An edge-computing-based wind turbine fault diagnosis method, system, device and medium

Abstract

This application belongs to the field of wind power fault diagnosis technology, and relates to a method, system, equipment, and medium for wind turbine fault diagnosis based on edge computing. The method includes: a perception layer collecting multi-source raw data through multi-parameter sensors deployed at various measuring points of the wind turbine; an edge layer preprocessing and extracting features from the data to obtain multi-dimensional feature parameters, which are then input into a lightweight fault diagnosis model. The model outputs the fault type, confidence level, and severity level through forward inference; the diagnostic results are graded according to a preset threshold strategy, with minor anomalies cached locally, and moderate or severe faults triggering local alarms and being synchronously uploaded to the cloud platform; the cloud platform incrementally trains and optimizes the model based on fault data and historical operation and maintenance samples, and the updated model is transmitted to the edge layer via encrypted transmission to complete model iteration. This application improves the real-time performance of fault diagnosis.

Description

Technical Field

[0001] This application belongs to the field of wind power fault diagnosis technology, and in particular relates to a method, system, equipment and medium for wind turbine fault diagnosis based on edge computing. Background Technology

[0002] Wind turbines are typically deployed in remote areas, operating in harsh environments and subjected to alternating loads, wind and sand erosion, lightning strikes, and other adverse factors over long periods. This leads to frequent failures of critical components such as blades, gearboxes, generators, and main shaft bearings. Statistics show that drivetrain failure is one of the main causes of unplanned shutdowns of wind turbines. Each major failure not only results in direct power loss but can also lead to high maintenance costs and safety hazards.

[0003] Currently, wind turbine fault diagnosis mainly relies on two types of systems: Supervisory Control and Data Acquisition (SCADA) systems and Condition Monitoring Systems (CMS) systems. SCADA systems collect parameters such as wind speed, power, temperature, and rotational speed through sensors deployed at key locations within the turbine, and trigger alarms for exceeding preset thresholds. CMS systems, on the other hand, collect high-frequency signals such as vibration and sound, which are then analyzed by technical personnel to determine the health status of rotating components such as gearboxes and bearings.

[0004] However, both types of systems have certain limitations in practical applications. SCADA systems, with their preset threshold alarm methods, lack the ability to accurately diagnose faults, only issuing alarms after parameters significantly exceed limits, making it difficult to detect potential faults in advance. While CMS systems can acquire high-frequency signals and perform spectrum analysis, this requires specialized technicians for interpretation, resulting in high labor costs, and the accuracy of the analysis results depends heavily on the experience and skill level of the technicians. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a wind turbine fault diagnosis method, system, equipment, and medium based on edge computing. By employing edge computing technology, some data processing and analysis tasks are moved from the cloud to the edge layer, which can effectively improve the real-time performance and accuracy of fault diagnosis.

[0006] Firstly, this application provides a wind turbine fault diagnosis method based on edge computing. The method includes, The perception layer collects multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine. The edge layer preprocesses the multi-source raw data to obtain multi-source processed data, and extracts features from the multi-source processed data to obtain multi-dimensional feature parameters. The multidimensional feature parameters are input into the lightweight fault diagnosis model, and the fault mode identification and fault severity assessment are completed through forward inference of the model, and the preliminary fault diagnosis results are output. The initial fault diagnosis results are classified and judged according to the fault level and preset threshold strategy. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform. Based on the fault data and historical operation and maintenance samples, the cloud platform performs incremental training and parameter optimization on the lightweight fault diagnosis model, generates an updated model weight file, and sends it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

[0007] Furthermore, The edge layer preprocesses the original multi-source data to obtain multi-source processed data, and extracts features from the multi-source processed data to obtain multi-dimensional feature parameters, specifically including: Receive the multi-source raw data, which includes at least raw vibration data, raw sound data, and raw temperature data; The multi-source raw data is subjected to at least data cleaning, filtering and noise reduction, and time alignment to obtain multi-source processed data. The data types corresponding to each processed data are matched with the preset data types in the preset feature extraction library. The feature extraction type is determined based on the matching results. Based on the feature extraction type, the processed data is used to extract features and obtain feature parameters.

[0008] Furthermore, Fault mode identification and fault severity assessment are performed through model forward inference, and preliminary fault diagnosis results are output, including: The multidimensional feature parameters are input into a lightweight fault diagnosis model, which includes a lightweight neural network or a gradient boosting tree. The lightweight fault diagnosis model performs forward calculation on the multidimensional feature parameters and outputs the fault type, fault confidence, and fault severity level. The severity level of the fault includes at least three levels: mild, moderate, and severe.

[0009] Furthermore, The preliminary fault diagnosis results are graded and judged according to the fault level and preset threshold strategy, specifically including: Obtain the fault confidence level and fault severity level from the preliminary fault diagnosis results; The fault confidence level was compared with the preset confidence threshold, and the fault severity level was compared with the preset severity level, respectively. When the fault confidence level is lower than the first confidence threshold and the fault severity level is low, it is judged as a minor anomaly; When the confidence level of the fault is higher than the second confidence threshold or the severity level of the fault is moderate, it is determined to be a moderate anomaly. When the fault confidence level is higher than the second confidence threshold and the fault severity level is severe, it is determined to be highly abnormal; The second confidence threshold is greater than the first confidence threshold.

[0010] Furthermore, Based on the judgment result, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform, specifically including: When a minor anomaly is identified, the preliminary fault diagnosis results and the corresponding fault data are stored in the local cache on the edge side and overwritten on a rolling basis according to a preset period. When a moderate or high anomaly is detected, a local alarm is immediately triggered on the edge side, and the preliminary fault diagnosis results and corresponding fault data are synchronously uploaded to the cloud management platform.

[0011] Furthermore, Based on the aforementioned fault data and historical operation and maintenance samples, the cloud platform performs incremental training and parameter optimization on the lightweight fault diagnosis model, specifically including: Receive several fault data and corresponding preliminary fault diagnosis results; The fault data is merged with the verified fault data in the historical operation and maintenance samples to construct an incremental training set; The lightweight fault diagnosis model is retrained using an incremental learning algorithm to update the model parameters and generate an updated model weight file. The historical operation and maintenance samples include verified fault type labeling data and corresponding feature data.

[0012] Furthermore, The lightweight fault diagnosis model is updated by transmitting data to the edge layer via encrypted transmission, specifically including: The updated model weight file is encrypted using an encryption algorithm to generate an encrypted model package; The encrypted model package is sent to the corresponding edge computing node in the edge layer. After receiving the encrypted model packet, the edge layer decrypts and verifies its integrity. If the verification is successful, it replaces the original local lightweight fault diagnosis model or performs incremental parameter updates on the local model to obtain the target fault diagnosis model.

[0013] Secondly, based on the same inventive concept, this application provides a wind turbine fault diagnosis system based on edge computing, the system comprising: The perception layer is used to collect multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine. The edge layer is used to preprocess the multi-source raw data to obtain multi-source processed data, and to extract features from the multi-source processed data to obtain multi-dimensional feature parameters. The multidimensional feature parameters are input into the lightweight fault diagnosis model, and the fault mode identification and fault severity assessment are completed through forward inference of the model, and the preliminary fault diagnosis results are output. The initial fault diagnosis results are classified and judged according to the fault level and preset threshold strategy. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform. The cloud platform is used to incrementally train and optimize the parameters of the lightweight fault diagnosis model based on the fault data and historical operation and maintenance samples, generate an updated model weight file, and send it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

[0014] Thirdly, this application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of any of the edge computing-based wind turbine fault diagnosis methods described above.

[0015] Fourthly, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the edge computing-based wind turbine fault diagnosis methods described above.

[0016] Compared with the prior art, this application has the following advantages: 1. This application deploys a lightweight fault diagnosis model at the edge layer, which completes data preprocessing, feature extraction and model inference in real time at the unit site. The diagnostic response time reaches the millisecond level, which overcomes the minute-level sampling lag of traditional SCADA systems and the transmission delay caused by centralized cloud processing. It can detect early weak faults such as bearing wear and gear damage in a timely manner, providing a sufficient time window for condition-based maintenance.

[0017] 2. This application collects multi-parameter data such as vibration, sound, and temperature through the sensing layer, achieves spatiotemporal alignment and feature-level fusion of multi-source data at the edge layer, and supports decision-level fusion strategies, comprehensively utilizing complementary information for fault identification. Compared with diagnostic schemes based on a single data source, it significantly improves the identification accuracy in multi-fault coupled scenarios and effectively solves the problem of data silos in existing monitoring systems.

[0018] 3. This application constructs a closed-loop mechanism of "edge diagnosis - cloud training - model distribution". The cloud platform uses real fault data uploaded from the edge layer to incrementally train the lightweight diagnostic model, and then encrypts and distributes the updated model to the edge nodes. This mechanism enables the diagnostic model to continuously absorb new fault samples, adapt to individual differences in units and changes in operating conditions, and overcomes the technical defects of traditional diagnostic models that are statically deployed and difficult to update.

[0019] 4. This application employs a tiered response mechanism. Minor anomalies are cached locally and periodically overwritten without being uploaded to the cloud; real-time uploads are triggered only for moderate and severe faults. Furthermore, edge nodes only upload diagnostic results and key feature data, rather than all high-frequency raw data. This mechanism significantly reduces network bandwidth consumption and cloud data storage, improving the system's economy and scalability.

[0020] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a wind turbine fault diagnosis method based on edge computing according to an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Figure 1 A flowchart illustrating a wind turbine fault diagnosis method based on edge computing according to an embodiment of this application is shown, as follows: Figure 1 As shown in the embodiment of this application, the wind turbine fault diagnosis method based on edge computing includes, S1, the perception layer collects multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine; In this embodiment, the multi-source raw data includes at least: raw vibration data collected by a vibration sensor, raw sound data collected by a sound sensor, and raw temperature data collected by a temperature sensor. In some embodiments, the multi-source raw data may also include one or more of preload data, oil monitoring data, rotational speed data, or power data, which can be determined according to the configuration of the wind turbine measuring points.

[0025] S2, the edge layer preprocesses the multi-source raw data to obtain multi-source processed data, and extracts features from the multi-source processed data to obtain multi-dimensional feature parameters; In this embodiment of the application, step S2 specifically includes: S21, Receive the multi-source raw data, wherein the multi-source raw data includes at least raw vibration data, raw sound data and raw temperature data; S22, Perform at least data cleaning, filtering and noise reduction and time alignment processing on the multi-source raw data to obtain multi-source processed data; S23, Match the data types corresponding to each processed data with the preset data types in the preset feature extraction library, determine the feature extraction type based on the matching results, extract features from the processed data based on the feature extraction type, and obtain feature parameters.

[0026] In this embodiment of the application, data cleaning is used to remove outliers caused by sensor anomalies and communication errors, and linear interpolation or nearest neighbor filling method is used to fill in the missing data; Filtering and denoising employ appropriate filters based on the signal type: for vibration signals, bandpass filters are used to extract characteristic frequency band components; for sound signals, spectral subtraction or Wiener filtering is used to suppress background noise. Time alignment is based on PTP (Precise Time Protocol) timestamps or GPS synchronized clocks to align data collected by different sensors to the same time reference, ensuring consistency in the time dimension for subsequent feature extraction and fusion analysis.

[0027] In some preferred embodiments, the preprocessing further includes a condition segmentation step: based on the wind speed signal collected by the wind speed encoder and / or the power signal output by the generator, the data is divided into different condition segments such as the start-up stage, rated operation stage, variable speed stage, and shutdown stage, so as to eliminate the interference of condition changes on fault feature extraction.

[0028] In this embodiment, the edge computing node calls a locally stored preset feature extraction library, which pre-configures mapping relationships between different data types and feature extraction types. The edge computing node matches the data type corresponding to each processed data with the preset data types in the preset feature extraction library, determines the feature extraction type corresponding to each data based on the matching result, and then performs feature extraction on each processed data according to the determined feature extraction type to obtain multi-dimensional feature parameters.

[0029] For example: When the data type is vibration data, the matching feature extraction types include: time domain feature extraction (peak value, effective value, kurtosis, waveform factor, etc.), frequency domain feature extraction (amplitude of each natural frequency, sideband energy, etc.), and time-frequency feature extraction (wavelet packet energy entropy). When the data type is sound data, the matching feature extraction types include: MFCC coefficients, formants, spectral centroids, and sound pressure level; When the data type is temperature data, the matching feature extraction types include: mean, rate of change, and difference from reference temperature.

[0030] S3, input the multidimensional feature parameters into the lightweight fault diagnosis model, complete the fault mode recognition and fault degree assessment through model forward inference, and output the preliminary fault diagnosis results; In this embodiment of the application, step S3 specifically includes: S31, the multidimensional feature parameters are input into the lightweight fault diagnosis model, which includes a lightweight neural network or a gradient boosting tree; S32, the lightweight fault diagnosis model performs forward calculation on the multidimensional feature parameters and outputs the fault type, fault confidence and fault severity level; The severity level of the fault includes at least three levels: mild, moderate, and severe.

[0031] In this embodiment, the lightweight fault diagnosis model is a lightweight machine learning model that is pre-built and deployed on edge computing nodes through offline training. Its model structure undergoes lightweight processing such as pruning, quantization, or knowledge distillation to adapt to the limited computing resources and storage space of edge devices.

[0032] In this embodiment, the lightweight fault diagnosis model employs a lightweight neural network, such as MobileNetV2, SqueezeNet, or a custom convolutional neural network with optimized structure. In another implementation, the lightweight fault diagnosis model uses a gradient boosting tree model, such as XGBoost or LightGBM. These models are characterized by high computational efficiency and strong interpretability, making them suitable for resource-constrained edge deployments.

[0033] In this embodiment of the application, the lightweight fault diagnosis model performs forward inference calculations on the input multidimensional feature parameters, specifically: For lightweight neural network models, feature parameters are propagated forward sequentially through the input layer, multiple lightweight convolutional layers, pooling layers, and fully connected layers, and finally the output layer outputs the fault category probability distribution through the Softmax function or the Sigmoid function. For the gradient boosting tree model, the feature parameters are propagated forward through multiple decision trees in sequence. Each tree outputs a predicted value based on the feature splitting condition. The prediction results of all trees are accumulated and then logically transformed to obtain the probability of the fault category.

[0034] In this embodiment of the application, the preliminary fault diagnosis results include at least: Fault type: The specific fault category identified by the model, such as "bearing wear", "gear damage", "loose connecting bolts", etc. Fault confidence level: Expressed as a percentage or probability value, this indicates how reliable the model's prediction of the fault type is. A higher confidence level means a more reliable judgment from the model. The value typically ranges from 0 to 1 or from 0% to 100%. Fault Severity Level: The level classification output by the model after assessing the severity of the identified faults, including at least three levels: mild, moderate and severe, and can be further subdivided into more levels according to actual needs.

[0035] In this embodiment, the lightweight fault diagnosis model supports the simultaneous output of identification results for multiple fault types in a single inference through a multi-label classification architecture, adapting to the actual situation of multiple coupled faults. When multiple fault types exist, each fault type corresponds to its own confidence level and severity level.

[0036] In one optional embodiment, the lightweight fault diagnosis model supports the identification of at least ten of the following typical abnormal operating conditions: imbalance, misalignment, bearing wear, bearing damage, poor bearing lubrication, gear wear, gear damage, loose connecting bolts, blade damage, overall 3P vibration, elastic support failure, component overheating, and gearbox oil leakage. This identification capability covers common fault modes of key transmission chains and structural components of wind turbines, meeting the diagnostic needs of engineering applications.

[0037] S4. Based on the fault level and preset threshold strategy, the preliminary fault diagnosis results are classified and judged. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is uploaded to the cloud management platform simultaneously. In this embodiment of the application, step S4 specifically includes: S41, Obtain the fault confidence level and fault severity level from the preliminary fault diagnosis results; S42, compare the fault confidence with the preset confidence threshold and the fault severity level with the preset severity, respectively; S43, when the fault confidence level is lower than the first confidence threshold and the fault severity level is low, it is determined to be a minor abnormality; S44, when the fault confidence level is higher than the second confidence threshold or the fault severity level is moderate, it is determined to be a moderate anomaly; S45, when the fault confidence level is higher than the second confidence threshold and the fault severity level is severe, it is determined to be highly abnormal; The second confidence threshold is greater than the first confidence threshold.

[0038] In the embodiments of this application, the fault confidence level represents the degree of credibility of the model's diagnostic conclusion, and is usually expressed as a probability value between 0 and 1 or a percentage between 0% and 100%. The higher the value, the more reliable the model's judgment on the fault type. The fault severity level represents the model's quantitative assessment of the severity of the identified faults. In some embodiments, the fault severity level is calculated either through independent nodes in the model output layer or through a mapping relationship between fault feature parameters and a preset severity level.

[0039] In this application embodiment, for example, the first confidence threshold can be set to 0.5 (50%), and the second confidence threshold can be set to 0.8 (80%).

[0040] In this embodiment of the application, step S4 further includes: S41, when a minor abnormality is determined, the preliminary fault diagnosis result and the corresponding fault data are stored in the local cache on the edge side and are rolled over according to a preset period. S42, when a moderate or high anomaly is determined, a local alarm is immediately triggered on the edge side, and the preliminary fault diagnosis results and corresponding fault data are synchronously uploaded to the cloud management platform.

[0041] In this embodiment of the application, when the network is interrupted or communication is restricted, the edge layer temporarily stores the preliminary fault diagnosis results and the corresponding fault data in a local cache, and automatically retransmits them after communication is restored.

[0042] In this embodiment of the application, the fault data includes at least: fault type, fault confidence level, fault severity level, diagnosis timestamp, and corresponding multidimensional feature parameters and raw data fragments.

[0043] In this embodiment, the local cache is managed using a circular queue or FIFO (First-In, First-Out) structure, and rolling over is performed according to a preset period or preset capacity. For example, the cache capacity can be set to the most recent 7 days of slightly abnormal data; when new data is written, the oldest historical data is automatically overwritten. Through the rolling over mechanism, recent abnormal records are retained for traceability analysis while effectively controlling the storage resource consumption of edge nodes.

[0044] In this embodiment of the application, the local alarm includes one or more of the following forms: Audible and visual alarms: Audible and visual signals are emitted through the indicator lights, buzzers, or external alarm devices of the edge node devices; Interface push: Alarm information pops up on the field operation and maintenance terminal (such as handheld PDA, field industrial control computer), showing the fault type, location, severity and suggested handling measures; Main control linkage: In case of abnormal conditions, edge nodes can send linkage signals to the wind turbine main control system, suggesting the execution of protective actions such as power reduction operation, operation restriction or emergency shutdown.

[0045] In some embodiments, edge nodes set different retransmission priorities based on the importance and timeliness of the data: highly abnormal data is retransmitted first, moderately abnormal data is retransmitted second, and slightly abnormal data is retransmitted with a summary when resources permit.

[0046] S5, the cloud platform performs incremental training and parameter optimization on the lightweight fault diagnosis model based on the fault data and historical operation and maintenance samples, generates an updated model weight file, and sends it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

[0047] In this embodiment of the application, step S5 specifically includes: S51 receives several fault data and corresponding preliminary fault diagnosis results; S52, merge the fault data with the verified fault data in the historical operation and maintenance samples to construct an incremental training set; S53 uses an incremental learning algorithm to retrain the lightweight fault diagnosis model, update the model parameters, and generate an updated model weight file. The historical operation and maintenance samples include verified fault type labeling data and corresponding feature data.

[0048] In this embodiment, after receiving the data, the cloud platform first performs data verification and deduplication. Specifically, the cloud platform generates a unique data identifier based on the timestamp, unit identifier, and measurement point location information in the data packet, compares it with the stored data records, and eliminates duplicate data packets. Simultaneously, it checks data integrity; for data packets missing key fields, it records an anomaly log and sends a retransmission request to the corresponding edge node.

[0049] In this embodiment of the application, the cloud platform employs the following strategy to construct the incremental training set: From the newly received fault data, data with high confidence (e.g., greater than 0.8) and classified as moderate or high anomalies are selected as candidate samples. The candidate samples are merged with samples of the same type and operating conditions in the historical operation and maintenance sample library and organized in chronological order. For fault types with a small number of samples, data augmentation (such as adding noise or time shifting) or oversampling techniques can be used to expand the sample size and avoid class imbalance during model training. Perform the same standardization process on the feature parameters as during initial training (such as Z-score normalization) to ensure consistent feature distribution.

[0050] In one implementation, the incremental learning algorithm employs one or a combination of the following strategies: Fine-tuning strategy: Based on the pre-trained model, the model is fine-tuned several times using an incremental training set. To preserve existing knowledge, a small learning rate is adopted (such as 1 / 10 of the initial learning rate), and constraints are imposed on the parameter updates of the model's underlying feature extraction layer.

[0051] Knowledge distillation strategy: The current model is used as the teacher model. The incremental training set is input into the teacher model to generate soft labels, which are then combined with the real hard labels of the new samples to train the student model (with the same or simplified structure as the original model). By distilling the loss function, the student model learns the features of the new samples while maintaining consistency with the output of the teacher model.

[0052] Experience replay strategy: During incremental training, in addition to the new samples in the current batch, some representative samples (such as a small number of typical samples for each fault type) are randomly selected from the historical operation and maintenance sample library and mixed into the training batch to participate in parameter updates, so as to prevent the model from forgetting historical knowledge.

[0053] Elastic weight solidification strategy: During training, the importance of different parameters in the model is estimated. For parameters that are important in historical tasks, a large regularization constraint is applied to limit their update range, while parameters with lower importance are allowed to be adjusted more significantly to learn new knowledge.

[0054] In this embodiment of the application, step S5 further includes: S54, The updated model weight file is encrypted using an encryption algorithm to generate an encrypted model package; S55, the encrypted model package is sent to the corresponding edge computing node of the edge layer; S56, after receiving the encrypted model packet, the edge layer decrypts and verifies its integrity. If the verification is successful, it replaces the original local lightweight fault diagnosis model or performs incremental parameter updates on the local model to obtain the target fault diagnosis model.

[0055] In this embodiment, the edge layer sends update confirmation information to the cloud platform after the model update is completed.

[0056] In one implementation, this embodiment uses a symmetric encryption algorithm (such as AES-256) to encrypt the model weight file. Symmetric encryption has the advantages of high computational efficiency and fast encryption / decryption speed, and is suitable for scenarios where the model file is large and needs to be frequently distributed. The cloud platform and edge nodes pre-negotiate and securely store a shared key, which is then used by the cloud to encrypt the model file.

[0057] In another implementation, this application employs a hybrid encryption scheme combining asymmetric and symmetric encryption. The cloud platform first generates a temporary symmetric key, uses this key to encrypt the model weight file, and then uses the public key of the edge nodes to encrypt the temporary symmetric key. The encrypted model file and the encrypted temporary symmetric key are then packaged together into an encrypted model package. This scheme combines the efficiency of symmetric encryption with the key distribution security of asymmetric encryption.

[0058] In this embodiment, after encryption, the cloud platform further adds a digital signature or message authentication code to the encrypted model packet, which is used by edge nodes to verify the authenticity of the model packet's origin and the integrity of the data before decryption. For example, the HMAC-SHA256 algorithm can be used to generate the signature, which is then appended to the end of the model packet.

[0059] In this embodiment, the cloud platform distributes the encrypted model packet to the corresponding edge computing nodes at the edge layer via a secure transmission protocol. For protocol selection, TLS / DTLS encrypted transmission channels can be used to carry protocols such as HTTPS and MQTT over TLS, providing additional security at the transport layer. The cloud platform selects a suitable distribution method based on the network status of the edge nodes and the model update strategy. For edge nodes that are in good online condition, real-time push notifications are used. For nodes with poor network quality, a breakpoint resume mechanism is adopted, supporting chunked transmission and retry on failure.

[0060] In this embodiment, the edge node performs integrity verification on the received data. If the model package includes a digital signature or message authentication code, the edge node uses a pre-set shared key or cloud public key to verify the signature, confirming that the model package has not been tampered with during transmission and that it is authorized for distribution from the cloud. If the verification fails, the edge node discards the model package and sends a verification failure message to the cloud platform, requesting a retransmission.

[0061] In this embodiment, the full replacement update includes the edge node replacing the locally stored original lightweight fault diagnosis model with the decrypted updated model weight file. To ensure the atomicity of the update process and system availability, the edge node adopts a dual-copy mechanism: first, the new model is written to a temporary directory, and after integrity verification and loading tests are completed, the model file path is atomically switched, and then the old model file is deleted. If the new model fails to load, the system automatically rolls back to the old model and reports the update failure information to the cloud platform.

[0062] Incremental parameter updates involve edge nodes updating only the changed parameters for model structures that support incremental updates (such as gradient boosting trees or neural networks with modular structures), rather than replacing the entire model file. Incremental updates significantly reduce the amount of data downloaded, making them suitable for scenarios with limited network bandwidth. When generating an updated model, the cloud platform calculates the difference (diff) between the parameters of the old and new models, encrypts the difference file, and sends it out. Edge nodes receive this file and merge the difference into their local models to obtain the updated model.

[0063] Based on the above method, this application embodiment also provides a wind turbine fault diagnosis system based on edge computing corresponding to the above method, the system comprising: The perception layer is used to collect multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine. The edge layer is used to preprocess the multi-source raw data to obtain multi-source processed data, and to extract features from the multi-source processed data to obtain multi-dimensional feature parameters. The multidimensional feature parameters are input into the lightweight fault diagnosis model, and the fault mode identification and fault severity assessment are completed through forward inference of the model, and the preliminary fault diagnosis results are output. The initial fault diagnosis results are classified and judged according to the fault level and preset threshold strategy. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform. The cloud platform is used to incrementally train and optimize the parameters of the lightweight fault diagnosis model based on the fault data and historical operation and maintenance samples, generate an updated model weight file, and send it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

[0064] Based on the same inventive concept disclosed above, this application also provides an electronic device. The electronic device of this application includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0065] It should be noted that the electrical connections between the above-mentioned units do not necessarily represent the connections between lines. Indirect connections are applicable to the embodiments of this application as long as they achieve the purpose of this application.

[0066] Based on the same inventive concept, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above method.

[0067] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A wind turbine fault diagnosis method based on edge computing, characterized in that, The method includes, The perception layer collects multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine. The edge layer preprocesses the multi-source raw data to obtain multi-source processed data, and extracts features from the multi-source processed data to obtain multi-dimensional feature parameters. The multidimensional feature parameters are input into the lightweight fault diagnosis model, and the fault mode identification and fault severity assessment are completed through forward inference of the model, and the preliminary fault diagnosis results are output. The initial fault diagnosis results are classified and judged according to the fault level and preset threshold strategy. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform. Based on the fault data and historical operation and maintenance samples, the cloud platform performs incremental training and parameter optimization on the lightweight fault diagnosis model, generates an updated model weight file, and sends it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

2. The method according to claim 1, characterized in that, The edge layer preprocesses the original multi-source data to obtain multi-source processed data, and extracts features from the multi-source processed data to obtain multi-dimensional feature parameters, specifically including: Receive the multi-source raw data, which includes at least raw vibration data, raw sound data, and raw temperature data; The multi-source raw data is subjected to at least data cleaning, filtering and noise reduction, and time alignment to obtain multi-source processed data. The data types corresponding to each processed data are matched with the preset data types in the preset feature extraction library. The feature extraction type is determined based on the matching results. Based on the feature extraction type, the processed data is used to extract features and obtain feature parameters.

3. The method according to claim 1, characterized in that, Fault mode identification and fault severity assessment are performed through forward inference of the model, and preliminary fault diagnosis results are output, including: The multidimensional feature parameters are input into a lightweight fault diagnosis model, which includes a lightweight neural network or a gradient boosting tree. The lightweight fault diagnosis model performs forward calculation on the multidimensional feature parameters and outputs the fault type, fault confidence, and fault severity level. The severity level of the fault includes at least three levels: mild, moderate, and severe.

4. The method according to claim 3, characterized in that, The preliminary fault diagnosis results are graded based on the fault level and a preset threshold strategy, specifically including: Obtain the fault confidence level and fault severity level from the preliminary fault diagnosis results; The fault confidence level was compared with the preset confidence threshold, and the fault severity level was compared with the preset severity level, respectively. When the fault confidence level is lower than the first confidence threshold and the fault severity level is low, it is judged as a minor anomaly; When the confidence level of the fault is higher than the second confidence threshold or the severity level of the fault is moderate, it is determined to be a moderate anomaly. When the fault confidence level is higher than the second confidence threshold and the fault severity level is severe, it is determined to be highly abnormal; The second confidence threshold is greater than the first confidence threshold.

5. The method according to claim 4, characterized in that, Based on the judgment result, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform, specifically including: When a minor anomaly is identified, the preliminary fault diagnosis results and the corresponding fault data are stored in the local cache on the edge side and overwritten on a rolling basis according to a preset period. When a moderate or high anomaly is detected, a local alarm is immediately triggered on the edge side, and the preliminary fault diagnosis results and corresponding fault data are synchronously uploaded to the cloud management platform.

6. The method according to claim 1, characterized in that, Based on the aforementioned fault data and historical operation and maintenance samples, the cloud platform performs incremental training and parameter optimization on the lightweight fault diagnosis model, specifically including: Receive several fault data and corresponding preliminary fault diagnosis results; The fault data is merged with the verified fault data in the historical operation and maintenance samples to construct an incremental training set; The lightweight fault diagnosis model is retrained using an incremental learning algorithm to update the model parameters and generate an updated model weight file. The historical operation and maintenance samples include verified fault type labeling data and corresponding feature data.

7. The method according to claim 1, characterized in that, The lightweight fault diagnosis model is updated by transmitting data to the edge layer via encrypted transmission, specifically including: The updated model weight file is encrypted using an encryption algorithm to generate an encrypted model package; The encrypted model package is sent to the corresponding edge computing node in the edge layer. After receiving the encrypted model packet, the edge layer decrypts and verifies its integrity. If the verification is successful, it replaces the original local lightweight fault diagnosis model or performs incremental parameter updates on the local model to obtain the target fault diagnosis model.

8. A wind turbine fault diagnosis system based on edge computing, characterized in that, The system includes: The perception layer is used to collect multi-source raw data based on multi-parameter sensors deployed at various measuring points of the wind turbine. The edge layer is used to preprocess the multi-source raw data to obtain multi-source processed data, and to extract features from the multi-source processed data to obtain multi-dimensional feature parameters. The multidimensional feature parameters are input into the lightweight fault diagnosis model, and the fault mode identification and fault severity assessment are completed through forward inference of the model, and the preliminary fault diagnosis results are output. The initial fault diagnosis results are classified and judged according to the fault level and preset threshold strategy. Based on the judgment results, the fault data is stored in the local cache on the edge side, or a local alarm is triggered, and the fault data is synchronously uploaded to the cloud management platform. The cloud platform is used to incrementally train and optimize the parameters of the lightweight fault diagnosis model based on the fault data and historical operation and maintenance samples, generate an updated model weight file, and send it to the edge layer through encrypted transmission to complete the update of the lightweight fault diagnosis model and obtain the target fault diagnosis model.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the edge computing-based wind turbine fault diagnosis method according to any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the edge computing-based wind turbine fault diagnosis method according to any one of claims 1-7.

Images