Wind turbine generator unit diagnosis system and method based on nested memory self-supervised learning

Abstract

The application discloses a wind turbine generator unit diagnosis system and method based on nested memory self-supervised learning, and the system comprises a data acquisition module, a cross-scale alignment and segmentation module for establishing hierarchical time indexes and time axis alignment, a multi-scale representation coding module for extracting feature representations of the multiple time scale layers, a continuous spectrum memory (CMS) module, surprise degree generated by a short-term layer triggers writing to a medium-term scenario memory, stable offset of the medium-term layer triggers consolidation to a long-term baseline memory, a self-supervised training module for obtaining an early warning model, and an inference and alarm module based on cross-scale anomaly scores, degradation trend scores and early warning levels. The application captures weak transient anomalies at a millisecond level or even a smaller level through construction of three time scale layers of a short-term layer, a medium-term layer and a long-term layer, and avoids early warning lag caused by early weak fault signals being submerged by means of a nested writing-consolidation mechanism of the continuous spectrum memory (CMS) module.

Description

Technical Field

[0001] This application relates to the field of wind turbine generator fault diagnosis, and more specifically, to a wind turbine generator fault diagnosis system and method based on nested memory self-supervised learning. Background Technology

[0002] With the continuous development of technology, wind power, as a clean and renewable energy source, is playing an increasingly important role in the global energy structure. Wind turbines are usually located in remote, high-altitude areas with harsh environments, making daily maintenance and repair difficult. Once a failure occurs, it will not only lead to high maintenance and installation costs, but may also cause fluctuations in the energy supply system due to the uncertainty of the failure. Therefore, monitoring and predicting the operating status of wind turbines is an important development direction for predictive maintenance and reducing energy supply system fluctuations caused by maintenance.

[0003] Currently, wind turbine fault diagnosis mainly employs statistical models based on low-frequency variables from SCADA, high-frequency signal feature engineering, and classifiers such as support vector machines and random forests for fault identification, as well as deep learning methods to predict the operating status of wind turbines. However, these methods have the following problems in the prediction process: On the one hand, existing methods usually model on a single time scale, making it difficult to simultaneously capture transient anomalies at the millisecond level or even smaller (such as impacts and crack initiation) and slow degradation at the cyclic and quarterly levels (such as wear, loosening, and aging). Increasing the window can expand the time coverage, but it will lead to a sharp increase in computational load and make engineering deployment difficult. On the other hand, most existing solutions rely on sliding windows or statically fixed model parameters, lacking a controllable short, medium, and long-term memory collaborative update mechanism, resulting in insufficient long-term degradation tracking capabilities.

[0004] Therefore, existing technologies have shortcomings and urgently need improvement. Summary of the Invention

[0005] In view of the above problems, the purpose of this invention is to provide a wind turbine generator diagnostic system and method based on nested memory self-supervised learning, so as to solve the problems in the prior art. On the one hand, existing methods usually model on a single time scale, which makes it difficult to simultaneously capture transient anomalies at the millisecond level or even smaller (such as impact, crack initiation) and slow degradation at the cyclic and quarterly levels (such as wear, loosening, aging). Increasing the window can expand the time coverage, but it will lead to a sharp increase in the amount of computation and make engineering deployment difficult. On the other hand, most existing solutions rely on sliding windows or static fixation of model parameters, lacking a controllable short, medium and long-term memory collaborative update mechanism, resulting in insufficient long-term degradation tracking capability.

[0006] The first aspect of this invention provides a wind turbine generator diagnostic system based on nested memory self-supervised learning, comprising: The data acquisition module is used to acquire high-frequency sensor data and SCADA data; The cross-scale alignment and segmentation module is used to establish a hierarchical time index, segment the high-frequency sensing data according to millisecond-level time windows, obtain data segments containing multiple sampling points, construct multiple time scale layers including short-term, medium-term and long-term layers, and align the time axis of SCADA data and event logs with the time axis of the high-frequency sensing data. A multi-scale representation encoding module is used to extract feature representations from the multiple time-scale layers; The Continuous Spectrum Memory (CMS) module is configured with a short-term volatile memory submodule, a medium-term contextual memory submodule, and a long-term baseline memory submodule. Each submodule is configured with a different update frequency and executes a nested write-consolidation mechanism based on surprise gating. Specifically, the surprise generated by the short-term layer triggers the write to the medium-term contextual memory, and the stable offset of the medium-term layer triggers the consolidation to the long-term baseline memory. The self-supervised training module is used to jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals to obtain the early warning model. The reasoning and alarm module is used to output cross-scale anomaly scores, degradation trend scores, and early warning levels based on the early warning model.

[0007] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the cross-scale alignment and segmentation module is also used to divide the data into different operating condition buckets based on the operating parameters of the wind turbine generator. For a single operating condition bucket, surprise gating is used to calculate surprise based on one or both of the short-term prediction residuals and cross-modal inconsistency.

[0008] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the short-term prediction residual is: the residual of the short-term encoder predicting or masking the next segment based on the current high-frequency segment. The cross-modal inconsistency is defined as the consistency loss between the mechanical state inferred from the high-frequency sensing data and the predicted state based on the SCADA data.

[0009] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the short-term volatile memory submodule is updated at a first frequency and performs circular caching or short context attention caching; the medium-term context memory submodule is updated at a second frequency and performs key-value storage; the long-term baseline memory submodule is updated at a third frequency and constructs a prototype library or parameter adapter. Wherein: the first frequency is greater than the second frequency and the second frequency is greater than the third frequency.

[0010] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the nested write-consolidation mechanism based on surprise gating includes: In response to the existence of a surprise level greater than or equal to the surprise level threshold in a certain working condition bucket, the short-term representation corresponding to the working condition bucket is written into the medium-term scenario memory submodule. In response to the mid-term scenario memory in any working condition bucket shifting within multiple consecutive cycles and the shift rate being less than a preset shift rate, the mid-term scenario memory summary corresponding to that working condition bucket is written into the long-term baseline memory submodule. The long-term baseline memory submodule is configured with a forgetting strategy and stability constraints.

[0011] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the self-supervised training module, based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals, jointly trains the multi-scale representation encoding module and the continuous spectrum memory CMS module to obtain an early warning model, including: The token sequence of high-frequency segments is randomly masked and the masked part is reconstructed, and the short-term mask model is constructed. The cross-scale prediction is as follows: based on the feature representation of the medium-term layer, the statistics or residual trend of the future window are predicted, or the feature representation of the long-term layer is used to predict the future daily distribution or seasonal baseline. The cross-modal consistency residual is: data obtained by comparative learning or mutual prediction of the numerical vector feature representations of encoder outputs of high-frequency sensing data and SCADA data within the same time window; The memory consistency residual is obtained by constraining long-term memory to be sparsely updated only when the stable consolidation condition is met, and by applying distillation constraints to the old and new baselines.

[0012] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the cross-scale anomaly score is a score derived from the prediction residual, reconstruction error, or distance to a prototype with similar operating conditions based on high-frequency segments. The medium-term degradation trend score is: for the same working condition bucket, the score is obtained by calculating its short-term anomaly score and performing robust aggregation and trend estimation; The long-term progress score is calculated based on the distance between the current interim summary and the prototype in long-term baseline memory, as well as the cross-day trend.

[0013] As a preferred technical solution for a wind turbine generator diagnostic system based on nested memory self-supervised learning, the wind turbine generator diagnostic system based on nested memory self-supervised learning is configured as an edge-cloud collaborative deployment architecture, with the data acquisition module, the cross-scale alignment and segmentation module, the short-term encoder, and the short-term anomaly score calculation function deployed on the edge side. Deploy and consolidate the mid-term encoder, long-term encoder, mid-term context memory submodule, and long-term baseline memory submodule in the cloud, and distribute the updated long-term prototype or adapter to the edge.

[0014] This invention also provides a wind turbine generator diagnostic method based on nested memory self-supervised learning, including: Step S1: Acquire high-frequency sensor data and SCADA data; Step S2: Establish a hierarchical time index, divide the high-frequency sensing data into millisecond-level time windows to obtain data segments containing multiple sampling points and construct multiple time scale layers; and align the SCADA data and event logs with the time axis of the high-frequency sensing data. Step S3: Extract the feature representations of the multiple time-scale layers; Step S4: Configure the short-term volatile memory submodule, the medium-term contextual memory submodule, and the long-term baseline memory submodule. Each submodule is configured with a nested write-consolidation mechanism based on surprise gating. Step S5: Jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module to obtain the early warning model; Step S6: Output cross-scale anomaly score, degradation trend score, and early warning level based on the early warning model.

[0015] Compared with the prior art, the beneficial effects of the present invention are that by constructing three time scale layers—short-term (millisecond- or even smaller, because the data acquisition frequency is much smaller than milliseconds or even smaller, so it can react to smaller changes) to second-term, medium-term (minute-day), and long-term (week-year)—the present invention captures weak transient anomalies at the millisecond level or even smaller. By using the nested write-consolidation mechanism of the continuous spectrum memory (CMS) module, short-term anomalies are gradually accumulated into traceable medium-term degradation evidence and long-term discernible progression stages, achieving full-cycle continuous coverage from fault initiation to stable degradation, and avoiding early warning lag caused by the submersion of early weak fault signals.

[0016] Furthermore, this invention trains the target by combining short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals. This allows the model to be trained even when there are few fault labels or only a few alarm records, significantly reducing the cost of manual annotation and improving the applicability and scalability of the solution in real wind field environments.

[0017] Furthermore, this invention classifies and compares data under similar operating conditions through a working condition binning strategy, combines cross-modal consistency residual constraints to cross-verify the representations of high-frequency sensor data and SCADA data, and utilizes the forgetting strategy of long-term baseline memory and stability constraints to accurately distinguish between true degradation trends and normal operating condition drift or seasonal changes, significantly improving alarm accuracy.

[0018] Furthermore, this invention achieves long-range dependency modeling in a layered nested manner without relying on an infinitely long context window. Instead, it uses three layers of memory sub-modules that work together with differentiated update frequencies: "Short-term memory updates quickly to ensure millisecond-level or even smaller response times, medium-term memory updates moderately to track daily offsets, and long-term memory updates slowly to maintain quarterly and grade-level baselines." This layered nested approach results in controllable computational overhead and is suitable for engineering deployment.

[0019] Furthermore, this invention simultaneously outputs short-term anomaly scores, mid-term degradation trend scores, and long-term progress scores through the reasoning and alarm modules, and supports backtracking of high-frequency segments triggered by writing and corresponding operating condition bucket information, providing maintenance personnel with a complete chain of evidence from instantaneous anomaly capture to long-term degradation judgment, thereby improving the credibility and traceability of fault diagnosis conclusions. Attached Figure Description

[0020] Figure 1 The diagram shows the structural block diagram of the wind turbine generator diagnostic system based on nested memory self-supervised learning provided by the present invention; Figure 2 The flowchart illustrates the steps of the wind turbine generator diagnosis method based on nested memory self-supervised learning provided by the present invention. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0023] Figure 1 The diagram shows the structural block diagram of the wind turbine generator diagnostic system based on nested memory self-supervised learning provided by the present invention.

[0024] like Figure 1 As shown, this invention discloses a wind turbine generator diagnostic system based on nested memory self-supervised learning, comprising: The data acquisition module is used to acquire high-frequency sensor data and SCADA data; The cross-scale alignment and segmentation module is used to establish a hierarchical time index, segment high-frequency sensing data according to millisecond-level time windows, obtain data fragments containing multiple sampling points, construct multiple time scale layers including short-term, medium-term and long-term layers, and align the time axis of SCADA data and event logs with the time axis of high-frequency sensing data. The multi-scale representation encoding module is used to extract feature representations from multiple time-scale layers; The Continuous Spectrum Memory (CMS) module is configured with a short-term volatile memory submodule, a medium-term contextual memory submodule, and a long-term baseline memory submodule. Each submodule is configured with a different update frequency and executes a nested write-consolidation mechanism based on surprise gating. Specifically, the surprise generated by the short-term layer triggers the write to the medium-term contextual memory, and the stable offset of the medium-term layer triggers the consolidation to the long-term baseline memory. The self-supervised training module is used to jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals to obtain the early warning model. The inference and alerting module is used to output cross-scale anomaly scores, degradation trend scores, and early warning levels based on the early warning model.

[0025] It should be noted that, in this embodiment of the invention, high-frequency sensing data includes, but is not limited to, vibration signals, acoustic emission signals, acceleration signals, and current ripple signals, with a sampling frequency of 40kHz. SCADA data refers to data collected by the wind turbine monitoring and data acquisition system, including but not limited to power, rotational speed, wind speed, pitch angle, yaw angle, temperature, and alarm information. The cross-scale alignment and segmentation module is used to establish a hierarchical time index. Specifically, the 40kHz high-frequency sensing data is divided into time windows of 20-200ms to obtain data segments containing multiple sampling points. Each segment contains 800-8000 sampling points, and overlap can be set between segments. The feature is represented as a numerical vector obtained by mapping the original high-frequency segments, low-frequency operating conditions, and cross-scale aggregation windows by the encoder. The numerical vector is used to characterize the high-frequency states such as vibration / acoustic emission / current ripple, SCADA operating conditions, and health status shifts within the corresponding time window. Based on this, the multiple time scale layers constructed in the cross-scale alignment and segmentation module are at least divided into three levels: short-term layer (millisecond level or even smaller (high-frequency sensing data is 40kHz, which is much longer than milliseconds, so levels smaller than milliseconds can also be achieved, and the millisecond level or even smaller level mentioned below is the same as this) - second level), medium-term layer (minutes - days), and long-term layer (weeks - years). At the same time, the time axis of SCADA data and event logs is aligned with the time axis of high-frequency sensing data to establish a unified hierarchical index (segment - window - day - quarter). The multi-scale representation encoding module extracts feature representations corresponding to three levels through short-term encoders, medium-term encoders, and long-term encoders. The short-term encoder encodes data fragments at the millisecond level or even smaller. During encoding, methods such as multi-window STFT, wavelet transform, and learnable filter banks can be used to convert these data fragments into token sequences. Alternatively, 1D convolution, state-space models can be used to directly generate tokens, or the Transformer's ViT architecture can be used to process fragments directly as patches. The specific method chosen depends on the actual situation. The medium-term encoder aggregates short-term tokens within a second to minute window into a medium-term representation using statistical convergence or attention mechanisms to reflect the slow shifts in operating conditions and health. The long-term encoder aggregates medium-term representations by day or week to form long-term event fragments (a combination of daily or weekly aggregated medium-term representations and metadata such as maintenance, season, and unit ID). These fragments are then associated with maintenance records, seasonal tags, and unit IDs to form a long-term baseline and individualized memory. In the Continuous Spectrum Memory (CMS) module, the Short-Term Volatile Memory (STM) submodule is used for millisecond or even smaller-scale context storage, maintaining the K most recent high-frequency segment tokens. It can be implemented as a circular cache or a short context attention cache, with an update frequency of each segment. The value of K is selected based on the actual situation, such as the processing power of the computer used to maintain the high-frequency segment tokens.The mid-term scenario memory submodule adopts a key-value storage structure. The key is the mid-term representation and operating condition index (wind speed / rotation speed / power range, etc.), and the value is the corresponding health characterization or prediction residual statistics. The update frequency is every minute or every window.

[0026] The long-term baseline memory submodule employs a slowly updated prototype library or parameter adapter (e.g., a small number of adapters, prototype vectors, and slow weights) to model seasonality, post-maintenance steady-state, and long-term degradation. Updates occur daily, weekly, or event-triggered. The short-term volatile memory submodule, intermediate-term contextual memory submodule, and long-term baseline memory submodule are all configured with different update frequencies and implement a nested write-consolidation mechanism based on surprise gating: whether to write to the intermediate-term contextual memory is determined by the surprise generated in the short-term volatile memory submodule; the intermediate-term contextual memory submodule is consolidated into the long-term baseline memory submodule under low-frequency or triggered conditions, forming a new long-term baseline; simultaneously, a forgetting strategy is implemented to prevent drift contamination of the long-term baseline. The self-supervised training module, with the training objective of combining short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals, jointly trains the multi-scale representation encoding module and the continuous spectrum memory (CMS) module.

[0027] Furthermore, the cross-scale alignment and segmentation module is also used to divide the data into different operating condition buckets based on the operating parameters of the wind turbine generator set; For a single operating condition bucket, surprise gating is used to calculate surprise based on one or both of the short-term prediction residuals and cross-modal inconsistency.

[0028] It should be noted that the operating parameters of wind turbine generators include, but are not limited to, wind speed, engine speed, and power. Based on the operating data, the data is divided into different operating condition bins, and the bin division is based on the formula... The data is divided into buckets, where B(*) is the bucketing function for the operating conditions. For the working condition bucket to which the k-th window belongs, The SCADA / operating condition vector, aligned with the k-th high-frequency window, includes wind speed, engine speed, power, pitch angle, temperature, alarm status, etc. The operating condition bucket can be divided by at least one of wind speed, engine speed, power, pitch angle, ambient temperature, and unit operating mode. Based on the completed operating condition bucket division, the surprise gating calculation is limited to a single operating condition bucket, thus comparing residuals only under similar operating conditions and reducing the influence of factors such as wind conditions. Surprise gating calculates surprise based on one or a combination of short-term prediction residuals and cross-modal inconsistency. Surprise gating for short-term prediction residuals means that when the residual obtained from the short-term encoder's prediction or mask reconstruction of the next high-frequency segment exceeds a threshold or is located at the upper bound of the quantile, a write operation is triggered. Surprise gating for cross-modal inconsistency means that when the consistency loss between the mechanical state inferred from high-frequency sensor data and the state predicted by SCADA / control variables (such as pitch and torque) increases, the write weight is increased.

[0029] This invention provides a formula for calculating the degree of surprise: ; In the formula, For the Sigmoid function, For bias terms, , , These are the weighting parameters for the comprehensive residual, long-term drift measure, and uncertainty / consistency breach measure, respectively. The values ​​of the weighting parameters are assigned by the computer, and the sum of all weighting parameters is 1. The inference phase synthesizes self-supervised residuals, which are obtained by combining window-level prediction residuals, mask reconstruction residuals, and cross-modal consistency residuals. To integrate drift measures representing the relative long-term baseline prototype, As a measure of model uncertainty or cross-modal consistency residual violation, The surprise / novelty score for the k-th window. Let the surprise threshold be the value of the bucket corresponding to the b-th working condition. The quantiles can be determined from the surprise distribution of historical normal operation data within the same operating condition bucket, such as the 95th or 99th quantiles; alternatively, they can be given by operational experience, or updated adaptively using a sliding window quantile. This is an indicator function.

[0030] The calculation formula is: ; In the formula: For multi-step or cross-scale prediction residuals; Reconstruct residuals from high-frequency fragment tokens or local statistics masks; This represents the cross-modal consistency residual between the high-frequency representation and the SCADA conditional representation; , , For residual weights, , , The value can be fixed, determined through historical normal data, or learned through model training. This is a low-frequency condition representation or operating condition representation, obtained from SCADA condition vector encoding. It represents short-term or high-frequency features and is a numerical vector obtained from model encoding, not the original physical quantity; Fusion represents a measure of drift relative to a long-term baseline prototype. The calculation formula is: ; In the formula: Let be the mean of the j-th long-term baseline prototype in the fused representation space. Let be the covariance of the j-th long-term baseline prototype in the fused representation space. The covariance of the long-term baseline prototype in the fused representation space. For fusion representation, by z k s With z k c The result is obtained through fusion mapping function, used for residual, drift, and memory write. The form of the fusion function is selected based on the actual situation to meet the requirements. It can be used for residual, drift, and memory writes; no specific limitations are imposed on the fusion function here.

[0031] Model uncertainty or cross-modal consistency residual violation measure The calculation formula is: One of them, in the formula: The variance of different prediction head outputs under ensemble model or MC-Dropout. This is the normalized value of the cross-modal consistency residual.

[0032] Furthermore, the short-term prediction residual is: the residual of the short-term encoder when predicting or masking the next segment based on the current high-frequency segment; Cross-modal inconsistency is defined as the consistency loss between the mechanical state inferred from high-frequency sensor data and the predicted state based on SCADA data.

[0033] It should be noted that when the short-term prediction residual value is greater than the set threshold or is located at the upper bound of the quantile, it is considered a surprising event. When the consistency loss increases, it is considered cross-modal inconsistency, indicating that anomalies may exist. The set threshold value is determined based on historical data, computer-generated values, or analysis of existing data.

[0034] Furthermore, the short-term volatile memory submodule is updated at a first frequency and performs circular caching or short contextual attention caching; the medium-term contextual memory submodule is updated at a second frequency and performs key-value storage; the long-term baseline memory submodule is updated at a third frequency and builds a prototype library or parameter adapter. Among them: the first frequency is greater than the second frequency and the second frequency is greater than the third frequency.

[0035] The short-term volatile memory submodule is updated based on the following formula: In the formula: This is a state of short-term memory loss. For short-term state update functions, These are short-term update coefficients.

[0036] It should be noted that the values ​​of the first, second, and third frequencies are related to the time span of the short-term, medium-term, and long-term layers, and the specific values ​​are selected according to the actual situation. In this embodiment of the invention, the first frequency is the time length corresponding to each segment. Based on this time length, the most recent K high-frequency segment tokens are periodically cached in a circular manner or with short context attention, thereby maintaining the real-time context at the millisecond to second level. The second frequency is the time length corresponding to minutes or each window, thereby storing the medium-term representation and operating condition index using key-value pairs. The third frequency is the value of days or weeks, used to build a prototype library or parameter adapter. The prototype library stores health benchmark representations under different operating conditions, seasons, and maintenance versions, while the parameter adapter can be a lightweight adapter module used to adapt to baselines under different operating states.

[0037] Furthermore, a nested write-consolidation mechanism based on surprise gating is implemented, including: In response to the existence of a surprise value greater than or equal to the surprise value threshold in a certain condition bucket, the short-term representation of the condition bucket is written into the medium-term context memory submodule. In response to the shift of the intermediate scenario memory within any working condition bucket within multiple consecutive cycles and the shift rate being less than a preset shift rate, the summary of the intermediate scenario memory corresponding to that working condition bucket is written into the long-term baseline memory submodule. The long-term baseline memory submodule is configured with a forgetting strategy and stability constraints.

[0038] It should be noted that the stability constraint includes: writing to the long-term baseline memory submodule is only allowed when the offset stability condition is met for multiple consecutive periods, thereby ensuring that the long-term baseline only records the true degradation trend and is not disturbed by temporary data disturbances. The period unit corresponding to the intermediate-term contextual memory is days, preferably three consecutive days for multiple consecutive periods. The forgetting strategy here does not mean simply deleting fault evidence, but rather that when maintaining the long-term healthy prototype, the long-term baseline memory submodule needs to de-weight, freeze, replace, merge, or delete long-term prototypes that are expired, no longer applicable, contaminated, or have decreased confidence, in order to avoid the long-term baseline being contaminated by historical old states, occasional anomalies, or seasonal / maintenance changes. This can be interpreted as a long-term memory maintenance mechanism. Those skilled in the art will understand that the preset offset rate is the maximum acceptable level of influence of short-term data on the intermediate-term contextual memory. The smaller the offset rate, the less the intermediate-term contextual memory is affected by short-term factors. Therefore, the value of the preset offset rate is assigned based on the actual situation. Preferably, the preset offset rate is 20% of the maximum offset rate.

[0039] Furthermore, the self-supervised training module, based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals jointly training the multi-scale representation encoding module and the continuous spectrum memory CMS module, obtains the early warning model, including: Randomly mask the token sequence of high-frequency segments and reconstruct the masked part, and construct a short-term mask model; Cross-scale prediction is as follows: based on the feature representation of the medium-term layer, predict the statistics or residual trend of the future window, or use the feature representation of the long-term layer to predict the future daily distribution or seasonal baseline. The cross-modal consistency residual is the cross-modal consistency residual obtained by comparing or predicting the numerical vector feature representations of encoder outputs of high-frequency sensing data and SCADA data within the same time window. The memory consistency residual is obtained by constraining long-term memory to be sparsely updated only when the stable consolidation condition is met, and by applying distillation constraints to the old and new baselines.

[0040] It should be noted that residuals are the difference between predicted and true values, and are a type of training data. The specific construction method for the self-supervised training module, based on joint training of short-term mask modeling, cross-scale prediction, cross-modal consistent residuals, and memory-consistent residuals, includes: Short-term masking modeling involves randomly masking the token sequence of high-frequency segments, training the model to reconstruct the masked token sequence, and forcing the model to learn the normal patterns of high-frequency signals.

[0041] Cross-scale prediction includes two directions: predicting the statistics or residual trends of future windows based on the feature representation of the medium-term layer, and predicting the future daily distribution or seasonal baseline using the feature representation of the long-term layer. Thus, through cross-scale prediction, the model has the ability to infer future degradation from the current state.

[0042] Cross-modal consistency residuals, by comparing and learning the numerical vector feature representations of encoder outputs of high-frequency sensing data and SCADA data within the same time window, can make the two modal representations as consistent as possible within the same time window. Mutual prediction of the numerical vector feature representations of encoder outputs of high-frequency sensing data and SCADA data within the same time window can use one modality to predict the features of another modality, thereby enhancing the robustness of the representation and the fault characterization capability.

[0043] Memory consistency residuals, by constraining long-term memory to be sparsely updated only when the stable consolidation condition is met and imposing distillation constraints on the old and new baselines, can keep the old and new representations of data consistent, thus ensuring the stability and continuity of long-term memory.

[0044] Furthermore, the cross-scale anomaly score is a score derived from the prediction residuals, reconstruction errors, or distances to similar working condition prototypes based on high-frequency segments; The medium-term degradation trend score is: for the same working condition bucket, the score is obtained by calculating its short-term anomaly score and performing robust aggregation and trend estimation; The long-term progress score is calculated based on the distance between the current interim summary and the prototype in long-term baseline memory, as well as the cross-day trend.

[0045] It should be noted that the cross-scale anomaly score is used to reflect the instantaneous anomaly level at the millisecond or even smaller level – the second level; the medium-term degradation trend score is used to reflect the continuous degradation signal after the normalization of the operating conditions; and the long-term progress score can examine the trend over a long period of time, thus forming a multi-stage discrimination and early warning level of short-term, medium-term and long-term.

[0046] Furthermore, the wind turbine generator diagnostic system based on nested memory self-supervised learning is configured with an edge-cloud collaborative deployment architecture, deploying a data acquisition module, a cross-scale alignment and segmentation module, a short-term encoder, and a short-term anomaly score calculation function on the edge side. Deploy and consolidate the mid-term encoder, long-term encoder, mid-term context memory submodule, and long-term baseline memory submodule in the cloud, and distribute the updated long-term prototype or adapter to the edge.

[0047] Furthermore, Figure 2 The flowchart illustrates the steps of the wind turbine generator diagnosis method based on nested memory self-supervised learning provided by the present invention, including: Step S1: Acquire high-frequency sensor data and SCADA data; Step S2: Establish a hierarchical time index, divide the high-frequency sensing data into millisecond-level time windows to obtain data segments containing multiple sampling points and construct multiple time scale layers; and align the SCADA data and event logs with the time axis of the high-frequency sensing data. Step S3: Extract feature representations from multiple time-scale layers; Step S4: Configure the short-term volatile memory submodule, the medium-term contextual memory submodule, and the long-term baseline memory submodule. Each submodule is configured with a nested write-consolidation mechanism based on surprise gating. Step S5: Jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module to obtain the early warning model; Step S6: Output cross-scale anomaly score, degradation trend score, and early warning level based on the early warning model.

[0048] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules can be combined, or integrated into another module, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be indirect coupling or communication connection through some interfaces, devices, or units, and can be electrical, mechanical, or other forms.

[0049] In addition, in the various embodiments of the present invention, all functional modules can be integrated into one processing module, or each module can be a separate module, or two or more modules can be integrated into one module; the integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0050] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] Alternatively, if the integrated modules of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A wind turbine generator diagnostic system based on nested memory self-supervised learning, characterized in that, include: The data acquisition module is used to acquire high-frequency sensor data and SCADA data; The cross-scale alignment and segmentation module is used to establish a hierarchical time index, segment the high-frequency sensing data according to millisecond-level time windows, obtain data segments containing multiple sampling points, construct multiple time scale layers including short-term, medium-term and long-term layers, and align the time axis of SCADA data and event logs with the time axis of the high-frequency sensing data. A multi-scale representation encoding module is used to extract feature representations from the multiple time-scale layers; The Continuous Spectrum Memory (CMS) module is configured with a short-term volatile memory submodule, a medium-term contextual memory submodule, and a long-term baseline memory submodule. Each submodule is configured with a different update frequency and executes a nested write-consolidation mechanism based on surprise gating. Specifically, the surprise generated by the short-term layer triggers the write to the medium-term contextual memory, and the stable offset of the medium-term layer triggers the consolidation to the long-term baseline memory. The self-supervised training module is used to jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals to obtain the early warning model. The reasoning and alarm module is used to output cross-scale anomaly scores, degradation trend scores, and early warning levels based on the early warning model. The self-supervised training module, based on short-term mask modeling, cross-scale prediction, cross-modal consistency residuals, and memory consistency residuals, jointly trains the multi-scale representation encoding module and the continuous spectrum memory CMS module to obtain an early warning model, including: The token sequence of high-frequency segments is randomly masked and the masked part is reconstructed, and the short-term mask is constructed. The cross-scale prediction is as follows: based on the feature representation of the medium-term layer, the statistics or residual trend of the future window are predicted, or the feature representation of the long-term layer is used to predict the future daily distribution or seasonal baseline. The cross-modal consistency residual is: the cross-modal consistency residual obtained by comparing or predicting the numerical vector feature representations of encoder outputs of high-frequency sensing data and SCADA data within the same time window. The memory consistency residual is the memory consistency residual loss obtained by constraining long-term memory to be sparsely updated only when the stable consolidation condition is met, and by applying distillation constraints to the old and new baselines.

2. The wind turbine generator diagnostic system based on nested memory self-supervised learning according to claim 1, characterized in that, The cross-scale alignment and segmentation module is also used to divide the data into different operating condition buckets based on the operating parameters of the wind turbine generator set. For a single operating condition bucket, surprise gating is used to calculate surprise based on one or both of the short-term prediction residuals and cross-modal inconsistency.

3. The wind turbine generator diagnostic system based on nested memory self-supervised learning according to claim 2, characterized in that, The short-term prediction residual is the residual obtained by the short-term encoder when predicting or reconstructing the next segment based on the current high-frequency segment. The cross-modal inconsistency is defined as the consistency loss between the mechanical state inferred from the high-frequency sensing data and the predicted state based on the SCADA data.

4. The wind turbine generator diagnostic system based on nested memory self-supervised learning according to claim 3, characterized in that, The short-term volatile memory submodule is updated at a first frequency and is subject to circular caching or short context attention caching. The intermediate-term scenario memory submodule is updated at a second frequency and stored as a key; the long-term baseline memory submodule is updated at a third frequency and used to build a prototype library or parameter adapter. Wherein: the first frequency is greater than the second frequency and the second frequency is greater than the third frequency.

5. The wind turbine generator diagnostic system based on nested memory self-supervised learning according to claim 4, characterized in that, The surprise-gated nested write-consolidation mechanism includes: In response to the existence of a surprise level greater than or equal to the surprise level threshold in a certain working condition bucket, the short-term representation corresponding to the working condition bucket is written into the medium-term scenario memory submodule. In response to the mid-term scenario memory in any working condition bucket shifting within multiple consecutive cycles and the shift rate being less than a preset shift rate, the mid-term scenario memory summary corresponding to that working condition bucket is written into the long-term baseline memory submodule. The long-term baseline memory submodule is configured with a forgetting strategy and stability constraints.

6. The wind turbine generator diagnostic system based on nested memory self-supervised learning according to claim 3, characterized in that, The wind turbine generator diagnostic system based on nested memory self-supervised learning is configured with an edge-cloud collaborative deployment architecture, and the data acquisition module, the cross-scale alignment and segmentation module, the short-term encoder, and the short-term anomaly score calculation function are deployed on the edge side. Deploy the mid-term encoder, long-term encoder, mid-term context memory submodule, and long-term baseline memory submodule in the cloud, and distribute the updated long-term prototype or adapter to the edge.

7. A wind turbine generator diagnostic method based on nested memory self-supervised learning, used to implement the wind turbine generator diagnostic system based on nested memory self-supervised learning as described in any one of claims 1-6, characterized in that, include: Step S1: Acquire high-frequency sensor data and SCADA data; Step S2: Establish a hierarchical time index, divide the high-frequency sensing data into millisecond-level time windows to obtain data segments containing multiple sampling points and construct multiple time scale layers; and align the SCADA data and event logs with the time axis of the high-frequency sensing data. Step S3: Extract the feature representations of the multiple time-scale layers; Step S4: Configure the short-term volatile memory submodule, the medium-term contextual memory submodule, and the long-term baseline memory submodule. Each submodule is configured with a nested write-consolidation mechanism based on surprise gating. Step S5: Jointly train the multi-scale representation encoding module and the continuous spectrum memory CMS module to obtain the early warning model; Step S6: Output cross-scale anomaly score, degradation trend score, and early warning level based on the early warning model.

Images