Distributed health assessment method for new energy equipment based on edge calculation

Abstract

The invention relates to the technical field of new energy operation and maintenance monitoring, in particular to a new energy equipment distributed health assessment method based on edge computing, and the method comprises the steps: collecting edge side equipment operation monitoring signals, and obtaining an equipment state feature set through preprocessing and feature extraction; calculating a predicted value based on an equipment physical model, and comparing the predicted value with a measured value to generate a residual feature; inputting the state feature set and the residual feature into a self-supervised embedder to generate a low-dimensional health embedding vector, and performing quantitative hash processing to form a health fingerprint; and comparing the health fingerprints with a neighborhood graph pre-constructed by a station side to obtain a consistency deviation score, calculating a health index in combination with residual features, and finally realizing the prediction of the residual life of the equipment. According to the method, through multi-dimensional feature fusion and self-supervised learning, the accuracy of equipment health state characterization and the reliability of residual life prediction are improved, and the method is suitable for intelligent operation and maintenance scenes of new energy equipment.

Description

Technical Field

[0001] This invention relates to the field of new energy operation and maintenance monitoring technology, specifically to a distributed health assessment method for new energy equipment based on edge computing. Background Technology

[0002] In the construction and operation of new energy power plants, the operating status of key equipment such as wind turbines, photovoltaic inverters, energy storage batteries, and converters directly affects the safety and economy of the system. During long-term operation, these devices are subject to complex operating conditions, environmental disturbances, and grid fluctuations, leading to risks of efficiency decline, component aging, and latent faults. Therefore, conducting health status assessments of new energy equipment is a crucial step in ensuring the safe operation of power plants and reducing operation and maintenance costs.

[0003] Currently, health assessments of new energy equipment largely rely on centralized data collection and processing. Large amounts of raw data need to be transmitted to the cloud or a central control center for feature extraction and model calculation. While this approach can achieve condition monitoring to some extent, it also suffers from insufficient real-time performance, excessive bandwidth consumption, and limitations in network conditions in remote areas. Furthermore, centralized methods often overlook the interrelationships between devices, resulting in limited effectiveness in identifying early-stage faults. In addition, some methods lack physical mechanism constraints, leading to poor interpretability of the results.

[0004] The main problem with existing technologies is that centralized health assessment methods are difficult to meet the requirements of new energy power plants for low latency, low communication pressure and high reliability. Especially in scenarios with heterogeneous multi-source data, frequent fluctuations in operating conditions and hidden equipment degradation, existing methods are prone to misjudgment or omission, making it difficult to provide effective support for the health management of equipment throughout its entire life cycle. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a distributed health assessment method for new energy equipment based on edge computing, which addresses the shortcomings of the prior art and solves the technical problem that the existing assessment methods are prone to misjudgment or omission.

[0006] The objective of this invention is achieved through the following technical solutions: In a first aspect, the present invention provides a distributed health assessment method for new energy equipment based on edge computing, comprising: Collect operation monitoring signals from edge-side new energy equipment; preprocess and extract features from the operation monitoring signals to obtain a set of equipment status features; Based on a preset physical model of the equipment, the predicted value is calculated from the set of equipment state features and compared with the corresponding measured value to generate residual features; The device state feature set and the residual features are input into a self-supervised embedding unit to generate a low-dimensional health embedding vector that characterizes the health status of the device. The low-dimensional health embedding vector is then quantized and hashed to obtain the health fingerprint of the corresponding new energy device. The health fingerprint is compared with the pre-constructed neighborhood graph on the station side to obtain a consistency deviation score. Based on the consistency deviation score and residual features, the health index of the new energy equipment is obtained. Based on the health index, the remaining lifespan of the new energy equipment is predicted.

[0007] As a further improvement of the present invention, the operating data includes electrical quantity signals, vibration signals, acoustic emission signals and temperature signals sampled at different frequencies; The feature extraction includes time-domain features, frequency-domain features, or time-frequency-domain features; the feature extraction process includes sequentially filtering, denoising, and normalizing the electrical quantity signal, the vibration signal, the acoustic emission signal, and the temperature signal.

[0008] As a further improvement of the present invention, the physical model of the device includes at least an equivalent model of LCL filtering for photovoltaic inverters and a two-mass torsional vibration model for wind turbine generators.

[0009] As a further improvement of the present invention, generating residual characteristics includes: calculating the predicted current using the LCL filter equivalent model, comparing the predicted current with the measured current to obtain current harmonic residual characteristics; estimating the predicted power using the two-mass torsional vibration model, comparing the predicted power with the measured power to obtain power residual characteristics.

[0010] As a further improvement of the present invention, the neighborhood graph construction process on the station side includes: The electrical distance between new energy equipment is calculated based on the rated line parameters and bus impedance at the station, and the operational correlation between new energy equipment is calculated based on historical health fingerprint sequences. The electrical distance and the operational correlation are combined and used as the weights of the connecting edges in the neighborhood graph.

[0011] As a further improvement of the present invention, the health fingerprint is compared with a pre-constructed neighborhood graph at the station to obtain a consistency deviation score, including: The health fingerprint information of the neighboring new energy equipment in the neighborhood graph is aggregated by a graph convolutional network to generate a neighborhood reference vector. The difference between the health fingerprint of the target new energy device and the neighborhood reference vector is calculated and used as the consistency deviation score.

[0012] As a further improvement of the present invention, a health index for new energy equipment is obtained based on the consistency deviation score and residual characteristics, including: An anomaly measure is calculated based on the healthy fingerprint or the intermediate vector used to generate the healthy fingerprint; The residual features, the consistency deviation score, and the anomaly measure are respectively mapped to healthy evidence, abnormal evidence, and uncertain evidence; The health belief value is obtained by fusing the health evidence, anomalous evidence, and uncertain evidence using the Dempster-Shafer evidence theory rules. The health index is calculated using a weighted function based on the health belief value, the anomaly measure, the consistency deviation score, and the residual characteristics.

[0013] As a further improvement of the present invention, after obtaining the health index, the method further includes uploading the health index to the cloud, and the cloud performs degradation prediction on the time series of the new energy equipment based on the health index and a preset health prototype library to obtain the remaining lifespan of the new energy equipment.

[0014] As a further improvement of the present invention, after obtaining the health index, the method further includes uploading the health fingerprint to the cloud according to a set threshold and a health index-triggered event-driven reporting mechanism. After the health fingerprint is uploaded to the cloud, it is used to update the preset health prototype library. The event-driven reporting mechanism is triggered when the health index continuously falls below a first threshold, or when the decrease in the health index within a unit of time exceeds a second threshold.

[0015] As a further improvement of the present invention, after the health fingerprint is uploaded to the cloud, a preset health prototype library is updated using a federated distillation method, including: The health fingerprints uploaded from the edge side are weighted and aggregated; during weighted aggregation, dynamic weights based on temporal decay are assigned to the health fingerprints reported at different times; The baseline information in the health prototype library is updated using the weighted aggregated knowledge.

[0016] The beneficial effects of this invention are as follows: This embodiment provides a distributed health assessment method for new energy equipment based on edge computing. It collects operational monitoring signals from edge-side new energy equipment, preprocesses and extracts features to form an equipment state feature set, achieving a structured representation of the equipment's raw data to improve the accuracy of subsequent analysis. Based on a preset equipment physical model, the method calculates predicted values ​​from the equipment state feature set and compares them with corresponding measured values ​​to generate residual features. The deviation between model predictions and measurements is used to quantify the degree of equipment operational anomalies. The equipment state feature set and residual features are input into a self-supervised embedding unit to generate a low-dimensional health embedding vector, which is then quantified and hashed to obtain a health fingerprint. This fingerprint is then used for dimensionality reduction and hashing. This technology provides an efficient compressed representation of equipment health status. It compares health fingerprints with a pre-constructed neighborhood graph at the station to obtain a consistency deviation score, and combines this with residual features to generate a health index. Multi-dimensional feature fusion enhances the reliability of equipment health assessment. Finally, based on the health index, it predicts the remaining lifespan of new energy equipment, achieving closed-loop management from equipment status monitoring to lifespan prediction. The synergistic effect of various technical features significantly reduces data processing complexity while maintaining the integrity of feature information. Compared with existing technologies, it boasts superior technical effects such as high health assessment accuracy, small lifespan prediction error, and excellent data storage efficiency. Furthermore, the dynamic verification mechanism of comparing health fingerprints with the neighborhood graph enables continuous optimization and monitoring of equipment health status. Attached Figure Description

[0017] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a distributed health assessment method for new energy equipment based on edge computing, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1 This embodiment provides a distributed health assessment method for new energy equipment based on edge computing. The method includes the following steps: collecting operation monitoring signals of new energy equipment on the edge side; preprocessing and extracting features from the operation monitoring signals to obtain a set of equipment status features; calculating predicted values ​​from the set of equipment status features based on a preset equipment physical model, and comparing them with corresponding measured values ​​to generate residual features; inputting the set of equipment status features and residual features into a self-supervised embedding unit to generate a low-dimensional health embedding vector to characterize the health status of the equipment; quantizing and hashing the low-dimensional health embedding vector to obtain the health fingerprint of the corresponding new energy equipment; comparing the health fingerprint with a pre-constructed neighborhood graph on the station side to obtain a consistency deviation score; obtaining a health index of the new energy equipment based on the consistency deviation score and residual features; and predicting the remaining lifespan of the new energy equipment based on the health index.

[0022] This embodiment collects operational monitoring signals from edge-side new energy equipment and preprocesses and extracts features to form an equipment status feature set, thereby achieving a structured representation of the raw equipment data and improving the accuracy of subsequent analysis. Based on a preset equipment physical model, the equipment status feature set is calculated to obtain predicted values, which are then compared with corresponding measured values ​​to generate residual features. The deviation between model prediction and actual measurement is used to quantify the degree of equipment operational anomalies. The equipment status feature set and residual features are input into a self-supervised embedding unit to generate a low-dimensional health embedding vector, which is then quantized and hashed to obtain a health fingerprint. Dimensional reduction hashing achieves efficient compressed representation of equipment health status. The health fingerprint is compared with a pre-constructed neighborhood graph at the station to obtain a consistency deviation score, which is combined with residual features to generate a health index. Multi-dimensional feature fusion improves the reliability of equipment health assessment. Finally, the remaining lifespan of the new energy equipment is predicted based on the health index, achieving closed-loop management from equipment status monitoring to lifespan prediction. The synergistic effect of various technical features significantly reduces data processing complexity while maintaining the integrity of feature information.

[0023] It should be noted that the operational data includes electrical quantity signals, vibration signals, acoustic emission signals, and temperature signals sampled at different frequencies. Feature extraction includes time-domain features, frequency-domain features, or time-frequency-domain features; the feature extraction process includes sequentially filtering, denoising, and normalizing the electrical quantity signals, vibration signals, acoustic emission signals, and the temperature signals. The equipment physical model includes an LCL filter equivalent model for photovoltaic inverters and a two-mass torsional vibration model for wind turbine generators.

[0024] Specifically, the predicted current is calculated using an equivalent model of LCL filtering, and the characteristics of the current harmonic residual are obtained by comparing the predicted current with the measured current; the predicted power is estimated using a two-mass torsional vibration model, and the characteristics of the power residual are obtained by comparing the predicted power with the measured power.

[0025] The process of constructing the neighborhood graph on the station side includes: calculating the electrical distance between new energy equipment based on the rated line parameters and bus impedance on the station side, and calculating the operational correlation between new energy equipment based on historical health fingerprint sequences; and integrating the electrical distance and the operational correlation as the weight of the connecting edges in the neighborhood graph.

[0026] The health fingerprint is used to compare with the pre-constructed neighborhood graph at the station to obtain a consistency deviation score, which includes: aggregating the health fingerprint information of the neighboring new energy equipment in the neighborhood graph through a graph convolutional network to generate a neighborhood reference vector; and calculating the difference between the health fingerprint of the target new energy equipment and the neighborhood reference vector as the consistency deviation score.

[0027] The health index of new energy equipment is obtained based on the consistency deviation score and residual features, and further includes: calculating anomaly measure based on health fingerprint or intermediate vector of health fingerprint generation; mapping residual features, consistency deviation score and anomaly measure to health evidence, anomaly evidence and uncertain evidence respectively; fusing health evidence, anomaly evidence and uncertain evidence through Dempster-Shafer evidence theory rules to obtain health belief value; and calculating health index based on health belief value, anomaly measure, consistency deviation score and residual features through weighted function.

[0028] After obtaining the health index, the process also includes uploading the health index to the cloud. The cloud then uses the health index and a preset health prototype library to predict the degradation of the new energy equipment over time to obtain the remaining lifespan of the new energy equipment.

[0029] It also includes an event-driven reporting mechanism that triggers an event based on a set threshold and health index to upload health fingerprints to the cloud. Once uploaded to the cloud, the health fingerprints are used to update a preset health prototype library. The event-driven reporting mechanism is triggered when the health index falls below the first threshold for an extended period or when the health index decreases by more than the second threshold within a unit of time.

[0030] Example 2 like Figure 1 As shown in the embodiment of the present invention, a distributed health assessment method for new energy equipment based on edge computing is proposed. The overall process includes multi-stage processing and assessment.

[0031] First, electrical quantities, temperature, vibration, and acoustic emission signals of the new energy equipment are collected at the equipment side. The collected data are then filtered, normalized, and feature extracted to remove noise and generate effective characterization features. Subsequently, the predicted output value is calculated based on a simplified physical model and compared with the measured data to obtain residuals. The residual features, together with the extracted time-frequency features, constitute the physical constraint feature input.

[0032] Based on this, a low-dimensional embedding vector is generated using a self-supervised embedder, and then quantized and hashed to obtain a health fingerprint and its corresponding verification digest. At the station level, a neighborhood graph structure is established based on electrical distance and historical correlation to compare the consistency of health fingerprints from different devices and obtain deviation scores.

[0033] By combining residual characteristics, anomaly metrics, and consistency bias, a health index for new energy equipment is obtained through fusion calculation using evidence theory. Local monitoring is maintained at the edge when the health index is within the normal range; event-driven reporting is triggered when the health index undergoes a sudden change or falls below a threshold, and the compressed health fingerprint and statistics are uploaded to the cloud.

[0034] Federated distillation is performed in the cloud to update the health prototype library and send back new threshold parameters to achieve dynamic calibration at the edge. Furthermore, time series models of the health index are built, and the degradation process is used to predict the remaining lifespan of the equipment. Finally, health status and lifespan assessment results are output, providing a basis for fault warning and operation and maintenance decisions for new energy equipment.

[0035] In summary, the method of this embodiment forms a complete distributed health assessment process through device-side data acquisition and processing, physical model constraints, embedding and fingerprint generation, neighborhood consistency analysis, evidence fusion calculation, event-driven reporting, and cloud-based federated distillation and lifespan prediction. To enable those skilled in the art to more clearly understand the specific technical implementation of this method, the following will provide a detailed description of its specific implementation methods in conjunction with each major step.

[0036] In this embodiment of the invention, the operating data of the new energy equipment is first collected at the equipment side. The collected data includes: electrical quantities (voltage, current), ambient and component temperatures, equipment vibration signals, and acoustic emission signals.

[0037] Electrical quantity sampling adopts a three-phase synchronous sampling method with a sampling frequency of 10 kHz to capture the harmonic components and short-time fluctuation characteristics of current and voltage waveforms. The vibration signal sampling frequency is set to 12.8 kHz, and accelerometers are deployed at key components of the inverter or wind turbine to reflect mechanical shock and early signs of failure. The acoustic emission signal is sampled at a frequency of 100 kHz, and transient characteristics generated by discharge, friction, or crack propagation are recorded by a high-sensitivity acoustic sensor. The temperature sampling frequency is 1 Hz, which is used to record the temperature rise of key nodes during equipment operation.

[0038] The edge computing unit performs preprocessing operations on the collected data, including: 1. Filtering and denoising: Band-stop filtering is applied to electrical signals to suppress power frequency harmonic spikes; 10 Hz to 1 kHz bandpass filtering is applied to vibration signals, combined with envelope demodulation to extract impact characteristics; threshold triggering and event duration detection are used for acoustic emission signals to eliminate environmental noise interference.

[0039] 2. Normalization processing: Based on the nameplate parameters and historical statistical results, the voltage, current, temperature and other signals are normalized to make the data comparable under different load, wind speed or irradiance conditions.

[0040] 3. Feature Extraction: Extract total harmonic distortion (THD), power factor, and short-time energy distribution from electrical signals; extract peak factor, kurtosis, and other impact indicators from vibration signals; extract event counts and energy distribution from acoustic emission signals; and extract moving average and rate of rise indicators from temperature data.

[0041] Through the above steps, a multi-dimensional feature set with time labels is formed, which serves as the input for subsequent physical residual modeling and fingerprint generation.

[0042] After completing data acquisition and preprocessing, this embodiment of the invention further performs calculations based on a simplified physical model to construct residual characteristics that reflect equipment operating deviations.

[0043] For photovoltaic inverters, an equivalent LCL filter circuit model is used to calculate the predicted output current based on the input voltage, current, and modulation parameters. and compared with the actual measured current By comparison, the current harmonic residuals are obtained:

[0044] The mean, standard deviation, and rate of change of the residual are used as characteristics to characterize deviations in electrical performance.

[0045] For wind turbine generator sets, a two-mass torsional vibration model is used to estimate the electromagnetic torque and shaft torque in the drivetrain. Based on this, and combined with wind speed, pitch angle, and rotational speed information, the predicted mechanical output power is calculated. and the actual output power By comparison, the power residual is obtained:

[0046] The magnitude and trend of power residuals reflect the health status of the transmission chain and control components.

[0047] In this process, the residuals are not only used alone as anomaly features, but also further processed by first-order difference operations to construct dynamic residual features, which enhance the sensitivity to short-term abrupt changes. Finally, the residual features and the time-frequency features obtained from the aforementioned preprocessing together constitute the input feature set, providing physically constrained input for subsequent health fingerprint generation.

[0048] After obtaining the feature set and residual features, this embodiment of the invention generates a health fingerprint by deploying a lightweight self-supervised embedder on the device side.

[0049] First, an embedding structure composed of a convolutional neural network and a temporal convolutional network (TCN) is adopted: The convolutional part is used to extract local temporal patterns. It uses two one-dimensional convolutional layers with a kernel length of 5 and 32 and 64 channels, respectively. The temporal convolutional network is used to capture long-term dependencies. It sets up two dilated convolutional layers with dilation rates of 2 and 4, and uses residual connections to ensure the continuity of features at different scales. The activation function used is LeakyReLU to enhance feature separation capabilities in scenarios with low magnitude residuals.

[0050] After embedding operations, a 16-dimensional low-dimensional embedding vector is obtained. This vector represents the health status of the device within the current time window.

[0051] To improve communication efficiency and security, this embodiment of the invention further quantizes the embedding vector: Real-valued embedding vectors are mapped to a high-dimensional space using random projection. The mapping result is symbolized to obtain a 256-bit binary vector, which serves as the health fingerprint. The health fingerprint is combined with a timestamp and feature digest information, and a digest signature is generated using a cryptographic hash function (such as BLAKE3) for integrity verification in subsequent transmissions.

[0052] The generated health fingerprint is stored locally on the device and selectively uploaded in an event-driven manner based on the health index assessment results, thereby preserving key health status information while ensuring privacy and compressing communication load.

[0053] After generating a health fingerprint at the edge, this embodiment of the invention uses a station-side coordination unit to perform neighborhood consistency comparisons on multiple devices in the same combiner box, the same busbar, or the same area to identify potential single-point anomalies.

[0054] First, a neighborhood graph is established based on the station's internal topology and operational data: Electrical distance calculation: Based on line parameters and bus impedance, the equivalent electrical distance between different devices is calculated. Devices with closer electrical distances have higher weights in the neighborhood graph. Correlation assessment: Pearson correlation coefficient was calculated using historical 24-hour health fingerprint sequences as a supplementary correlation weight; Comprehensive weight determination: The electrical distance weight and the correlation weight are weighted and merged to form the edge weight of the neighborhood graph. The neighborhood relationship is updated every 15 minutes and reconstructed immediately when an emergency occurs.

[0055] After the neighborhood graph is constructed, a lightweight graph convolutional network is used to aggregate the health fingerprints of neighboring devices to obtain the neighborhood reference embedding vector of the target device. Then, the difference between the target device embedding vector and the neighborhood reference embedding vector is calculated and defined as the consistency deviation score.

[0056]

[0057] in, Embed vectors for the target device. For neighborhood normalized weights, The neighborhood set of the target device.

[0058] Furthermore, this embodiment introduces an operating condition gating mechanism to avoid misjudgments caused by differences in environmental or operating conditions. For example, consistency comparison is only enabled when adjacent wind turbines are in the same wind speed range or adjacent inverters are in similar irradiance levels; if the difference in operating conditions exceeds a preset threshold, the neighborhood comparison calculation is suspended.

[0059] By comparing the consistency of the surrounding area as described above, it is possible to effectively identify situations where the health status of a certain device deviates significantly from that of neighboring devices, thereby providing support for the early detection of device anomalies.

[0060] After completing the neighborhood consistency comparison, this embodiment of the invention performs fusion calculation on multi-source indicators to form a unified health index.

[0061] First, define the local anomaly degree: calculate the Mahalanobis distance based on the healthy embedding vector and its historical mean and covariance to measure the degree of deviation between the current state and the historical normal state; use the aforementioned consistency deviation score to reflect the difference between the target device and neighboring devices; combine the mean and rate of change of the physical residual to describe the deviation between the device operation and the theoretical model.

[0062] Subsequently, the above indicators were used as sources of evidence and fused using evidence theory (Dempster-Shafer, DS): Each indicator is first mapped to a basic probability allocation of three categories: healthy (H), abnormal (F), and uncertain (U). The mapping process is implemented in segments based on a set threshold. During the fusion process, the confidence of health evidence increases as residuals decrease and consistency improves, while the confidence of anomalous evidence increases as Mahalanobis distance and bias score increase. For conflicting evidence, the DS rule uses normalization to assign the conflict, thereby obtaining a comprehensive health credibility and anomaly credibility.

[0063] Finally, a Health Index (HI) is constructed based on the fusion results:

[0064] in, For the Sigmoid function, For health belief value, For local anomaly degree, For consistency deviation, For the average residual, The weight parameters are initially set to 0.6, 0.2, 0.15, and 0.05, and are dynamically updated based on the results transmitted back from the cloud.

[0065] The health index ranges from 0 to 1, with higher values ​​indicating a higher level of health. When the health index falls below a preset threshold, the system considers the device to have a potential anomaly and initiates an event-driven reporting process; when the health index is within the normal range, local monitoring continues.

[0066] When the health index calculated at the edge side shows an abnormality, this embodiment of the invention triggers a report through an event-driven mechanism, thereby achieving global collaborative calibration while ensuring communication efficiency.

[0067] The trigger condition is: when the health index is below a threshold for a consecutive period of time, for example, below 0.6; If the health index changes abruptly within a short period of time, and the magnitude of the change exceeds the preset value, such as a decrease of more than 0.1; if any of the above conditions are met, the edge unit will trigger an event report.

[0068] To reduce bandwidth consumption, the reported content is a highly compressed feature summary, including: a binarized health fingerprint vector; key statistics (such as residual mean and anomaly measure); and a verification summary signature for verifying data integrity.

[0069] All data is encrypted and then transmitted to the cloud via a lightweight communication protocol, with a single report payload not exceeding 512 bytes.

[0070] After receiving the reported data, the cloud service unit performs the following processing: The new fingerprint is compared with the baseline vector in the health prototype library to identify its similarity to the normal state or failure mode. By using federated distillation, knowledge is integrated from features uploaded from the edge devices to update the health prototype library and avoid model deviation caused by a single device malfunction. Robust calibration of the model is performed during the distillation process, and a differential privacy mechanism is introduced to prevent feature leakage from a single device.

[0071] After completing the global distillation update, the cloud sends the corrected threshold parameters and statistical compensation factors back to each edge unit. The edge units then use this information to calibrate subsequent health index assessments, thereby ensuring the dynamic consistency and accuracy of the overall system.

[0072] Through the aforementioned event-driven reporting and cloud distillation mechanism, this invention achieves dynamic updates of the global model and threshold optimization while ensuring autonomous real-time evaluation at the edge, thereby improving the accuracy and long-term stability of health assessment.

[0073] After completing the calculation and cloud calibration of the health index, this embodiment of the invention further models the time series of the health index to predict the remaining lifespan of the equipment.

[0074] The edge units perform sliding storage of health indices, forming time-series trajectories with minute-level or hourly granularity. The site-side servers archive the health index trajectories of multiple devices, supporting trend analysis across days and weeks. By comparing short-term fluctuations with long-term degradation trends, potential performance degradation patterns can be preliminarily identified.

[0075] A Gamma-based degradation process model was used to fit the trajectory of the health index. It was assumed that the health index exhibited a monotonically decreasing trend over time, and its degradation rate followed a Gamma distribution. The degradation rate parameters were updated using a Bayesian method. The prior distribution is given by historical degradation data of similar equipment; When new observation data arrives, the edge unit or station-side server performs a posteriori correction on the degradation rate parameter to ensure that the model is consistent with actual operation.

[0076] When the health index approaches the warning threshold, the degradation model predicts the time distribution required to reach the threshold, thus obtaining a confidence interval for the remaining useful life (RUL). For example, when the health index drops to 0.45 and the degradation rate stabilizes, the model may give a 95% probability that the equipment will reach the critical threshold within the next 10 to 14 days. This prediction provides a quantitative reference for maintenance planning and spare parts preparation.

[0077] While predicting remaining lifespan, this invention also generates root cause suggestions based on the similarity between the healthy fingerprint and the faulty prototype. For example, if the fingerprint is closest to the "cooling fan degradation" prototype, a "cooling subsystem risk" label is attached to the lifespan prediction result so that maintenance personnel can make targeted maintenance plans.

[0078] Through the above-described health trajectory modeling and life prediction process, this invention not only achieves real-time assessment of equipment health status, but also provides future evolution trends and maintenance suggestions, effectively supporting proactive maintenance and life management of new energy equipment.

[0079] In summary, the edge computing-based distributed health assessment method for new energy equipment proposed in this invention achieves a collaborative health assessment system across edge, station, and cloud environments. This system performs data acquisition and preprocessing, physical model calculation and residual construction, and generates health fingerprints using a self-supervised embedding mechanism at the device level. At the station level, it conducts neighborhood consistency comparison and operational condition gating. Evidence theory is introduced at the fusion layer to calculate the health index. Furthermore, event-driven reporting and cloud-based federated distillation are combined to realize this collaborative system. Further, this method utilizes the time series of the health index for degradation modeling and remaining lifetime prediction, which not only improves the real-time performance and accuracy of the health assessment but also enhances the physical interpretability and operational guidance value of the results. Compared with existing health management methods that rely on centralized processing, this invention provides effective support for the reliable operation and intelligent maintenance of new energy equipment while reducing communication overhead, improving anomaly identification accuracy, and protecting data privacy.

Claims

1. A distributed health assessment method for new energy equipment based on edge computing, characterized in that, include: Collect operational monitoring signals from edge-side new energy equipment; The operation monitoring signals are preprocessed and feature extracted to obtain a set of equipment status features; Based on a preset physical model of the equipment, the predicted value is calculated from the set of equipment state features and compared with the corresponding measured value to generate residual features; The device state feature set and the residual features are input into a self-supervised embedding unit to generate a low-dimensional health embedding vector that characterizes the health status of the device. The low-dimensional health embedding vector is then quantized and hashed to obtain the health fingerprint of the corresponding new energy device. The health fingerprint is compared with the pre-constructed neighborhood graph on the station side to obtain a consistency deviation score. Based on the consistency deviation score and residual features, the health index of the new energy equipment is obtained. Based on the health index, the remaining lifespan of the new energy equipment is predicted.

2. The distributed health assessment method for new energy equipment based on edge computing according to claim 1, characterized in that, The operational data includes electrical signals, vibration signals, acoustic emission signals, and temperature signals sampled at different frequencies; The feature extraction includes time-domain features, frequency-domain features, or time-frequency-domain features; The feature extraction process includes sequentially filtering, denoising, and normalizing the electrical quantity signal, the vibration signal, the acoustic emission signal, and the temperature signal.

3. The distributed health assessment method for new energy equipment based on edge computing according to claim 1, characterized in that, The physical model of the equipment includes an equivalent model of LCL filtering for photovoltaic inverters and a two-mass torsional vibration model for wind turbine generators.

4. The distributed health assessment method for new energy equipment based on edge computing according to claim 3, characterized in that, The process of generating residual features includes: calculating the predicted current using the LCL filter equivalent model, comparing the predicted current with the measured current to obtain the current harmonic residual features; and estimating the predicted power using the two-mass torsional vibration model, comparing the predicted power with the measured power to obtain the power residual features.

5. The distributed health assessment method for new energy equipment based on edge computing according to claim 1, characterized in that, The process of constructing the neighborhood graph on the station side includes: The electrical distance between new energy equipment is calculated based on the rated line parameters and bus impedance at the station, and the operational correlation between new energy equipment is calculated based on historical health fingerprint sequences. The electrical distance and the operational correlation are combined and used as the weights of the connecting edges in the neighborhood graph.

6. The distributed health assessment method for new energy equipment based on edge computing according to claim 5, characterized in that, The health fingerprint is compared with a pre-constructed neighborhood graph at the station to obtain a consistency deviation score, including: The health fingerprint information of the neighboring new energy equipment in the neighborhood graph is aggregated by a graph convolutional network to generate a neighborhood reference vector. The difference between the health fingerprint of the target new energy device and the neighborhood reference vector is calculated and used as the consistency deviation score.

7. The distributed health assessment method for new energy equipment based on edge computing according to claim 1, characterized in that, The health index of new energy equipment is obtained based on the consistency deviation score and residual characteristics, including: An anomaly measure is calculated based on the healthy fingerprint or the intermediate vector used to generate the healthy fingerprint; The residual features, the consistency deviation score, and the anomaly measure are respectively mapped to healthy evidence, abnormal evidence, and uncertain evidence; The health belief value is obtained by fusing the health evidence, anomalous evidence, and uncertain evidence using the Dempster-Shafer evidence theory rules. The health index is calculated using a weighted function based on the health belief value, the anomaly measure, the consistency deviation score, and the residual characteristics.

8. The distributed health assessment method for new energy equipment based on edge computing according to claim 7, characterized in that, After obtaining the health index, the method further includes uploading the health index to the cloud, whereby the cloud performs degradation prediction on the time series of the new energy equipment based on the health index and a preset health prototype library, thereby obtaining the remaining lifespan of the new energy equipment.

9. The distributed health assessment method for new energy equipment based on edge computing according to claim 8, characterized in that, After obtaining the health index, the process also includes uploading the health fingerprint to the cloud based on a set threshold and a health index-triggered event-driven reporting mechanism. The health fingerprint is then used to update the preset health prototype library. The event-driven reporting mechanism is triggered when the health index continuously falls below a first threshold, or when the decrease in the health index within a unit of time exceeds a second threshold.

10. The distributed health assessment method for new energy equipment based on edge computing according to claim 9, characterized in that, After the health fingerprint is uploaded to the cloud, the preset health prototype library is updated using a federated distillation method, including: The health fingerprints uploaded from the edge side are weighted and aggregated; during weighted aggregation, dynamic weights based on temporal decay are assigned to the health fingerprints reported at different times; The baseline information in the health prototype library is updated using the weighted aggregated knowledge.

Images