Centralized controller gateway and remote operation and maintenance system integrating AI fault diagnosis

Abstract

This invention discloses a centralized controller gateway and remote operation and maintenance system integrating AI fault diagnosis, belonging to the field of automation control and system management technology. By deploying a centralized controller gateway with edge computing and intelligent decision-making capabilities at the equipment site and deeply collaborating with a cloud-based remote operation and maintenance system, this invention achieves early and accurate prediction and rapid closed-loop handling of equipment faults. This invention establishes a two-way model optimization mechanism of edge-side personalized adaptation and cloud-based group experience evolution. This cloud-edge collaborative continuous learning mode ensures the long-term effectiveness and high accuracy of the diagnostic system throughout the entire equipment lifecycle. This invention integrates fault prediction, maintenance work order generation, and global resource scheduling into a fully automated and intelligent process, realizing the transformation of operation and maintenance management from passive response to proactive value creation, and improving the operational economy of the entire system.

Description

Technical Field

[0001] This invention relates to the field of automation control and system management technology, specifically to a centralized controller gateway and remote operation and maintenance system that integrates AI fault diagnosis. Background Technology

[0002] In the field of large-scale industrial asset management, such as wind power generation, centralized controller gateways and remote operation and maintenance systems are the core infrastructure for realizing automated monitoring and control of equipment clusters. Centralized controller gateways are typically deployed at the equipment site, responsible for collecting various sensor data and communicating with a remote center, while remote operation and maintenance systems analyze massive amounts of data in the cloud, display equipment status, and assist operation and maintenance personnel in decision-making and management to ensure the safe and stable operation of the equipment.

[0003] In existing technologies, a common approach is to use a data pass-through gateway. Its main function is to transmit raw operational data or simply aggregated data collected by field sensors to a remote cloud service platform in real time via a communication network. All complex calculations and analyses, including fault diagnosis and prediction, are performed by the cloud server. Based on the received data, the cloud system uses a fixed algorithm model or rule base to make judgments. When an anomaly is detected, it sends alarm information to maintenance personnel, who then intervene to create work orders and schedule maintenance. Existing technologies, such as the invention application patent with announcement number CN113988189B, focus on the general optimization of the cloud model, lacking personalized adaptation based on the general model (i.e., the backbone network). They cannot perform real-time correction based on the historical data of individual devices to achieve accurate diagnosis that combines general and personalized approaches. Existing technologies, such as the invention application patent with announcement number CN114004091B, disclose diagnostic approaches based on single-point, specific algorithm optimization, but cannot adaptively select the optimal feature extraction rules based on different operating conditions, such as startup and rated operation, making the diagnostic strategy more flexible and better suited to the real-time status of the equipment.

[0004] The aforementioned existing technical solutions have several drawbacks in practical applications. First, uploading massive amounts of raw data to the cloud places extremely high demands on the bandwidth and stability of the communication network, easily leading to data congestion and transmission delays. Second, the reliance on centralized cloud processing introduces an inherent time lag between data collection and the acquisition of analysis results, making it difficult to meet the control requirements for rapid response to sudden failures. Furthermore, general diagnostic models deployed in the cloud often fail to adequately adapt to the individual differences and performance degradation characteristics of each device, resulting in limited diagnostic accuracy, excessive manual steps in the maintenance process, long response chains, and low efficiency. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to provide a centralized controller gateway and remote operation and maintenance system that integrates AI fault diagnosis.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The first aspect of the present invention provides a centralized controller gateway for integrating AI fault diagnosis, including: a data acquisition module, used to acquire real-time operating status data from multiple sensors of at least one connected wind turbine, wherein the sensors include at least a vibration sensor, a speed sensor and a power sensor.

[0007] The edge computing processing module is used to preprocess and extract features from the running status data to generate a time-series feature dataset.

[0008] The AI ​​fault diagnosis module is integrated into the edge computing processing module and includes a pre-set fault prediction model. The AI ​​fault diagnosis module is used to input the time-series feature dataset into the fault prediction model and output the fault prediction result of the wind turbine. The fault prediction result includes at least the probability of fault occurrence and the fault type.

[0009] The local decision-making and control module is connected to the AI ​​fault diagnosis module. It is used to determine the fault prediction results based on preset judgment rules, and generate and issue corresponding local control commands or warning signals when the judgment meets the warning or control conditions.

[0010] The communication module, connected to the edge computing processing module and the local decision and control module, is used to transmit the fault prediction results, the early warning signals and the processed operating status data to the remote operation and maintenance system, and to receive remote instructions from the remote operation and maintenance system.

[0011] A second aspect of the present invention provides a remote operation and maintenance system configured to utilize the centralized controller gateway for integrated AI fault diagnosis, comprising: the centralized controller gateway, which is deployed on the wind farm side to perform real-time data acquisition, edge-side fault prediction and local decision-making for at least one wind turbine.

[0012] A cloud service platform, communicatively connected to at least one of the central controller gateways, the cloud service platform comprising: The model management and training unit is used to issue or update the fault prediction model to the centralized controller gateway, and to retrain the model based on historical operating data and fault data collected from multiple centralized controller gateways to generate the corresponding model update package.

[0013] The remote operation and maintenance management unit is used to receive and centrally display the fault prediction results, early warning signals and operating status data from the centralized controller gateway, and generate preventive maintenance work orders based on the fault prediction results.

[0014] The operation and maintenance coordination scheduling unit is used to analyze and schedule maintenance resources based on the content of the preventive maintenance work order, generate an optimized maintenance plan, and send it to the target user terminal.

[0015] The distinguishing features of this invention are: (1) an AI fault diagnosis module, including a dynamic feature library, backbone network, and adaptation module, is integrated on the gateway side to achieve real-time and personalized preliminary diagnosis at the edge; (2) a local decision-making and control module is set up, which can perform simulation and deduction based on the diagnosis results and execute optimal temporary control; (3) the cloud system has the function of aggregating personalized edge experience to optimize the backbone network. These features together practically solve the technical problem of how to achieve low-latency, high-precision, and continuously evolving device fault prediction and autonomous response at the resource-constrained edge.

[0016] The beneficial effects of this invention are as follows: (1) By deploying a centralized controller gateway with edge computing and intelligent decision-making capabilities at the equipment site and deeply collaborating with the cloud-based remote operation and maintenance system, this invention achieves early and accurate prediction and rapid closed-loop handling of equipment faults. This method brings data processing and preliminary diagnostic capabilities down to the equipment side, enabling immediate response to fault symptoms and execution of localized optimal temporary control strategies, shortening the response time from risk identification to preliminary intervention, effectively suppressing the further development of faults, and ensuring the safety of equipment operation.

[0017] (2) This invention establishes a two-way model optimization mechanism that combines edge-side personalized adaptation with cloud-based group experience evolution. The field gateway can fine-tune the diagnostic model based on the operating characteristics of a single device, while the cloud aggregates massive amounts of device operating data and local adaptation experience to globally optimize and iterate the basic model. This cloud-edge collaborative continuous learning mode enables the diagnostic accuracy of the entire system to spontaneously improve with the accumulation of time and data, ensuring the long-term effectiveness and high accuracy of the diagnostic system throughout the entire device lifecycle.

[0018] (3) This invention integrates fault prediction, maintenance work order generation, and global resource scheduling into a fully automated and intelligent process. The system can automatically generate preventative maintenance work orders with clear execution windows based on the prediction results, and formulate optimized maintenance plans that take into account equipment risk, resource availability, and overall production efficiency through multi-objective optimization algorithms. This not only significantly improves the utilization efficiency of maintenance resources, but also reduces the impact of maintenance activities on production through forward-looking planning, realizing the transformation of operation and maintenance management from passive response to proactive value creation, and improving the operational economy of the entire system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the centralized controller gateway of the present invention.

[0021] Figure 2 This is a schematic diagram of the remote operation and maintenance system of the present invention. Detailed Implementation

[0022] 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.

[0023] Reference Figure 1 As shown, the first aspect of the present invention provides a centralized controller gateway for integrating AI fault diagnosis, including: a data acquisition module for acquiring real-time operating status data from multiple sensors of at least one connected wind turbine, wherein the sensors include at least a vibration sensor, a speed sensor and a power sensor.

[0024] The edge computing processing module is used to preprocess and extract features from the running status data to generate a time-series feature dataset.

[0025] In a specific embodiment of the present invention, the preprocessing and feature extraction of the operating status data to generate a time-series feature dataset specifically involves: acquiring raw data streams from vibration sensors, speed sensors, and power sensors.

[0026] The original data stream is cleaned and denoised to generate regular time-series data.

[0027] It should be noted that the raw data streams of the wind turbine collected by the vibration sensor, speed sensor, and power sensor are cleaned and denoised using the following steps to generate regularized time-series data: A1. Standardization of key operating parameters: For vibration signals in the raw data stream, unit: m / s 2To address the differences in parameter formats between speed signals (r / min) and power signals (kW), data accuracy was standardized. Data was arranged in the order of acquisition timestamps to construct a three-dimensional time-series data sequence of timestamp-vibration value-speed value-power value, ensuring the temporal correlation and format uniformity of data from different sensors.

[0028] A2. Effective signal denoising and purification: Wavelet threshold denoising algorithm is used to remove environmental interference noise from the original data, such as false vibration pulses caused by sudden wind speed changes, external electromagnetic interference signals and sensor hardware noise such as zero drift signals and circuit noise. Effective signal components directly related to the mechanical operation of wind turbines, such as bearing rotation, blade rotation and power output, are retained, such as the filtered bearing horizontal / vertical vibration signal and the continuous power output signal under stable operation.

[0029] A3. Data integrity completion and correction: For data loss caused by temporary sensor disconnection, a linear interpolation algorithm is used to complete the missing values ​​in the time series; for abnormal values ​​that exceed the normal operating range of the equipment, such as invalid data with speed > 120% of rated speed or power < 0kW, reasonable estimation and correction are performed based on the historical data trend of a set number of consecutive timestamps before and after, such as 5, to ensure the continuity and validity of the time series data and avoid interference from abnormal values ​​to subsequent feature extraction.

[0030] A4. Structured Labels and Dimensional Division: Add classification labels to the regularized time series data, including sensor type labels such as vibration-horizontal, vibration-vertical, speed, power, and equipment operation stage labels such as startup stage, rated operation stage, and shutdown stage. The data is structured according to monitoring dimensions and operation scenarios to form a structured time series dataset of timestamp-label-standardized values, which facilitates the edge computing processing module to quickly locate target data and extract fault correlation features.

[0031] Identify the current operating mode of the device reflected by the regularized timing data.

[0032] It should be noted that the operating modes include, but are not limited to, the startup phase, the rated operation phase, and the shutdown phase.

[0033] Select the corresponding feature calculation rule from the feature calculation rule set stored in the dynamic feature library, based on the current operating mode of the device.

[0034] It should be noted that the dynamic feature library predefines the most fault-sensitive feature indicators and their calculation methods under different operating modes. Specifically, it is a dynamically adaptable feature calculation rule set database built based on the differences in fault-sensitive features under different operating modes of wind turbines. Its purpose is to provide targeted and high-precision feature calculation basis for AI fault diagnosis under different operating modes. The specific construction method is as follows: B1. Classification of Equipment Operation Modes: Based on the operating mechanism of wind turbines and actual operation and maintenance data, combined with the operation stage labels in the regularized time series data, the equipment operation modes are divided into preset categories such as startup stage and rated operation stage. Each operation mode is clearly defined by the threshold of key operating parameters, such as startup stage: speed 0-800r / min, power 0-200kW; rated operation stage: speed 1000-1200r / min, power 800-1000kW.

[0035] B2. Analyze the fault sensitivity characteristics under each operating mode. For each type of operating mode, through historical fault data statistics, simulation experiments and mechanism analysis, determine the types of faults that the equipment is prone to occur under this operating mode. For example, during the start-up phase, the starting motor is prone to jamming and abnormal speed increase; during the rated operation phase, bearing wear and blade imbalance are prone to occur. Extract the fault sensitivity characteristics that are strongly correlated with this type of fault, such as the time domain characteristics and frequency domain characteristics of vibration signals, and the rate of change characteristics of speed signals.

[0036] B3. Define the feature calculation rules corresponding to each operating mode. For the fault-sensitive features in each type of operating mode, formulate corresponding feature calculation rules. Each rule should clearly include: the calculation object, such as vibration signal and speed signal in regular time series data; the calculation method, such as time domain statistics, frequency domain analysis, and trend fitting; the calculation parameters, such as time window length, frequency range, and fitting order; and the output result format, such as feature value and feature vector.

[0037] B4. The storage structure of the dynamic feature library adopts a structured storage method, which associates and stores the running mode identifier, the corresponding feature calculation rule set including the calculation object, method, parameters, output format, rule validity identifier, update timestamp, and other information to form a mapping relationship table between running mode and feature calculation rule, which supports quick indexing of the corresponding rule set according to the running mode.

[0038] B5. Regularly collect equipment operation data and fault diagnosis results. If it is found that the existing feature calculation rules have a fault identification accuracy rate of less than a preset threshold such as 90% for a certain type of operation mode, or if new equipment operation modes or fault types are added, the dynamic feature library will be iteratively updated by re-analyzing fault-sensitive features and optimizing calculation rules to ensure the adaptability and effectiveness of feature calculation rules.

[0039] Examples of dynamic feature libraries are shown in Table 1. Table 1 - Dynamic Feature Library

[0040] The regularized time series data are calculated using the aforementioned feature calculation rules to generate the time series feature dataset.

[0041] It should be noted that by using this set of selected feature calculation rules, targeted mathematical operations are performed on the regularized time series data, ultimately generating a highly condensed time series feature dataset that is strongly correlated with the current operating state, which serves as the input for the subsequent fault prediction model.

[0042] The principle behind the above lies in implementing adaptive feature engineering based on the real-time operating conditions of equipment. This replaces static analysis methods by dynamically selecting the optimal feature extraction strategy according to the current operating mode of the equipment. The beneficial effects are that it can accurately extract the most indicative features of early faults from massive amounts of raw data, improving the quality of data input to the artificial intelligence model, thereby enhancing the sensitivity and accuracy of fault prediction. Simultaneously, because the computational objective is clearly defined, it also reduces the computational load on edge devices, maximizing the analysis benefits with limited resources.

[0043] The AI ​​fault diagnosis module is integrated into the edge computing processing module and includes a pre-set fault prediction model. The AI ​​fault diagnosis module is used to input the time-series feature dataset into the fault prediction model and output the fault prediction result of the wind turbine. The fault prediction result includes at least the probability of fault occurrence and the fault type.

[0044] It should be noted that when the system is first deployed, a pre-set general feature library and model are used, and personalized adaptations are gradually made as running data accumulates.

[0045] In a specific embodiment of the present invention, the step of inputting the time-series feature dataset into the fault prediction model and outputting the fault prediction result of the wind turbine is as follows: the time-series feature dataset is input into the fault prediction model, and the fault prediction model includes a backbone network and an adaptation module.

[0046] The backbone network processes the time-series feature dataset to generate an initial prediction vector.

[0047] It should be noted that the backbone network is a pre-trained general-purpose deep learning network trained on a large amount of historical fault data. It excels at mining universal and deep-seated fault symptoms from input data. When the time-series feature dataset flows through the backbone network, an initial prediction vector is generated. This vector is a preliminary and generalized judgment of the current equipment state by the model, including the failure probability of key components such as mechanical / electrical systems, the correlation between typical fault types such as bearing wear and power anomalies, the confidence level of normal equipment operation, the fault development stage coefficient, the abnormal contribution value of each feature such as vibration / speed / power, and the reliability coefficient of the prediction result. The whole is a structured numerical vector that reflects universal fault symptoms. The fault development stage coefficient refers to a value between 0 and 1 calculated based on the matching degree between the historical similar fault development curves and the current feature trend, which is used to quantify the development stage of the fault.

[0048] The system acquires the historical operating data of the device stored locally, and uses the adaptation module to perform personalized correction on the initial prediction vector using the historical operating data of the device, thereby generating a corrected prediction vector.

[0049] It should be noted that the historical operating data of this equipment records the unique operating trajectory and performance degradation information of this specific wind turbine. This personalized information is used to perform fine-tuning and correction on the initial prediction vector, thereby generating a corrected prediction vector that better matches the actual condition of this equipment. The adaptation module can be updated online.

[0050] The specific implementation is as follows: The adaptation module first filters the historical operating data of the device stored locally, extracts historical time-series data that is consistent with the current operating mode and covers normal and various fault states, and constructs a personalized sample set after standardization and noise reduction preprocessing; based on this sample set, a device-specific correction model is generated by training the gradient descent algorithm. The core of the correction model is the mapping relationship weight between each dimension of the initial prediction vector and the actual state label of the device in the historical data; during correction, the adaptation module inputs the initial prediction vector into the correction model, and dynamically adjusts the weights of each dimension of the initial prediction vector, such as fault probability, fault correlation, and feature anomaly contribution value, in combination with the differences between the current real-time time-series features and historical samples, correcting the generalization prediction bias; finally, the corrected prediction vector that matches the operating characteristics of the device is output, improving the personalized accuracy of fault diagnosis. The actual state label refers to the supervision label that corresponds to the extracted historical time-series data and quantifies the actual operating state of the device. Its dimensions are precisely aligned with the indicators of the initial prediction vector, such as fault probability, fault correlation, and feature anomaly contribution value, and are used to provide supervision for the training of the correction model.

[0051] The fault prediction result is output based on the corrected prediction vector.

[0052] It should be noted that the personalized correction prediction vector is fed into the output layer of the model and decoded into a clear fault prediction result. This result clearly includes the specific probability of fault occurrence, such as 80%, and the specific fault type, such as bearing wear or abnormal gearbox oil temperature.

[0053] The decoding here refers to the process by which the model output layer, through predefined mapping rules and parsing logic, transforms the personalized, quantitatively-characteristic prediction vector into a concrete fault diagnosis result that the operation and maintenance system can directly understand. Specifically, based on the predefined correspondence between vector dimensions and fault types, as well as the conversion rules between numerical values ​​and probability percentages, the output layer parses the quantitative values ​​of each dimension in the corrected prediction vector: on the one hand, it converts the values ​​reflecting the probability of a fault into explicit percentage forms, such as parsing 0.8 in the corresponding dimension of the vector as an 80% probability of a fault occurring; on the other hand, it matches the dimensional values ​​pointing to fault features to specific fault type identifiers, such as parsing the dimensional values ​​corresponding to abnormal mechanical components as explicit fault names such as bearing wear or abnormal gearbox oil temperature. Finally, it outputs a structured and unambiguous fault prediction result, realizing the transformation from an abstract quantitative vector to a concrete diagnostic conclusion.

[0054] The underlying principle is the construction of an intelligent diagnostic framework that combines general knowledge with individual experience. A pre-trained backbone network utilizes extensive fault knowledge learned from numerous devices for rapid initial assessment. Then, an online-updable adaptation module combines this general judgment capability with the unique lifecycle characteristics of each individual device for personalized and precise correction. The beneficial effects of this method are improved accuracy and reliability of fault prediction, as the prediction results are tailored to the specific condition of each device. Simultaneously, because the adaptation module can be updated online, the model can self-learn and evolve as the device operates and ages, ensuring that its diagnostic capabilities remain at a high level throughout the device's lifecycle. This effectively solves the problem of decreased diagnostic accuracy caused by traditional static models' inability to adapt to changes in device condition.

[0055] The local decision-making and control module is connected to the AI ​​fault diagnosis module. It is used to determine the fault prediction results based on preset judgment rules, and generate and issue corresponding local control commands or warning signals when the judgment meets the warning or control conditions.

[0056] In a specific embodiment of the present invention, the fault prediction result is judged based on a preset judgment rule, and when the judgment meets the warning or control conditions, a corresponding local control command or warning signal is generated and issued. Specifically, the fault occurrence probability and the fault type are obtained from the fault prediction result.

[0057] Based on the fault type, query the stored control strategy library to obtain at least one candidate control strategy.

[0058] It should be noted that the control strategy library is a structured database storing corresponding handling solutions for various faults in wind turbines. It is used by the local decision-making and control modules to quickly match executable safety handling strategies after fault identification, ensuring the standardization, accuracy, and timeliness of fault handling. Specifically, it classifies the impact level based on the fault types identifiable by the AI ​​fault diagnosis module, combined with the fault occurrence probability, impact range, and development speed; based on equipment operation and maintenance specifications, safe operation thresholds, and engineering practices, it clarifies control objectives for various faults and levels, and formulates targeted control strategies including execution conditions, operation steps, and parameter thresholds; it uses a three-dimensional mapping structure of fault type-impact level-control strategy for structured storage, supporting rapid indexing; it regularly collects fault handling effect data and operation and maintenance feedback to optimize handling logic and parameter thresholds, achieving dynamic iterative updates of the strategy library and ensuring the adaptability and effectiveness of the strategies.

[0059] For example, if the fault type is abnormal power output, the impact level is moderate, and the triggering conditions are a fault probability of 70%-85% and the equipment is in its rated operating phase, three candidate control strategies are proposed: Strategy 1 focuses on maintaining power generation efficiency, with control logic involving fine-tuning the blade angle by ±2° to optimize wind energy capture efficiency, simultaneously monitoring power output fluctuations, and recording recovery status every 5 minutes; Strategy 2 focuses on equipment protection, with control logic involving reducing power from 800-1000kW to 700-800kW, reducing the load on core components, and initiating electrical system status monitoring; Strategy 3 focuses on rapid troubleshooting, with control logic involving a brief 10-minute switch to low-load operation to detect the matching degree between grid voltage, wind speed, and equipment operating parameters, generating a parameter adaptation analysis report. Subsequently, the three strategies are simulated and executed to evaluate the impact of each strategy on equipment health status (e.g., component temperature rise, vibration changes) and power generation efficiency (e.g., power recovery rate, energy consumption ratio), and the strategy evaluation results are output.

[0060] If the fault type is abnormal vibration signal, the impact level is moderate, and the triggering conditions are a fault probability of 75%-85% and a fault development stage of level 2 (initial fault). Two candidate control strategies are proposed: Strategy 1 prioritizes vibration suppression, with control logic involving activating the active vibration reduction module, adjusting damping parameters, reducing bearing vibration amplitude, and continuously collecting effective vibration values ​​and component temperature data; Strategy 2 prioritizes load optimization, with control logic involving a slight reduction in rotational speed of 5%-8% to reduce vibration caused by mechanical friction, while simultaneously monitoring power output loss. By simulating the execution of both strategies, the reduction in equipment health risk and the proportion of power generation efficiency loss are compared and analyzed to generate a basis for recommending the optimal strategy.

[0061] By combining the current operating load of the equipment with environmental parameters, each candidate control strategy is simulated and executed to predict its impact on the health status and power generation efficiency of the equipment, and strategy evaluation results are generated.

[0062] It should be noted that, based on the current operating load of the equipment and environmental parameters, and using a pre-set rule base corresponding to the fault type and the candidate control strategies, including equipment operation mechanism models, fault impact association rules, and efficiency loss calculation logic, the predicted change trends of equipment health status such as component wear rate, vibration amplitude change, fault development delay degree, and power generation efficiency such as power output stability, energy consumption ratio, and efficiency loss rate are deduced after the execution of each candidate control strategy. In this way, the comprehensive impact of different candidate control strategies on equipment health status and power generation efficiency is quantitatively evaluated through weighted fusion, and a strategy evaluation result including strategy priority, implementation risk level, and comprehensive benefit score is generated.

[0063] Based on the probability of failure and the strategy evaluation results, the optimal temporary handling plan is selected.

[0064] Specifically, the optimal temporary response plan is selected as follows: Based on the fault occurrence probability classification into three emergency levels—high ≥80%, medium 60%-80%, and low <60%—and combined with the implementation risk level, strategy priority, and comprehensive benefit score in the strategy evaluation results, the selection is carried out according to the progressive logic of risk controllability → priority matching → optimal benefit: for high emergency levels, candidate strategies with high implementation risk levels are excluded first, and only low / medium risk strategies are retained; for medium / low emergency levels, medium risk strategies are compatible, and then the strategies that meet the risk threshold are sorted in descending order of priority. If there are strategies with the same priority, the one with the highest comprehensive benefit score is used as the final judgment criterion. Finally, the optimal temporary response plan that simultaneously satisfies the requirements of fault urgency matching, implementation risk controllability, and comprehensive balance between equipment safety and power generation efficiency is determined.

[0065] When it is determined that the triggering conditions of the optimal temporary handling plan are met, the plan is executed and the local control command or the warning signal is generated.

[0066] Specifically, if the previously obtained probability of failure has exceeded a threshold that is dynamically adjusted based on the equipment status, or if the preset triggering conditions in the selected optimal temporary handling plan are met, then the system will no longer wait, immediately execute the plan, and generate corresponding local control commands or warning signals. The former directly acts on the equipment controller for adjustment, while the latter notifies the on-site personnel.

[0067] It should be noted that the specific implementation of this dynamic threshold adjustment is as follows: Based on a preset threshold adjustment model, which integrates the fitting rules of equipment operation mechanism and historical operation and maintenance data, the system collects real-time data on the current operating mode of the equipment (e.g., rated load / low load), real-time health status (e.g., component wear rate, fault development stage coefficient), historical data on the handling effect of similar faults, and environmental parameters (e.g., temperature, humidity, wind speed). The system dynamically calculates the appropriate threshold through the following logic: If the equipment is in a high-load operating mode or the health status score is low (e.g., component wear rate exceeds the benchmark value by 30%), the threshold is lowered to improve fault response sensitivity; if the equipment health status is good and the historical false trigger rate of similar faults is higher than the preset value, the threshold is appropriately raised; at the same time, the threshold gradient is corrected by combining the fault escalation rate under similar operating conditions in historical data, ensuring that the threshold accurately matches the real-time operating condition and health level of the equipment, taking into account both the timeliness of fault response and the need to suppress false triggers.

[0068] The principle behind the above is to introduce a simulation-based, proactive decision-making mechanism at the equipment site. This enables the centralized controller gateway not only to detect problems but also to autonomously predict and select the optimal temporary solution. Its benefits lie in enhancing the system's autonomy and intelligence in responding to sudden risks, allowing it to take the most appropriate temporary control measures at the initial stage of a fault without human intervention. This immediate and optimized local intervention effectively suppresses the further development of the fault, buying valuable time for subsequent remote maintenance. Simultaneously, while ensuring equipment safety, it minimizes power generation losses caused by unnecessary downtime or excessive power restriction, achieving a dynamic balance between equipment protection and production efficiency.

[0069] In a specific embodiment of the present invention, the determination of whether the triggering condition of the optimal temporary treatment plan is met includes comparing the probability of the fault occurrence with a dynamic adjustment threshold. The method for determining the dynamic adjustment threshold includes monitoring the time change trend of the multidimensional health vector recently generated by the device.

[0070] It should be noted that the multidimensional health vector refers to a set of multidimensional key parameters that are integrated in vector form to quantify the health status of various key systems of equipment. It is used to predict failure risks through time-series change trends. Its dimensions specifically include: wear rate, effective value of vibration amplitude, temperature rise rate, and failure development stage coefficient of key equipment components such as bearings, converters, and blades, as well as quantifiable indicators such as energy consumption stability of electrical systems, mechanical transmission efficiency, and key parameters such as voltage and speed deviation rate. Each dimension parameter corresponds to a certain key health attribute of the equipment, and the vector as a whole can comprehensively reflect the overall health level of the equipment.

[0071] Based on the time change trend and the obtained cumulative operating time of the equipment, the health status decay coefficient is calculated.

[0072] It should be noted that the specific calculation method is as follows: Taking the multi-dimensional health vector of the equipment in a newly manufactured or fault-free state as the benchmark value, the time change trend is first quantified, that is, the rate of change of each dimension of health parameters with time is calculated, such as the wear rate increase value and the vibration amplitude increase rate within a unit cumulative operating time. Then, the rate of change of each dimension is weighted and integrated with the cumulative operating time of the equipment. The weight of the cumulative operating time is preset according to the decay law of the equipment life cycle. The longer the operating time, the higher the weight. The normalization calculation is completed by the formula health status decay coefficient = Σ (the rate of change of each dimension of health parameters × the corresponding weight) / benchmark value × 100%. The final coefficient value is in the range of 0-1. The closer it is to 1, the more serious the decay. It can quantify the degree of decay of the equipment health status with the cumulative operating time.

[0073] A basic threshold for risk assessment is obtained, and the basic threshold is combined with the health status decay coefficient to obtain the dynamic adjustment threshold.

[0074] It should be noted that the health status safety threshold calibrated at the device's factory is used as the base threshold. A preset sensitivity adjustment coefficient, such as 0.6-0.9, is introduced to adapt to different device types and can be fine-tuned. The dynamic adjustment threshold is calculated using the formula: Base Threshold × (1 - Health Status Decay Coefficient × Sensitivity Adjustment Coefficient). The health status decay coefficient directly reflects the current degree of health decay of the device: when the coefficient is close to 1, indicating severe health decay, the dynamically adjusted threshold is significantly lower than the base threshold, improving fault response sensitivity; when the coefficient is close to 0, indicating good device health, the dynamically adjusted threshold is close to the base threshold, avoiding false triggering. This calculation logic uses the health status decay coefficient to perform scenario-based correction of the base threshold, achieving precise adaptation of the threshold to the device's real-time health level.

[0075] The principle behind the above lies in introducing an adaptive early warning mechanism for equipment health status awareness. This transforms a fixed alarm threshold into an intelligent threshold that dynamically adjusts based on the recent performance degradation trend and lifecycle stage of the equipment. The beneficial effect is that it makes the early warning system more sensitive to slow and continuously deteriorating, insidious faults, triggering responses at earlier stages of fault manifestation. It also has a stronger ability to suppress false alarms caused by temporary fluctuations in the health of equipment. This dynamic risk assessment strategy, tailored to each piece of equipment, improves the accuracy and foresight of early warnings, enabling local decision-making to more rationally balance the timeliness of response with unnecessary production intervention, thereby optimizing the effectiveness of the entire predictive maintenance system.

[0076] The communication module, connected to the edge computing processing module and the local decision and control module, is used to transmit the fault prediction results, the early warning signals and the processed operating status data to the remote operation and maintenance system, and to receive remote instructions from the remote operation and maintenance system.

[0077] In a specific embodiment of the present invention, the step of transmitting the fault prediction result, the early warning signal, and the processed operating status data to the remote operation and maintenance system specifically involves: encapsulating the fault prediction result, the early warning signal, the time-series feature dataset, and the local decision log to generate an upload data packet.

[0078] Add a device identifier and a data priority tag to the uploaded data packet.

[0079] It should be noted that the data priority label is a hierarchical identifier based on the urgency and importance of the uploaded data for remote operation and maintenance decisions. The core purpose is to enable the remote operation and maintenance system to process data packets in an orderly manner according to priority, ensuring that critical data is responded to first and non-critical data is used reasonably for resources, thus adapting to the real-time and bandwidth optimization requirements of wind turbine operation and maintenance.

[0080] For example, in conjunction with the technical scenario of the present invention, data priority tags are divided according to the following logic: High-priority tags: used for critical data such as emergency fault warning signals, severe fault prediction results, and real-time control command feedback. For example, severe bearing wear faults + shutdown warnings require immediate handling by the remote system to prevent the fault from worsening.

[0081] Medium priority label: Used for important data such as medium fault prediction results, equipment health status reports, and strategy evaluation results, such as power output abnormality and medium fault handling logs, which need to be responded to in a short time to support the formulation of maintenance plans.

[0082] Low priority tags: Used for non-urgent data such as routine operating status data, historical fault archive logs, and non-critical parameter statistics, such as routine parameter summaries for 24 hours of operation under rated load. Processing can be delayed to save bandwidth resources.

[0083] After tags are bound to data packets, remote operation and maintenance systems can quickly sort data based on tags, prioritize the scheduling of computing power to process high-priority data, and achieve orderly and efficient data transmission and processing.

[0084] The uploaded data packet is sent to the cloud service platform of the remote operation and maintenance system via the communication module.

[0085] The principle behind the above is to achieve an efficient and intelligent edge-to-cloud data communication mechanism. It abandons the inefficient practice of transmitting massive amounts of raw data, instead performing in-depth processing and refinement on-site, packaging and reporting only the most valuable analysis results, decision records, and key feature data. The beneficial effects include reduced network bandwidth consumption, reduced computational pressure on the cloud platform, and improved timeliness of reporting critical fault information. By adding priority tags to data packets, it ensures that the most urgent alarms are handled first by the remote system, thus providing the remote operations and maintenance team with a clear, complete decision-making scenario with accompanying handling suggestions, effectively avoiding information overload, and laying a solid data foundation for rapid and accurate remote intervention and resource scheduling.

[0086] The technical principle of this invention lies in constructing an intelligent operation and maintenance architecture that deeply collaborates between the edge and the cloud. The intelligence of fault diagnosis is brought forward to the centralized controller gateway at the wind turbine site, enabling it not only to collect data but also to process it locally. It uses artificial intelligence models to predict faults in real time and autonomously executes preliminary control or early warning based on the prediction results. Subsequently, it uploads these refined key prediction results and decision information to the remote operation and maintenance system. The remote operation and maintenance system acts as a global command and resource optimization center. It receives precise intelligence from the field, automatically transforms it into specific preventative maintenance work orders, and further combines it with global resource status for intelligent analysis and scheduling. Finally, it generates and issues the optimal maintenance plan, thus forming a complete closed-loop automated process from data perception, on-site prediction, local handling, cloud planning, and resource scheduling.

[0087] The technical benefits of this invention are: improved operation and maintenance efficiency and reliability of wind power equipment. By performing real-time analysis and prediction at the equipment site, early detection and rapid response to potential faults are possible, effectively avoiding missed opportunities for optimal intervention due to data transmission delays or cloud processing congestion, thus suppressing the escalation of faults. Simultaneously, seamless integration of fault prediction results with maintenance work order generation and resource scheduling achieves full-process automation from diagnosis to action, shortening the maintenance response cycle and reducing reliance on manual analysis and decision-making. This combination of edge intelligence and cloud intelligence transforms operation and maintenance activities from passive fault repair to proactive preventative maintenance, ultimately improving equipment availability, optimizing maintenance resource allocation, and enhancing the overall operational efficiency of the wind farm.

[0088] Reference Figure 2 As shown, a second aspect of the present invention provides a remote operation and maintenance system configured to utilize the centralized controller gateway for fusion AI fault diagnosis, comprising: the centralized controller gateway, which is deployed on the wind farm side to perform real-time data acquisition, edge-side fault prediction and local decision-making for at least one wind turbine.

[0089] A cloud service platform, communicatively connected to at least one of the central controller gateways, the cloud service platform comprising: The model management and training unit is used to issue or update the fault prediction model to the centralized controller gateway, and to retrain the model based on historical operating data and fault data collected from multiple centralized controller gateways to generate the corresponding model update package.

[0090] In a specific embodiment of the present invention, the step of retraining the model to generate the corresponding model update package specifically involves: extracting the time-series feature dataset, the fault prediction results, and the corresponding actual operating results from the collected uploaded data packets to construct a model retraining dataset.

[0091] For example, the triggering condition for model retraining is as follows: Model retraining is started when any of the following conditions are met: ① 500 sets of valid uploaded data packets are collected in total; ② The average accuracy of 30 consecutive sets of fault prediction results is <95%; ③ Two or more new uncovered fault types are added; The upper limit of the number of model retraining iterations is 500 times. If the model does not converge after 500 iterations, the current optimal parameters are used to generate the model update package.

[0092] The backbone network in the fault prediction model is retrained based on the model retraining dataset to generate optimized backbone network parameters.

[0093] It needs to be explained that the backbone network is the basic architecture of general deep learning networks such as CNN and LSTM. The layer structure and operation logic can serve as the initial framework of the backbone network. However, its original parameters are generalized parameters generated based on general datasets and cannot accurately match the feature distribution of equipment fault prediction scenarios, such as vibration signals and power time series data. The optimized backbone network parameters are obtained through iterative optimization using a fault prediction-specific retraining dataset, which includes equipment operation features, fault samples, and operating parameters. Essentially, it is a scenario-specific adaptation and adjustment of general network parameters. It retains the basic operation architecture of the general network and, through retraining, makes the parameters accurately match the feature extraction requirements of fault prediction, replacing the generalized parameters of the general network. Ultimately, it achieves the adaptation and implementation of the general network architecture to the fault prediction scenario, ensuring the accuracy and specificity of the model prediction.

[0094] If the backbone network is a Convolutional Neural Network (CNN), it includes the kernel weights / biases of each convolutional layer, the sampling window size and stride of the pooling layer, and the activation function parameters. If it is a Recurrent Neural Network (RNN / LSTM / GRU), it includes the hidden layer weight matrix, forget gate / input gate / output gate adjustment parameters, and the state transition matrix. It also includes the mean / variance parameters of the batch normalization layer, the inactivation probability of the regularization layer, and other regularization parameters, as well as the inter-layer connection weights and output layer mapping parameters. After retraining and optimization, these parameters can improve the backbone network's accuracy in extracting equipment fault features and its model generalization ability, ensuring the specificity and accuracy of fault prediction.

[0095] The individualized correction patterns of the adaptation modules in multiple centralized controller gateways were analyzed, and common optimization strategies were extracted.

[0096] Specifically, this involves analyzing the personalized calibration patterns of adaptation modules in multiple centralized controller gateways and extracting common optimization strategies. Specifically, this includes collecting personalized calibration data from each adaptation module, including calibration triggering conditions, parameter adjustment logic, effect feedback, and operational status information. This data is then grouped and statistically analyzed by gateway model, adaptation device type, load level, and other dimensions. Common sensitive influencing factors, overlapping parameter adjustment ranges, and unified conditions for achieving calibration results in different modules are identified. This process then refines general optimization rules applicable to most modules, including unified adaptation parameter benchmark values, standardized calibration triggering logic, common protocol adaptation templates compatible with multiple device types, and a load-balanced calibration scheduling mechanism. This ensures that common strategies cover and adapt to the personalized needs of each module.

[0097] The cloud periodically collects the update amount ΔW of the calibration matrix W of each gateway adaptation module. A clustering algorithm is used to analyze the distribution pattern of ΔW, classifying devices with similar ΔW values ​​into groups. For each group, the mean vector ΔW_mean of ΔW is calculated, and this mean vector is encapsulated as a component of the common optimization strategy for that group of devices. This provides better initial calibration parameters when updating new devices or models of that group.

[0098] The optimized backbone network parameters and the common optimization strategy are encapsulated into the model update package.

[0099] The principle behind the above is to establish a two-way empowering cycle of individual learning and collective evolution. The adaptation module of the on-site gateway learns individually to adapt to the characteristics of a single device, while the cloud platform aggregates the learning experiences of all individuals to achieve collective evolution, refining more universal knowledge and more efficient learning methods. The beneficial effect is that it breaks the limitation of traditional AI models whose performance degrades over time after a single deployment, enabling the entire diagnostic system to possess self-evolution and continuous learning capabilities. Through this cloud-edge collaborative model, the overall diagnostic accuracy of the system not only does not decrease but also continuously improves with the increase in the number of connected devices and the length of runtime, forming a virtuous cycle of increasing accuracy with use. This ensures that the diagnostic system can always adapt to equipment aging, environmental changes, and newly emerging unknown fault modes, achieving forward-looking and robust diagnostic capabilities.

[0100] The remote operation and maintenance management unit is used to receive and centrally display the fault prediction results, early warning signals and operating status data from the centralized controller gateway, and generate preventive maintenance work orders based on the fault prediction results.

[0101] In a specific embodiment of the present invention, the step of generating a preventive maintenance work order based on the fault prediction result specifically involves receiving and parsing the uploaded data packet from the centralized controller gateway.

[0102] When the fault risk in the fault prediction result meets the work order generation conditions, the system associates the stored equipment historical fault database and maintenance knowledge base to generate a primary maintenance work order containing recommended measures and a list of required resources.

[0103] For example, when the probability of a fault occurrence in the fault prediction result is greater than or equal to a preset work order triggering threshold (e.g., 60%) and the fault impact level is greater than or equal to moderate, the work order generation conditions are met. The equipment historical fault database stores historical fault records of this equipment and equipment of the same model, including fault type, fault cause, past handling solutions and implementation effect data. The maintenance knowledge base stores equipment operation and maintenance specifications, spare parts model parameters, standard operating procedures and expert handling experience. The system generates a primary maintenance work order by matching the fault type with the best handling cases of similar faults in the historical fault database, combined with the standard requirements of the maintenance knowledge base. The primary maintenance work order includes targeted recommended measures, such as replacing bearing of model XX bearing for bearing wear faults, lubrication system maintenance, and a list of required resources including spare parts models, tool specifications, and manpower configuration.

[0104] Analyze the local decision log in the uploaded data packet to assess the available maintenance time window.

[0105] Based on the local decision log, the optimal temporary handling plan executed by the gateway is obtained, and based on the stabilizing effect of the plan on the device status, the maintenance time window available for repair is accurately evaluated.

[0106] The specific implementation method for assessing the available maintenance time window is as follows: Based on the fault type, impact level, execution effect of the optimal temporary disposal plan (such as fault suppression degree), equipment health stability, current operating mode (such as rated load / low load), and effective maintenance duration of the temporary disposal plan contained in the local decision log in the uploaded data packet, combined with maintenance adaptation records of similar faults in historical operation and maintenance, the following logic is used for evaluation: First, extract the longest duration during which the temporary disposal plan can stably suppress the fault deterioration, i.e., the latest maintenance start time limit; then, screen the time period when the equipment operating load is low and environmental parameters such as wind speed and temperature are suitable for maintenance operations, excluding the period when power consumption is high, extreme weather, etc., which are not suitable for downtime maintenance; finally, combine the fault development trend prediction (such as whether the fault will exceed the safety threshold under temporary disposal) to determine the continuous time period that satisfies both the effective time limit of temporary disposal and minimizes power generation loss and ensures the safety of maintenance operations, which is the available maintenance time window.

[0107] The maintenance time window is merged with the primary maintenance work order to generate the preventive maintenance work order.

[0108] It's important to note that the principle behind the above lies in automating and intelligently managing the entire process from fault warning to maintenance task creation. By deeply integrating predictive data, historical knowledge, and real-time on-site handling information, an abstract risk signal is transformed into a specific, detailed maintenance task with a clearly defined execution time window. Its beneficial effects include reducing the time delay and manpower costs between problem discovery and solution development, ensuring high accuracy and executability of maintenance tasks from the outset. By clearly defining the required resources and optimal timing for action in advance, the success rate of initial repairs is improved, avoiding secondary tower operations due to insufficient preparation, thereby achieving efficient utilization of maintenance resources and minimizing equipment downtime.

[0109] The operation and maintenance coordination scheduling unit is used to analyze and schedule maintenance resources based on the content of the preventive maintenance work order, generate an optimized maintenance plan, and send it to the target user terminal.

[0110] In a specific embodiment of the present invention, the step of performing maintenance resource analysis and scheduling based on the content of the preventive maintenance work order to generate an optimized maintenance plan specifically includes: obtaining all the preventive maintenance work orders to be processed and extracting their fault risk level and maintenance time window.

[0111] Obtain information on available maintenance personnel status, spare parts inventory, and external environment.

[0112] The fault risk level, the maintenance time window, the maintenance personnel status, the spare parts inventory information, and the external environment information are input into a multi-objective dynamic optimization model that aims to minimize overall power generation loss and maximize resource utilization, and the optimized maintenance plan is generated.

[0113] It should be noted that the multi-objective dynamic optimization model is a pre-trained dedicated model. The training and solution process is as follows: During training, historical maintenance data of the same type of equipment, fault handling records, power generation efficiency statistics, and resource scheduling logs are used as training samples. The model input is defined as five types of constraint parameters, including fault risk level and maintenance time window. The dual objective functions are minimizing overall power generation loss (quantified as power generation revenue loss during maintenance period) and maximizing resource utilization (quantified as personnel / spare part idle rate). Constraints including work safety threshold, resource supply limit, and time window boundary are constructed. The weighted summation method is used to transform the multi-objective problem into a single-objective optimization problem. The model parameters are iteratively optimized using the gradient descent algorithm until the deviation between the predicted output and the historical best maintenance plan is lower than the preset threshold, thus completing the model training.

[0114] The optimized maintenance plan output after solving includes: the optimal maintenance start time, the appropriate maintenance personnel configuration scheme including the number of skilled personnel, the on-duty sequence, the spare parts requisition list including the model, quantity and allocation path, the phased operation process, the estimated downtime and the power generation loss control target.

[0115] It's important to note that the principle behind the above is to transform the complex operation and maintenance scheduling problem into a multi-objective optimization problem solvable through a mathematical model. This integrates previously isolated fault warning, resource management, and production planning into a unified whole, allowing for comprehensive coordination and strategic maneuvering. The beneficial effect lies in changing the traditional first-come-first-served or purely experience-based scheduling model, enabling data-driven decisions that optimize the economic benefits and operational efficiency of the entire wind farm. For example, it can intelligently merge multiple geographically close maintenance tasks with the same team, or schedule non-urgent maintenance during predicted low-wind-speed periods. This ensures equipment reliability while improving the utilization efficiency of maintenance resources and reducing power generation losses due to downtime, achieving a strategic transformation of operation and maintenance activities from passive cost expenditure to proactive value creation.

[0116] It should also be noted that the execution priority order is as follows: local emergency adjustments, such as the optimal temporary handling plan for fault triggering, > cloud-based global maintenance adjustments, such as preventive maintenance plans and parameter optimization instructions, following the principles of safety first and emergency first.

[0117] Priority application conditions: ① When the equipment has an immediate risk of failure, i.e., the probability of failure is greater than or equal to the dynamic adjustment threshold, the local adjustment is executed immediately and autonomously, and conflicting cloud commands are blocked; ② When the equipment is not faulty or the local adjustment has been terminated, the cloud adjustment command is executed first; ③ When the two conflict, the local adjustment corresponding to the high-risk failure takes priority, and the local adjustment corresponding to the medium- and low-risk failure temporarily stores the cloud command and executes it after it terminates.

[0118] Priority exception: If local adjustments cause the device's health status to deteriorate or execution to malfunction, local adjustments will be automatically suspended and reported to the cloud. Emergency intervention commands from the cloud will temporarily have the highest priority.

[0119] By implementing the above priority rules, we can ensure that the equipment responds quickly in emergency scenarios and achieves global optimization in normal operation and maintenance scenarios, thus avoiding security risks or operation and maintenance failures caused by adjustment conflicts.

[0120] It should be added that the formulas mentioned above, through the principle of dimensional consistency and mathematical standardization methods such as normalization, dimensionless parameter conversion, or unit system unification, can translate physical quantities with different properties into unitless standard values ​​or parameters that can be superimposed in the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the invention.

[0121] It should also be noted that the setting basis for the other types of thresholds described in this patent application, including but not limited to fault prediction probability thresholds, health status decay coefficient thresholds, maintenance work order trigger thresholds, temporary handling effective maintenance duration thresholds, and multi-objective optimization model constraint thresholds, is based on historical data of wind turbine operation, fault handling records, and a large number of simulation tests and field experimental results. At the same time, it combines the experience of experts in the field of wind turbine operation and maintenance, and comprehensively considers factors such as the actual operating environment of the equipment, such as wind speed, temperature, humidity, the design life and performance indicators of key components of the equipment such as bearings, blades, and converters, and the grid operation requirements, to calibrate the thresholds. It has a solid scientific basis and practical operability, and the relevant threshold calibration technology in the prior art is relatively mature, so it will not be elaborated here.

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0125] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A centralized controller gateway integrating AI fault diagnosis, characterized in that, include: The data acquisition module is used to acquire real-time operating status data from multiple sensors of at least one connected wind turbine, wherein the sensors include at least a vibration sensor, a speed sensor and a power sensor. The edge computing processing module is used to preprocess and extract features from the running status data to generate a time-series feature dataset; The AI ​​fault diagnosis module is integrated into the edge computing processing module and includes a pre-set fault prediction model. The AI ​​fault diagnosis module is used to input the time-series feature dataset into the fault prediction model and output the fault prediction result of the wind turbine. The fault prediction result includes at least the probability of fault occurrence and the fault type. The local decision-making and control module is connected to the AI ​​fault diagnosis module. It is used to judge the fault prediction results based on preset judgment rules, and generate and issue corresponding local control commands or warning signals when the judgment meets the warning or control conditions. The communication module, connected to the edge computing processing module and the local decision and control module, is used to transmit the fault prediction results, the early warning signals and the processed operating status data to the remote operation and maintenance system, and to receive remote instructions from the remote operation and maintenance system.

2. The centralized controller gateway integrating AI fault diagnosis according to claim 1, characterized in that, The preprocessing and feature extraction of the running status data to generate a time-series feature dataset specifically includes: Acquire raw data streams from vibration sensors, speed sensors, and power sensors; The original data stream is cleaned and denoised to generate regular time-series data; Identify the current operating mode of the device reflected in the regularized timing data; Select the corresponding feature calculation rule from the feature calculation rule set stored in the dynamic feature library according to the current operating mode of the device; The regularized time series data are calculated using the aforementioned feature calculation rules to generate the time series feature dataset.

3. The centralized controller gateway integrating AI fault diagnosis according to claim 1, characterized in that, The specific content of inputting the time-series feature dataset into the fault prediction model and outputting the fault prediction result of the wind turbine is as follows: The time-series feature dataset is input into the fault prediction model, which includes a backbone network and an adaptation module. The time-series feature dataset is processed by the backbone network to generate an initial prediction vector; The system acquires the historical operating data of the device stored locally, and uses the adaptation module to perform personalized correction on the initial prediction vector using the historical operating data of the device to generate a corrected prediction vector. The fault prediction result is output based on the corrected prediction vector.

4. The centralized controller gateway integrating AI fault diagnosis according to claim 1, characterized in that, The fault prediction result is judged based on preset judgment rules, and when the judgment meets the early warning or control conditions, a corresponding local control command or early warning signal is generated and issued, the specific content of which is as follows: The probability of occurrence of the fault and the type of fault are obtained from the fault prediction results; Based on the fault type, query the stored control strategy library to obtain at least one candidate control strategy; By combining the current operating load of the equipment with environmental parameters, each candidate control strategy is simulated to predict its impact on the health status of the equipment and power generation efficiency, and strategy evaluation results are generated. Based on the probability of failure and the strategy evaluation results, the optimal temporary handling plan is selected; When it is determined that the triggering conditions of the optimal temporary handling plan are met, the plan is executed and the local control command or the warning signal is generated.

5. The centralized controller gateway integrating AI fault diagnosis according to claim 1, characterized in that, The specific content of transmitting the fault prediction results, the early warning signal, and the processed operating status data to the remote operation and maintenance system is as follows: The fault prediction result, the early warning signal, the time-series feature dataset, and the local decision log are encapsulated to generate an upload data packet; Add a device identifier and a data priority tag to the uploaded data packet; The uploaded data packet is sent to the cloud service platform of the remote operation and maintenance system via the communication module.

6. A remote operation and maintenance system configured to utilize the centralized controller gateway with integrated AI fault diagnosis as described in claim 1, characterized in that, include: The centralized controller gateway as described in claim 1 is used to be deployed on the wind farm side to perform real-time data acquisition, edge-side fault prediction and local decision-making for at least one wind turbine. A cloud service platform, communicatively connected to at least one of the central controller gateways, the cloud service platform comprising: The model management and training unit is used to send or update the fault prediction model to the centralized controller gateway, and to retrain the model based on historical operating data and fault data collected from multiple centralized controller gateways to generate the corresponding model update package. The remote operation and maintenance management unit is used to receive and centrally display the fault prediction results, early warning signals and operating status data from the centralized controller gateway, and generate preventive maintenance work orders based on the fault prediction results. The operation and maintenance coordination scheduling unit is used to analyze and schedule maintenance resources based on the content of the preventive maintenance work order, generate an optimized maintenance plan, and send it to the target user terminal.

7. The remote operation and maintenance system integrating AI fault diagnosis according to claim 6, characterized in that, The preventive maintenance work order generated based on the fault prediction results has the following specific content: Receive and parse the uploaded data packet from the central controller gateway; When the fault risk in the fault prediction result meets the work order generation conditions, the stored equipment historical fault database and maintenance knowledge base are associated to generate a primary maintenance work order containing recommended measures and a list of required resources. Analyze the local decision log in the uploaded data packet to assess the available maintenance time window; The maintenance time window is merged with the primary maintenance work order to generate the preventive maintenance work order.

8. The remote operation and maintenance system integrating AI fault diagnosis according to claim 6, characterized in that, The step involves analyzing and scheduling maintenance resources based on the content of the preventive maintenance work order to generate an optimized maintenance plan, the specific content of which is as follows: Obtain all pending preventative maintenance work orders and extract their fault risk level and maintenance time window; Obtain information on available maintenance personnel status, spare parts inventory, and external environment; The fault risk level, the maintenance time window, the maintenance personnel status, the spare parts inventory information, and the external environment information are input into a multi-objective dynamic optimization model that aims to minimize overall power generation loss and maximize resource utilization, and the optimized maintenance plan is generated.

9. The remote operation and maintenance system integrating AI fault diagnosis according to claim 6, characterized in that, The specific content of retraining the model to generate the corresponding model update package is as follows: Extract the time-series feature dataset, the fault prediction results, and the corresponding actual operating results from the collected uploaded data packets, and construct a model retraining dataset; Based on the model retraining dataset, the backbone network in the fault prediction model is retrained to generate optimized backbone network parameters. Analyze the individualized correction patterns of the adaptation modules in multiple centralized controller gateways and extract common optimization strategies; The optimized backbone network parameters and the common optimization strategy are encapsulated into the model update package.

10. The centralized controller gateway integrating AI fault diagnosis according to claim 4, characterized in that, The determination that the triggering condition of the optimal temporary handling plan is met includes comparing the probability of the fault occurrence with a dynamically adjusted threshold, wherein the method for determining the dynamically adjusted threshold includes: The time-varying trend of the multidimensional health vector recently generated by the monitoring equipment; Based on the time change trend and the obtained cumulative equipment operating time, calculate the health status decay coefficient; A basic threshold for risk assessment is obtained, and the basic threshold is combined with the health status decay coefficient to obtain the dynamic adjustment threshold.

Images