential abnormal fluctuations is predicted, and an early warning signal containing the failure probability and occurrence time is generated. Assess the risk level of the warning signal and activate the local response mechanism based on the risk level to output an immediate fault alarm.

[0007] In one optional implementation, the step of calculating the time series gradient corresponding to each parameter based on the preliminary feature set using time series analysis includes: Based on the vibration, temperature, and current parameters in the initial feature set, the acquisition time of each parameter is aligned by timestamps to establish the mapping relationship between vibration parameters, temperature parameters, current parameters and time. The operating data corresponding to each parameter of vibration, temperature, and current are sorted in chronological order to form a continuous single-parameter time series. Based on the single-parameter time series, by calculating the data difference between two adjacent moments and combining it with the time interval between the two corresponding moments, the gradient value of parameter change within the corresponding time period is obtained. Based on the parameter change gradient values ​​of each time period, they are integrated in chronological order to form the time series gradients corresponding to each parameter of vibration, temperature, and current.

[0008] In one optional implementation, the step of introducing a gradient descent optimization algorithm to iteratively optimize the time series gradient and determine the parameter change rate sequence includes: Using the time series gradient as the initial calculation result, a gradient descent optimization algorithm is introduced, and the iterative optimization objective is set as minimizing the deviation between the initial gradient and the gradient pattern of normal equipment operation. Retrieve the parameter gradient reference sequence when the equipment is running normally, and define the difference between the current time series gradient and the parameter gradient reference sequence as the gradient calculation error; According to the iterative rules of the gradient descent algorithm, the calculation parameters of the time series gradient are gradually adjusted based on the gradient calculation error; wherein, the calculation parameters include the weight coefficients of adjacent time data and the time interval conversion factor; After each parameter adjustment, the time series gradient is recalculated and the gradient calculation error is updated. If the error does not reach the preset optimization threshold, the parameter adjustment step is returned to continue the iteration. When the gradient calculation error of the iterative update meets the preset optimization threshold, the iteration stops, and the final optimized time series gradient is determined as the parameter change rate sequence.

[0009] In one optional implementation, the step of smoothing the parameter change rate sequence using a sliding window method to obtain a stable change trend includes: Using the parameter change rate sequence as the processing object, a sliding window of preset length is set; the sliding window is used to cover continuous parameter change rate data in the sequence; The sliding window is moved in chronological order so that it covers different data segments in the sequence in turn, and the average level of the parameter change rate data in each window is calculated. The average levels of each window are arranged sequentially according to the window movement order to form a smooth sequence that eliminates short-term fluctuations; By analyzing the overall trend of the smoothed sequence, the stable trend of the parameter change rate over time can be extracted.

[0010] In one optional implementation, the step of calling a preset historical operating condition database, matching the stable change trend with the historical operating condition data for features, and using a support vector machine to classify the matched features for anomaly types to obtain the risk level under the current operating condition includes: The system calls upon a pre-set historical operating condition database; the database contains parameter change characteristics, anomaly type records, corresponding risk level standards, and feature classification labels for different operating conditions. Extract the trend turning point, fluctuation cycle, and change amplitude as key features from the stable change trend to construct the feature vector of the current trend; The cosine similarity algorithm is used to calculate the similarity between the feature vector and the historical working condition features in the historical working condition database, and high-matching reference samples with similarity exceeding a preset similarity threshold are selected. The high-matching reference samples and the feature vectors of the current trend are fused to form a classification dataset, which is then input into a support vector machine classification model trained based on historical anomaly samples to output the current anomaly type. Based on the current anomaly type, the risk level standard in the historical operating condition database is queried. The risk is then adjusted by combining the differences in environmental parameters and equipment runtime between the current and historical operating conditions to determine the risk level under the current operating condition.

[0011] In one optional implementation, the real-time resource status of the edge node is obtained, and the risk level is matched and evaluated with the resource status. If it is determined that the edge node has sufficient resources, a deep analysis process is initiated to construct a refined anomaly prediction model, including: Obtain the real-time resource status of edge nodes; the resource status includes the current load level of computing units, the remaining capacity of storage space, the transmission capacity of network links, and the fluctuation of resource usage; The risk level is compared with the real-time resource status of the edge nodes in multiple dimensions to assess whether the edge nodes have the resource redundancy to carry out in-depth analysis tasks. If the assessment results indicate that the edge node resources are sufficient, that is, the computing load is in a low occupancy range, the remaining storage capacity meets the data processing needs, and the network transmission capability is stable and redundant, then the deep analysis process will be initiated. Extract short-term fluctuation characteristics, long-term trend characteristics, and parameter correlation characteristics of equipment operation, and integrate them to form a multi-dimensional analysis dataset; A refined anomaly prediction model is constructed based on the multi-dimensional analysis dataset; the refined anomaly prediction model includes parameter change thresholds, anomaly evolution paths, and correlation influence rules under different risk levels.

[0012] In one optional implementation, constructing a refined anomaly prediction model based on the multi-dimensional analysis dataset includes: Extract short-term fluctuation features, long-term trend features, and parameter correlation features from the multi-dimensional analysis dataset to determine the key feature dimensions of the model input. Based on the physical characteristics of wind power equipment operation, a model is established to identify abnormal evolution paths; the identification rules include the correlation of parameter changes and the logic of abnormal propagation. Based on the key feature dimensions and identification rules, an anomaly prediction model framework that can output the probability of failure and the time of occurrence is constructed. The anomaly prediction model framework is trained and optimized using historical anomaly case data stored at edge nodes to form a refined anomaly prediction model that reflects the correspondence between parameter changes and fault occurrence.

[0013] In one optional implementation, the step of predicting the evolution trend of potential abnormal fluctuations based on the refined anomaly prediction model and generating an early warning signal containing the failure probability and occurrence time includes: The short-term fluctuation characteristics, long-term trend characteristics, and parameter linkage characteristics of equipment operation are integrated into a multi-dimensional analysis dataset, which is then input into the constructed fine anomaly prediction model. The refined anomaly prediction model analyzes the correlation between features and the logic of parameter changes based on the identification rules of anomaly evolution paths, and extrapolates the development trend of potential abnormal fluctuations, outputting the extrapolation results. Based on the simulation results, assess the likelihood of the abnormal fluctuations causing equipment failure and determine the failure probability; By combining the current state and evolution pattern of abnormal fluctuations, analyze the time required for development to reach the fault state and determine the time of fault occurrence; The fault probability, fault occurrence time, and corresponding anomaly type information are integrated to form the early warning signal.

[0014] In one optional implementation, assessing the risk level of the warning signal and activating a local response mechanism based on the risk level to output an immediate fault alarm includes: Based on the preset risk assessment criteria, the probability of failure and the expected occurrence time are extracted from the early warning signal, the current risk status is assessed, and the risk level is output. Based on the risk level, a corresponding local response mechanism is matched; the response mechanism includes the alarm notification scope, equipment emergency handling methods, and operation and maintenance cooperation strategies. According to the local response mechanism, the corresponding response process is initiated; the response process includes pushing alarms to relevant terminals, triggering device protection mode, and contacting maintenance personnel. The risk level, the probability of failure in the warning signal, the time of failure occurrence, and the response process are integrated into standard alarm content and output as the fault alarm.

[0015] Secondly, the present invention also provides a wind power equipment fault early warning system based on edge computing, comprising: Data acquisition module: Collects operational data of vibration, temperature, and current parameters through sensors configured on the wind power equipment, generates raw data sequences, and preprocesses the raw data sequences to obtain a preliminary feature set; Feature analysis module: Based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and a gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence; Smoothing module: When the parameter change rate sequence exceeds the preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. Risk analysis module: Calls a preset historical operating condition database, performs feature matching between the stable change trend and the historical operating condition data, and uses a support vector machine to classify the anomaly types of the matched features to obtain the risk level under the current operating condition; Model building module: Obtain the real-time resource status of edge nodes, match and evaluate the risk level with the resource status, and if it is determined that the edge node has sufficient resources, start the deep analysis process to build a refined anomaly prediction model. Early warning module: Based on the refined anomaly prediction model, it predicts the evolution trend of potential abnormal fluctuations and generates an early warning signal containing the failure probability and occurrence time. Fault assessment module: assesses the risk level of the warning signal and activates the local response mechanism based on the risk level to output an immediate fault alarm.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By collecting multi-parameter operating data such as vibration, temperature and current through sensors and generating raw sequences, and combining them with preprocessing to form a preliminary feature set, a comprehensive perception of the operating status of wind power equipment is realized. Multi-dimensional parameters cover the core operating indicators of the equipment. Data cleaning and standardization filter out sensor anomalies and transmission interference, providing high-quality data support for subsequent analysis. This solves the problems of single traditional monitoring data and large noise interference, and improves the reliability of early warning data.

[0017] (2) Based on time series analysis, the parameter gradient is calculated, and the rate of change sequence is determined through iterative optimization using the gradient descent algorithm. Combined with sliding window smoothing, a stable trend is obtained, thus achieving accurate identification of abnormal fluctuations. The algorithm optimization eliminates the interference of natural parameter fluctuations, and the sliding window technique removes short-term noise, accurately capturing the essential trend of parameter changes. This avoids false alarms and missed alarms caused by simple threshold judgments, and improves the accuracy of anomaly identification.

[0018] (3) By calling the historical operating condition database for feature matching, determining the risk level through support vector machine classification, and launching in-depth analysis in conjunction with edge resource assessment, operating condition adaptation and hierarchical early warning are realized. Historical data matching adapts to different operating scenarios, machine learning classification accurately classifies anomaly types, and edge resource dynamic assessment balances analysis accuracy and response speed, solving the problems of traditional early warning lack of operating condition adaptability and low analysis efficiency.

[0019] (4) A refined anomaly prediction model is constructed to generate early warning signals containing fault probability and occurrence time. The risk level is matched to activate the local response mechanism, which realizes accurate prediction and rapid handling of faults. The prediction model quantifies the fault evolution trend, the hierarchical response mechanism adapts to different risk scenarios, and the edge localization processing shortens the early warning lag time, which greatly reduces equipment downtime losses and maintenance costs and improves operation and maintenance response efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a wind power equipment fault early warning method based on edge computing provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a wind power equipment fault early warning system based on edge computing provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1 This invention provides a wind power equipment fault early warning method based on edge computing, comprising the following steps: S11, collect operational data of vibration, temperature, and current parameters through sensors configured on the wind power equipment, generate a raw data sequence, and preprocess the raw data sequence to obtain a preliminary feature set; S12, based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence. S13, when the parameter change rate sequence exceeds the preset change threshold, it is determined to be a potential abnormal fluctuation, and the sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. S14, call the preset historical working condition database, match the stable change trend with the historical working condition data, and use support vector machine to classify the anomaly type of the matched features to obtain the risk level under the current working condition. S15, obtain the real-time resource status of the edge node, match and evaluate the risk level with the resource status, and if it is determined that the edge node has sufficient resources, start the deep analysis process and build a refined anomaly prediction model. S16, Based on the refined anomaly prediction model, predict the evolution trend of potential abnormal fluctuations and generate an early warning signal containing the failure probability and occurrence time. S17, assess the risk level of the warning signal, and activate the local response mechanism according to the risk level to output an immediate fault alarm.

[0023] In step S11, the operating data of vibration, temperature and current parameters are collected by the sensors configured on the wind power equipment to generate a raw data sequence, and the raw data sequence is preprocessed to obtain a preliminary feature set.

[0024] In one embodiment, this example uses a 2MW direct-drive wind turbine generator set in an onshore wind farm as the application object, and specifically implements the above-mentioned data acquisition and preprocessing process: Based on the key fault-prone parts of the equipment, one triaxial vibration acceleration sensor is installed at each of the generator's main shaft bearing (radial vibration sensitive point) and the generator front end cover (axial vibration monitoring point). Magnetic installation ensures stability, with a sampling frequency of 800Hz (adapted to the equipment's main operating frequency) and a measurement range of 0-40g (covering the range from normal operation to minor faults); two platinum resistance temperature sensors are installed at each of the gearbox oil sump and generator windings, with a measurement range of -50℃ to 150℃ and an accuracy of ±0.3℃; one Hall current sensor is installed at the low-voltage side outgoing terminal of the main transformer, with a measurement range of 0-600A and an accuracy of ±0.5%.

[0025] All sensors communicate with the edge computing node in the middle of the tower via the industrial-grade Modbus-RTU protocol. The node uses an Intel Atom processor (low power consumption adapted to edge scenarios), and is equipped with 8GB of memory and 128GB of solid-state storage. The sensors synchronously collect data at a fixed interval of 100ms. The node encapsulates the data in the format of "collection timestamp-vibration X-axis-vibration Y-axis-vibration Z-axis-gearbox temperature 1-gearbox temperature 2-generator temperature 1-generator temperature 2-output current" to generate the raw data sequence.

[0026] Edge node preprocessing: Electromagnetic interference jump values ​​from vibration sensors (beyond 1.5 times the interquartile range) were removed using box plotting; two missing values ​​≤300ms were filled using the nearest neighbor mean method; features were extracted in 2-minute windows—statistical features included vibration mean, temperature variance, and current extreme values; time-domain features included vibration peak value and temperature fluctuation amplitude; min-max standardization mapped the features to the [0,1] interval to eliminate dimensional differences. Finally, 18 vibration features, 16 temperature features, and 5 current features were integrated to form a preliminary 39-dimensional feature set, providing standardized data support for subsequent time series gradient calculations.

[0027] In step S12, based on the preliminary feature set, the time series gradient corresponding to each parameter is calculated using time series analysis, and the gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence.

[0028] In one implementation, building upon the 39-dimensional preliminary feature set established earlier, the edge computing node initiates a parameter change rate analysis process. This process, tailored to the "low latency and lightweight" characteristics of edge nodes, comprises two main steps: first, calculating the first-order time series gradient (i.e., the rate of change per unit time) of each parameter based on time series analysis theory to establish the basic framework for change rate analysis; and second, introducing a lightweight gradient descent optimization algorithm to iteratively correct the initial gradient, ultimately determining an accurate parameter change rate sequence while controlling computational resource consumption, thus providing a reliable basis for subsequent anomaly identification.

[0029] It is worth noting that the specific implementation in the time series gradient calculation stage is as follows: Based on the core principle of "temporal correlation" of time series data, with vibration, temperature, and current as the three major categories of parameters in the initial feature set as the core of analysis, the time series alignment of each parameter is first completed by collecting timestamps. Since the sensor acquisition interval is 100ms, all feature data are precisely matched with "millisecond-level timestamps" to establish three independent mapping relationships: "vibration parameter-time", "temperature parameter-time", and "current parameter-time". For example, the vibration X-axis mean feature corresponds to the time series mapping of "14:00:00.000→0.28, 14:00:00.100→0.31...", ensuring the consistency of data in the time dimension. Subsequently, the three types of parameter features are sorted in chronological order: the vibration category includes 3 axes × 6 types of features (mean, variance, extreme values, peak values, etc.) with a total of 18 subsequences; the temperature category includes 4 sensors × 4 types of features with a total of 16 subsequences; and the current category includes 5 types of features with a total of 5 subsequences, forming a structured single-parameter time series library. For each subsequence, the "difference-time ratio" calculation logic is used to obtain the parameter change gradient value for the corresponding time period using the following formula: , where r i,j Let x represent the time series gradient (parameter rate of change) of the j-th subsequence of the i-th type of parameter (vibration / temperature / current). i,j (t) represents the characteristic value at time t, x i,j (t+Δt) represents the feature value at time t+Δt, where Δt is a fixed acquisition interval (Δt=100ms in this embodiment). For example, when the temperature feature changes from 0.42 to 0.45, r=(0.45-0.42) / 0.1=0.3. Finally, the gradient values ​​of all time periods are integrated in chronological order to form the time series gradient specific to each parameter, completing the transformation from "static feature data" to "dynamic gradient data," thus meeting the fault early warning analysis requirements for "parameter change trends."

[0030] It is worth noting that the implementation process in the gradient descent optimization stage is as follows: Using the calculated time-series gradient as the initial result, a lightweight gradient descent algorithm adapted to the edge nodes is introduced. The core optimization objective is set as "minimizing the deviation between the initial gradient and the gradient pattern of normal equipment operation," while controlling the computation time of a single iteration to not exceed 50ms. The parameter gradient reference sequence G of the 2MW direct-drive unit under rated load (2MW) and normal temperature (25℃) is retrieved from the historical operating condition database stored locally on the edge nodes. ref =[g ref,1 ,g ref,2 ,...,g ref,n The vibration parameter gradient reference range is ±0.15, the temperature parameter is ±0.08, and the current parameter is ±0.12. The current gradient sequence... G curr =[ g curr,1 , g curr,2 ,..., g curr,n The Euclidean distance from the reference sequence is defined as the gradient calculation error, and the formula is: Where E is the gradient calculation error, n is the sequence length, and g curr,k For the k-th value of the current gradient sequence, g ref,kThe smaller the error value, the closer the gradient is to the normal operating pattern of the equipment. Based on this error, two key calculation parameters are adjusted: the weighting coefficient of adjacent time-series data (initially set to 0.5, giving equal importance to data from consecutive time-series moments), and the time interval conversion factor (initially 1.0, adapting to a fixed acquisition interval). The iterative adjustment rules are set according to the "error direction - parameter correction" correspondence. The adjustment magnitude is calculated based on the learning rate (η=0.01) and parameter sensitivity coefficient of the gradient descent algorithm: weighting coefficient adjustment = η × time dependence coefficient (0.01 × 5 = 0.05), time interval conversion factor adjustment = η × fluctuation suppression coefficient (0.01 × 3 = 0.03). These coefficients are calibrated using 100 sets of historical gradient data to ensure that each iteration conforms to the convergence logic. If the error is positive and exceeds the preset threshold of 0.02, it indicates that the current gradient is too high. The weight of the data from the previous time step needs to be increased (by 0.05 = 0.01 × 5 per round) and the time interval conversion factor needs to be decreased (by 0.03 = 0.01 × 3 per round) to strengthen the constraints of historical data and weaken the impact of short-term fluctuations. If the error is negative and exceeds the threshold, the parameters are adjusted in the opposite direction. After each parameter adjustment, the time series gradient is immediately recalculated and the error value is updated. If the error still does not fall below the threshold, the iteration returns to the parameter adjustment step. After 12 iterations, the gradient calculation error decreased from the initial 0.08 to 0.018, meeting the preset requirements. The iteration then stops, and the final optimized time series gradient is determined as the parameter change rate sequence—the optimized vibration parameter change rate sequence stabilizes in the 0.12-0.16 range, the temperature parameter in the 0.05-0.07 range, and the current parameter in the 0.09-0.11 range, effectively eliminating random interference components in the initial gradient and improving the reliability of the change rate data.

[0031] In step S13, when the parameter change rate sequence exceeds a preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend.

[0032] In one implementation, following the parameter change rate sequence (i.e., the optimized time series gradient) obtained through gradient descent optimization, the edge computing node, considering the continuous operation characteristic of wind power equipment's parameter fluctuations, immediately executes an abnormal fluctuation judgment and trend smoothing process: First, it retrieves a preset multi-dimensional change threshold system from the local historical operating condition database—this threshold is based on the parameter fluctuation range under the equipment's rated operating conditions, where the vibration parameter change rate threshold is ±0.2 (adapting to the normal fluctuation range of mechanical vibration), the temperature parameter is ±0.1 (matching the temperature change amplitude under the equipment's thermal equilibrium state), and the current parameter is ±0.15 (conforming to the current change range corresponding to load fluctuations), and a judgment condition of "three consecutive acquisition intervals exceeding the threshold" is set to avoid misjudgment due to instantaneous interference. Through real-time comparison, it is found that the change rate sequence of gearbox temperature 1 reaches 0.18 for three consecutive 100ms acquisition intervals, significantly exceeding the ±0.1 temperature threshold, meeting the abnormal judgment condition, and is immediately judged as a potential abnormal fluctuation. The sliding window smoothing process is then initiated, specifically executing the logic of covering data according to a preset length window, calculating the mean, and extracting the trend.

[0033] In practice, the gearbox temperature change rate sequence is used as the processing object. Considering the 100ms acquisition frequency and the initial "small fluctuations, overall upward trend" characteristic of the anomaly, a sliding window containing 5 consecutive data points is set. Based on historical data statistics, sensor transient interference typically lasts for 1-2 acquisition intervals (100-200ms). Therefore, a window length of 5 data points (500ms) can filter short-term transient jumps while preserving trend details, avoiding a trend lag exceeding 100ms due to an excessively long window (e.g., 7 points). The sliding window, with a step size of 1 data point, sequentially covers all data segments from the anomaly initiation time (the first time the change rate exceeds 0.1) to the present. The arithmetic mean is used to calculate the center value of the 5 change rate data within each window. For example, if the data in the first window are 0.18, 0.19, 0.17, 0.20, and 0.18, the average value is calculated to be 0.184, thus mitigating the impact of single-point outliers.

[0034] The average values ​​of all windows were arranged sequentially according to the window movement order to form a smoothed sequence that eliminated the interference of instantaneous jumps. The sudden peak value of 0.20 in the original sequence was reduced to about 0.184 after smoothing, and the data fluctuation range narrowed from 0.03 to 0.015. Through trend fitting analysis of the smoothed sequence, it can be seen that the rate of change of gearbox temperature 1 gradually climbed from 0.12 before the anomaly to 0.18, and then maintained a stable increase of about 0.01 every 10 minutes. This accurately extracted the stable trend of "continuous increase with a slow growth rate", effectively removing random noise from the original sequence and providing a clear and reliable trend basis for subsequent feature matching and anomaly classification.

[0035] In step S14, a preset historical operating condition database is invoked, the stable change trend is matched with the historical operating condition data, and a support vector machine is used to classify the matched features into anomaly types to obtain the risk level under the current operating condition.

[0036] In one implementation, following the stable trend of the extracted gearbox temperature 1 "continuously rising with a gradual increase," the edge computing node initiates a feature matching and anomaly classification process. First, it calls a locally pre-set historical operating condition database. This database stores the 2MW unit's full-condition operating data for the past three years, archived in a tiered manner ("normal - minor anomaly - fault"), covering more than 20 anomaly types such as "gradual temperature rise due to gearbox oil contamination" and "sudden temperature rise due to cooling system blockage." Each type is labeled with a risk level standard and feature tags of "trend slope - fluctuation period - change amplitude," adapting to the feature matching requirements of temperature-related anomalies.

[0037] Based on the current temperature change trend, key information is extracted from the core characteristics of the "gradually rising anomaly": the trend inflection point is when the rate of change increases from 0.12 to 0.15 (corresponding to the critical state of the anomaly's inception); the fluctuation period is determined to be "aperiodic" due to the absence of reciprocating fluctuations (quantitative standard: a fluctuation period > 30 minutes is considered "aperiodic," corresponding to a value of 0); and the change amplitude is the difference of 0.06 from the start of the anomaly to the present (reflecting the intensity of the anomaly's development). Based on this, a three-dimensional feature vector of the current trend is constructed. (Corresponding to turning points, fluctuation cycles, and amplitudes of change, respectively). The cosine similarity algorithm is used to calculate the similarity between this vector and historical feature vectors in the database. The core calculation formula is: Where cosθ is the similarity value (ranging from [0,1], with the value closer to 1 indicating a higher degree of feature matching). Here, n represents the three-dimensional feature vector of historical operating conditions (n=3 in this embodiment). The three selected high-matching reference samples are all cases of "gradual temperature rise due to decrease in gearbox oil viscosity," and their historical feature vectors are as follows: (Based on a case from August 2022, ambient temperature 25℃, operating time 12,000 hours) (Based on a case from March 2023, ambient temperature 26℃, operating time 15,000 hours) (Based on a case from October 2023, ambient temperature 24℃, operating time 14,000 hours). For example, the similarity calculated by the formula reached 0.88, 0.91 and 0.87 respectively, all of which exceeded the preset similarity threshold of 0.85.

[0038] The feature data of three reference samples, such as "inflection point 0.14 - no period - amplitude 0.07", are fused with the current feature vector "inflection point 0.15 - no period - amplitude 0.06" to form a 1×6 dimensional classification dataset, which is then input into a locally pre-trained support vector machine classification model. This model is specifically optimized for wind power temperature anomaly scenarios. It uses 1000 sets of historical anomaly samples covering types such as oil contamination and cooling system failures as training data. A radial basis function kernel function is used to adapt the nonlinear feature relationship, and the penalty parameter is set to 10 to balance the model's fit and generalization ability, enabling it to accurately distinguish temperature anomalies from different sources.

[0039] The model maps low-dimensional classification data to a high-dimensional feature space using a kernel function, automatically constructing an optimal separating hyperplane to classify anomalies. The final output anomaly type is "abnormal temperature rise caused by a decrease in gearbox oil viscosity." A query of the historical operating condition database reveals that the baseline risk level for this anomaly type is "medium risk." Adjustments are made based on differences in operating conditions: the current ambient temperature is 5°C higher than historical samples (accelerating oil degradation), and the equipment operating time is 1200 hours longer than historical samples (component wear exacerbates the anomaly's impact). According to the preset adjustment rule of "risk increases by 1 level for every 5°C increase in ambient temperature and 1 level increase in operating time exceeding 1000 hours," the baseline risk level is increased by one level, ultimately determining the current risk level as "higher risk" (this is a preliminary assessment, to be further validated using early warning signals). This provides accurate risk information for subsequent resource assessment and in-depth analysis.

[0040] In step S15, the real-time resource status of the edge node is obtained, and the risk level is matched and evaluated with the resource status. If it is determined that the edge node has sufficient resources, the deep analysis process is started to construct a refined anomaly prediction model.

[0041] In one implementation, after determining the current risk level to be "high risk," the edge computing node initiates a resource assessment and deep analysis process. First, the node's built-in resource monitoring module collects real-time data on four core resource statuses: CPU utilization is statistically analyzed using a 1-minute sliding window, with a current average of 32% (preset low utilization threshold ≤ 40%), and process scheduling latency is stable within 5ms; memory utilization is 28%, with 1.8GB remaining in the reserved analysis-dedicated memory pool (preset sufficient threshold ≥ 1GB); storage space is queried via the file system interface, with 45GB remaining (deep analysis requires caching nearly 24 hours of high-sampling-rate raw data, expected to occupy 18GB, sufficient threshold ≥ 20GB); network links are evaluated using a combination of ICMPping and throughput testing, with communication latency with the local data center stable at 20ms, transmission rate maintained at 12Mbps (intermediate results to be uploaded during analysis are approximately 500MB, sufficient threshold ≥ 10Mbps), and resource utilization fluctuation ≤ 5% (stable threshold ≤ 10%).

[0042] Based on the analysis complexity matrix corresponding to "higher risk," a multi-dimensional weighted evaluation method is used to determine resource adequacy: weight allocation is based on the "low latency priority" characteristic of edge computing, with computing resources (40%, affecting analysis speed) > storage (25%, affecting data caching) > network (20%, affecting result transmission) > fluctuation stability (15%, affecting process stability). The comprehensive score is calculated as "actual value / adequacy threshold × weight percentage": computing resource score = 32% / 40% × 40 = 32, storage score = 28% / 50% × 25 = 14 (memory and storage are evaluated together, with a preset comprehensive adequacy threshold ≤ 50%), network score = 12Mbps / 10Mbps × 20 = 24, fluctuation stability score = 5% / 10% × 15 = 7.5, and total score = 32 + 14 + 24 + 7.5 = 77.5 points (≥ 70 points is considered sufficient). This confirms that the edge nodes have resource redundancy, and the deep analysis process is then initiated.

[0043] In the in-depth analysis phase, features were extracted layer by layer according to time granularity: short-term fluctuation features (1-hour data) included three micro-oscillation peaks of gearbox temperature within 5 minutes (0.03℃) and the proportion of high-frequency components on the X-axis of vibration (12%); long-term trend features (7-day data) covered the average daily temperature growth rate (0.8℃ / day) and the periodicity coefficient of vibration amplitude (0.72); parameter linkage features were obtained through cross-correlation analysis, such as the lag time of temperature rise after a 10A increase in current (approximately 8 seconds) and the peak frequency of the cross-power spectrum of vibration and current (50Hz). These features were integrated into a 42-dimensional analysis dataset, including 35 numerical features and 7 logistic features, providing multi-dimensional input for model construction.

[0044] It is worth noting that the implementation process in the construction of the refined anomaly prediction model is as follows: By ranking features by importance (based on the Gini coefficient of random forest), 18 key feature dimensions were selected from the 42-dimensional dataset, including core indicators such as the short-term fluctuation frequency of gearbox temperature (weight 0.12), the long-term growth rate of vibration amplitude (weight 0.11), and the current-temperature linkage coefficient (weight 0.09). Combined with the physical characteristics of wind power equipment, three types of anomaly evolution identification rules were set: first, parameter association rules (e.g., "temperature change rate > 0.15 for 2 hours → vibration amplitude ≥ 1.2 times the baseline"); second, diffusion path rules (the propagation sequence of "gearbox anomaly → generator stator temperature rise → converter current fluctuation"); and third, environmental impact rules ("for every 10℃ increase in ambient temperature, the anomaly evolution speed increases by 15%").

[0045] Based on the above features and rules, a three-layer model framework of "feature preprocessing - LSTM time series prediction - rule constraint correction" is constructed. The input layer embeds "parameter change threshold" (e.g., temperature change rate ≥ 0.15 triggers deep inference), the hidden layer learns "abnormal evolution path" through LSTM network, and the output layer connects "correlation influence rules" (i.e. rule constraint correction module): the input layer performs standardization processing on 18-dimensional features (Z-score transformation); the hidden layer adopts a 3-layer LSTM network (64 neurons in the first layer, 32 neurons in the second layer, and 16 neurons in the third layer) to capture time dependence through gating mechanism; the output layer is designed with a dual-branch structure to output the fault probability (Sigmoid activation) and the expected occurrence time (linear activation) respectively. The model was trained using 500 historical cases of "gearbox oil anomalies" (including 200 complete fault evolution processes) stored at edge nodes. The Adam optimizer was used with an initial learning rate of 0.001, decaying by 10% every 50 iterations. After 200 iterations, the fault occurrence time prediction error on the validation set stabilized within ±30 minutes, and the F1-score for fault probability determination reached 0.92. Finally, a rule engine was used to refine the model output. For example, when the ambient temperature was 5°C higher than the historical case average, the predicted occurrence time was shortened by 10%, ultimately forming a refined anomaly prediction model that combines data-driven approaches with physical rule constraints.

[0046] In step S16, the evolution trend of potential abnormal fluctuations is predicted based on the refined anomaly prediction model, and an early warning signal containing the failure probability and occurrence time is generated.

[0047] In one implementation, the edge computing node initiates evolution trend prediction according to a preset process, connecting the refined anomaly prediction model constructed above. First, from the 42-dimensional multi-dimensional analysis dataset constructed in step S15, 18 key features (including short-term fluctuations, long-term trends, and parameter linkage features) are extracted and integrated to form the model input dataset. These 18 key features specifically include core indicators such as the 5-minute oscillation peak of gearbox temperature, the high-frequency proportion of vibration on the X-axis, and the average daily temperature growth rate. They are input into the model in a fixed order of "short-term features - long-term features - linkage features." Simultaneously, the model input layer performs standardization processing (Z-score transformation) on these 18 key features. By calling the feature mean and standard deviation transformation data saved during the training phase, the distribution of the input data is ensured to remain consistent with that during model training, meeting the basic requirements for model operation.

[0048] The model first invokes pre-defined rules for identifying abnormal evolution paths, including core criteria such as "vibration amplitude increases over time after the temperature change rate exceeds a threshold" and "the hysteresis correlation between current and temperature." Then, a three-layer LSTM network with hidden layers deeply analyzes the correlation between features and the logic of parameter changes: the network calls optimized weight matrices, bias values, and other parameters, focusing on the growth curve of the gearbox temperature change rate from 0.12 to 0.18. Combining this with the "hysteresis feature of temperature increase after current increase," the model simulates the parameter change trajectory through a 200-step time-series extrapolation, outputting the extrapolation result that "the temperature change rate increases by 0.01 per hour, and the vibration amplitude increases by 0.05g every 2 hours." This extrapolation result, combined with the annotation rule from model training that "a fault is triggered when the temperature change rate ≥ 0.25g and the vibration amplitude ≥ 0.5g," provides a quantitative basis for subsequent calculations of fault probability and occurrence time.

[0049] Based on the deduction of "continuous deterioration of temperature and vibration parameters," the model output layer uses the Sigmoid activation function for quantitative evaluation: quantifying the "evolution from the current rate of change of 0.18 to the critical value of 0.25" as an 87% probability of failure, accurately reflecting the possibility of abnormal fluctuations causing equipment failure. Simultaneously, taking the current state of abnormal fluctuations as the starting point of analysis, specifically including key parameters such as the temperature change rate of 0.18 and vibration amplitude of 0.32g; combined with the deduced evolutionary pattern—that is, the hourly temperature change rate increases by 0.01 and the vibration amplitude increases synchronously—the model's built-in linear regression branch calculates the "time difference from the current state to the failure state"—based on the deduction that "it takes 7 hours for the temperature change rate to reach 0.25 and 7.2 hours for the vibration amplitude to reach 0.5g," taking the shorter time and combining it with rule corrections, the preliminary estimate of the failure time after time series calculation is 6 hours and 20 minutes after the current moment.

[0050] The rules engine was activated to correct the initial prediction results—because the current ambient temperature is higher than the historical average, the estimated occurrence time was shortened according to preset rules; because the actual equipment load is higher than the average, the failure probability was adjusted upwards by querying the correction matrix, ultimately determining a failure probability of 90% and an estimated occurrence time of 5 hours and 54 minutes. These two core results were integrated with the corresponding anomaly type information (gearbox oil viscosity decrease causing abnormal temperature rise), and supplemented with fields such as equipment identification, current key parameters, correction instructions, and model confidence level to form a standardized early warning signal, providing complete data support for subsequent risk response.

[0051] The entire process follows a logical progression of "feature input - situational deduction - probability and time determination - signal encapsulation". Each step is closely aligned with the core requirements of anomaly prediction and early warning, ensuring the accurate implementation of the technical solution.

[0052] In step S17, the risk level of the warning signal is assessed, and the local response mechanism is activated based on the risk level to output an immediate fault alarm.

[0053] In one implementation, following the generated early warning signal, the edge computing node initiates a risk assessment and response process. First, a precise risk level verification is performed: Connecting the initial "higher risk" level determined in step S14, a quantitative verification is conducted using the "failure probability 90%, estimated occurrence time 5 hours 54 minutes" from the early warning signal—comparing it to the locally stored risk level matrix (divided into three dimensions: "failure probability + estimated occurrence time + scope of abnormal impact"). The scenario of "core component abnormality (gearbox) + failure probability 80%-95% + estimated occurrence time 3-8 hours" clearly corresponds to "higher risk." Since the two determinations are consistent, the risk level is confirmed to remain "higher risk."

[0054] According to the pre-set local response mechanism matched to this level: the alarm notification scope is limited to the main control system of the wind farm operation and maintenance center, the mobile terminal of the on-duty engineer, and the area manager to ensure full coverage of key positions; the emergency handling of equipment adopts the "protection first" strategy, activates the enhanced mode of the gearbox cooling system (fan speed increased by 30%), and limits the unit output power to 80% of the rated value to reduce the load on components by reducing load; the operation and maintenance cooperation strategy specifies time nodes, requiring the allocation of spare parts such as gearbox special lubricating oil within 2 hours, and the operation and maintenance team to arrive at the site for maintenance within 4 hours.

[0055] The edge node then initiates a response process according to the priority order of "equipment protection first, notification follow-up": ① Triggering the local PLC controller to execute cooling enhancement and power limiting commands (completed within 100ms, prioritizing equipment safety and preventing the escalation of the anomaly); ② Synchronously pushing structured alarm data (including real-time curves and predicted trend charts of abnormal parameters) to the main control system of the operation and maintenance center via the MQTT protocol, and sending SMS and APP push notifications with GPS positioning to the engineer's terminal (completed within 2 seconds, ensuring that operation and maintenance personnel can quickly obtain information); ③ Calling the voice notification interface to dial the on-duty engineer and the area manager (completed within 5 seconds, strengthening the response of key personnel). All three actions are completed within 10 seconds, and the status panel at the equipment tower base displays "early warning mode - reduced load operation" in real time for on-site personnel to view intuitively.

[0056] Finally, a standardized alarm content was formed: the title is "High Risk Warning for 2MW-Direct Drive-07 Unit," and the body of the text lists the risk level, failure probability (90%), estimated occurrence time, anomaly type, current key parameters (gearbox temperature 38℃, vibration amplitude 0.32g), initiated emergency measures (enhanced cooling, power limitation), and recommended maintenance actions (on-site repair within 4 hours). The alarm is simultaneously released through three channels: local audible and visual alarms at edge nodes (flashing red warning light + intermittent buzzer), tower base display screen, and remote terminals, achieving full coverage of "local alerts + remote notifications" to ensure rapid response and handling by relevant parties.

[0057] refer to Figure 2 The second embodiment of the invention provides a wind power equipment fault early warning system based on edge computing, comprising: Data acquisition module: Collects operational data of vibration, temperature, and current parameters through sensors configured on the wind power equipment, generates raw data sequences, and preprocesses the raw data sequences to obtain a preliminary feature set; Feature analysis module: Based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and a gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence; Smoothing module: When the parameter change rate sequence exceeds the preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. Risk analysis module: Calls a preset historical operating condition database, performs feature matching between the stable change trend and the historical operating condition data, and uses a support vector machine to classify the anomaly types of the matched features to obtain the risk level under the current operating condition; Model building module: Obtain the real-time resource status of edge nodes, match and evaluate the risk level with the resource status, and if it is determined that the edge node has sufficient resources, start the deep analysis process to build a refined anomaly prediction model. Early warning module: Based on the refined anomaly prediction model, it predicts the evolution trend of potential abnormal fluctuations and generates an early warning signal containing the failure probability and occurrence time. Fault assessment module: assesses the risk level of the warning signal and activates the local response mechanism based on the risk level to output an immediate fault alarm.

[0058] It should be noted that the edge computing-based wind power equipment fault early warning system provided in this embodiment of the invention is used to execute all the process steps of the edge computing-based wind power equipment fault early warning method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0059] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for early warning of wind power equipment faults based on edge computing, characterized in that, include: The operation data of vibration, temperature and current parameters are collected by sensors configured on the wind power equipment to generate raw data sequences, and the raw data sequences are preprocessed to obtain a preliminary feature set. Based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and a gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence. When the parameter change rate sequence exceeds a preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. The system calls a preset historical operating condition database, performs feature matching between the stable change trend and the historical operating condition data, and uses a support vector machine to classify the anomaly types of the matched features to obtain the risk level under the current operating condition. The real-time resource status of edge nodes is obtained, and the risk level is matched and evaluated with the resource status. If it is determined that the edge node has sufficient resources, a deep analysis process is initiated to build a refined anomaly prediction model. Based on the refined anomaly prediction model, the evolution trend of potential abnormal fluctuations is predicted, and an early warning signal containing the failure probability and occurrence time is generated. Assess the risk level of the warning signal and activate the local response mechanism based on the risk level to output an immediate fault alarm.

2. The wind power equipment fault early warning method based on edge computing according to claim 1, characterized in that, The step of calculating the time series gradient corresponding to each parameter based on the preliminary feature set using time series analysis includes: Based on the vibration, temperature, and current parameters in the initial feature set, the acquisition time of each parameter is aligned by timestamps to establish the mapping relationship between vibration parameters, temperature parameters, current parameters and time. The operating data corresponding to each parameter of vibration, temperature, and current are sorted in chronological order to form a continuous single-parameter time series. Based on the single-parameter time series, by calculating the data difference between two adjacent moments and combining it with the time interval between the two corresponding moments, the gradient value of parameter change within the corresponding time period is obtained. Based on the parameter change gradient values ​​of each time period, they are integrated in chronological order to form the time series gradients corresponding to each parameter of vibration, temperature, and current.

3. The wind power equipment fault early warning method based on edge computing according to claim 1, characterized in that, The introduction of the gradient descent optimization algorithm to iteratively optimize the time series gradient and determine the parameter change rate sequence includes: Using the time series gradient as the initial calculation result, a gradient descent optimization algorithm is introduced, and the iterative optimization objective is set as minimizing the deviation between the initial gradient and the gradient pattern of normal equipment operation. Retrieve the parameter gradient reference sequence when the equipment is running normally, and define the difference between the current time series gradient and the parameter gradient reference sequence as the gradient calculation error; According to the iterative rules of the gradient descent algorithm, the calculation parameters of the time series gradient are gradually adjusted based on the gradient calculation error; wherein, the calculation parameters include the weight coefficients of adjacent time data and the time interval conversion factor; After each parameter adjustment, the time series gradient is recalculated and the gradient calculation error is updated. If the error does not reach the preset optimization threshold, the parameter adjustment step is returned to continue the iteration. When the gradient calculation error of the iterative update meets the preset optimization threshold, the iteration stops, and the final optimized time series gradient is determined as the parameter change rate sequence.

4. The wind power equipment fault early warning method based on edge computing according to claim 1, characterized in that, The step of smoothing the parameter change rate sequence using a sliding window method to obtain a stable change trend includes: Using the parameter change rate sequence as the processing object, a sliding window of preset length is set; the sliding window is used to cover continuous parameter change rate data in the sequence; The sliding window is moved in chronological order so that it covers different data segments in the sequence in turn, and the average level of the parameter change rate data in each window is calculated. The average levels of each window are arranged sequentially according to the window movement order to form a smooth sequence that eliminates short-term fluctuations; By analyzing the overall trend of the smoothed sequence, the stable trend of the parameter change rate over time can be extracted.

5. The wind power equipment fault early warning method based on edge computing according to claim 1, characterized in that, The process involves calling a preset historical operating condition database, matching the stable trend with historical operating condition data for features, and using a support vector machine to classify the matched features for anomaly types to obtain the risk level under the current operating condition, including: The system calls upon a pre-set historical operating condition database; the database contains parameter change characteristics, anomaly type records, corresponding risk level standards, and feature classification labels for different operating conditions. Extract the trend turning point, fluctuation cycle, and change amplitude as key features from the stable change trend to construct the feature vector of the current trend; The cosine similarity algorithm is used to calculate the similarity between the feature vector and the historical working condition features in the historical working condition database, and high-matching reference samples with similarity exceeding a preset similarity threshold are selected. The high-matching reference samples and the feature vectors of the current trend are fused to form a classification dataset, which is then input into a support vector machine classification model trained based on historical anomaly samples to output the current anomaly type. Based on the current anomaly type, the risk level standard in the historical operating condition database is queried. The risk is then adjusted by combining the differences in environmental parameters and equipment runtime between the current and historical operating conditions to determine the risk level under the current operating condition.

6. The wind power equipment fault early warning method based on edge computing according to claim 1, characterized in that, The process involves acquiring the real-time resource status of edge nodes, matching and evaluating the risk level against the resource status, and if it is determined that the edge node resources are sufficient, initiating a deep analysis process to construct a refined anomaly prediction model, including: Obtain the real-time resource status of edge nodes; the resource status includes the current load level of computing units, the remaining capacity of storage space, the transmission capacity of network links, and the fluctuation of resource usage; The risk level is compared with the real-time resource status of the edge nodes in multiple dimensions to assess whether the edge nodes have the resource redundancy to carry out in-depth analysis tasks. If the assessment results indicate that the edge node resources are sufficient, that is, the computing load is in a low occupancy range, the remaining storage capacity meets the data processing needs, and the network transmission capability is stable and redundant, then the deep analysis process will be initiated. Extract short-term fluctuation characteristics, long-term trend characteristics, and parameter correlation characteristics of equipment operation, and integrate them to form a multi-dimensional analysis dataset; A refined anomaly prediction model is constructed based on the multi-dimensional analysis dataset; the refined anomaly prediction model includes parameter change thresholds, anomaly evolution paths, and correlation influence rules under different risk levels.

7. A wind power equipment fault early warning method based on edge computing according to claim 6, characterized in that, The construction of a refined anomaly prediction model based on the multi-dimensional analysis dataset includes: Extract short-term fluctuation features, long-term trend features, and parameter correlation features from the multi-dimensional analysis dataset to determine the key feature dimensions of the model input. Based on the physical characteristics of wind power equipment operation, a model is established to identify abnormal evolution paths; the identification rules include the correlation of parameter changes and the logic of abnormal propagation. Based on the key feature dimensions and identification rules, an anomaly prediction model framework that can output the probability of failure and the time of occurrence is constructed. The anomaly prediction model framework is trained and optimized using historical anomaly case data stored at edge nodes to form a refined anomaly prediction model that reflects the correspondence between parameter changes and fault occurrence.

8. A wind power equipment fault early warning method based on edge computing according to claim 7, characterized in that, The prediction of the evolution trend of potential abnormal fluctuations based on the refined anomaly prediction model, generating an early warning signal including the failure probability and occurrence time, includes: The short-term fluctuation characteristics, long-term trend characteristics, and parameter linkage characteristics of equipment operation are integrated into a multi-dimensional analysis dataset, which is then input into the constructed fine anomaly prediction model. The refined anomaly prediction model analyzes the correlation between features and the logic of parameter changes based on the identification rules of anomaly evolution paths, and extrapolates the development trend of potential abnormal fluctuations, outputting the extrapolation results. Based on the simulation results, assess the likelihood of the abnormal fluctuations causing equipment failure and determine the failure probability; By combining the current state and evolution pattern of abnormal fluctuations, analyze the time required for development to reach the fault state and determine the time of fault occurrence; The fault probability, fault occurrence time, and corresponding anomaly type information are integrated to form the early warning signal.

9. A method for early warning of wind power equipment faults based on edge computing according to claim 1, characterized in that, The process of assessing the risk level of the warning signal and activating a local response mechanism based on the risk level to output an immediate fault alarm includes: Based on the preset risk assessment criteria, the probability of failure and the expected occurrence time are extracted from the early warning signal, the current risk status is assessed, and the risk level is output. Based on the risk level, a corresponding local response mechanism is matched; the response mechanism includes the alarm notification scope, equipment emergency handling methods, and operation and maintenance cooperation strategies. According to the local response mechanism, the corresponding response process is initiated; the response process includes pushing alarms to relevant terminals, triggering device protection mode, and contacting maintenance personnel. The risk level, the probability of failure in the warning signal, the time of failure occurrence, and the response process are integrated into standard alarm content and output as the fault alarm.

10. A wind power equipment fault early warning system based on edge computing, characterized in that, include: Data acquisition module: Collects operational data of vibration, temperature, and current parameters through sensors configured on the wind power equipment, generates raw data sequences, and preprocesses the raw data sequences to obtain a preliminary feature set; Feature analysis module: Based on the preliminary feature set, time series analysis is used to calculate the time series gradient corresponding to each parameter, and a gradient descent optimization algorithm is introduced to iteratively optimize the time series gradient to determine the parameter change rate sequence; Smoothing module: When the parameter change rate sequence exceeds the preset change threshold, it is determined to be a potential abnormal fluctuation. The sliding window method is used to smooth the parameter change rate sequence to obtain a stable change trend. Risk analysis module: Calls a preset historical operating condition database, performs feature matching between the stable change trend and the historical operating condition data, and uses a support vector machine to classify the anomaly types of the matched features to obtain the risk level under the current operating condition; Model building module: Obtain the real-time resource status of edge nodes, match and evaluate the risk level with the resource status, and if it is determined that the edge node has sufficient resources, start the deep analysis process to build a refined anomaly prediction model. Early warning module: Based on the refined anomaly prediction model, it predicts the evolution trend of potential abnormal fluctuations and generates an early warning signal containing the failure probability and occurrence time. Fault assessment module: assesses the risk level of the warning signal and activates the local response mechanism based on the risk level to output an immediate fault alarm.

Images