Distributed energy storage equipment predictive maintenance system and method based on AI

Abstract

The embodiment of the invention relates to the technical field of computers, in particular to an AI-based distributed energy storage equipment predictive maintenance system and method, and the method achieves the precise monitoring of the operation state of distributed energy storage equipment through the integration of the comprehensive collection of multi-dimensional operation data and the deep analysis of an AI prediction model. In detail, the method can effectively identify the potential abnormal characteristics of the equipment, generates a targeted predictive maintenance instruction in combination with the quantitative evaluation of the maintenance urgency index, remarkably improves the timeliness and accuracy of the equipment maintenance, reduces the risk of unplanned shutdown, prolongs the service life of the equipment, and optimizes the maintenance resource distribution efficiency at the same time.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more specifically, to an AI-based predictive maintenance system and method for distributed energy storage equipment. Background Art

[0002] In the existing field of distributed energy storage equipment condition monitoring technology, traditional methods mainly rely on single-dimensional data analysis and simple threshold-based alarm mechanisms. For example, discrete temperature monitoring points are often used in conjunction with voltage over-limit detection, and alarms are triggered by setting fixed thresholds. This method is difficult to capture weak correlation characteristics during equipment degradation.

[0003] In addition, the input feature dimension of the traditional fault classification method based on support vector machines is limited, and it is impossible to effectively integrate the temperature gradient and high-frequency current waveform characteristics of the temporal and spatial distribution, resulting in insufficient sensitivity in early subtle anomaly detection. In addition, traditional maintenance decision systems mostly use static maintenance strategy tables, such as planned maintenance based on the accumulation of equipment operating time. This time-driven model is difficult to adapt to the dynamic degradation process under complex working conditions.

[0004] It can be seen that the above-mentioned technical defects seriously restrict the timeliness and pertinence of predictive maintenance of distributed energy storage systems. Summary of the invention

[0005] In order to at least overcome the above-mentioned deficiencies in the prior art, one of the objectives of the present application is to provide an AI-based distributed energy storage equipment predictive maintenance system and method.

[0006] The embodiment of the present application provides an AI-based distributed energy storage device predictive maintenance method, which is applied to a distributed energy storage device predictive maintenance system. The method includes: Collecting multi-dimensional operating data of distributed energy storage devices, wherein the multi-dimensional operating data includes device temperature series, charge and discharge current waveforms, and voltage fluctuation characteristics; Input the multi-dimensional operation data into a pre-trained AI prediction model to generate an operation status prediction result of the distributed energy storage device, wherein the operation status prediction result includes a device health score and a set of potential abnormal features; Based on the operation status prediction result, the correlation between the potential abnormal feature set and the preset fault mode library is analyzed to determine the maintenance urgency index of the distributed energy storage device; According to the comparison result between the maintenance urgency index and the preset maintenance threshold, a predictive maintenance instruction for the distributed energy storage device is generated, and the predictive maintenance instruction includes a maintenance time window and a maintenance operation type.

[0007] By applying the embodiments of this application, the method realizes accurate monitoring of the operating status of distributed energy storage equipment by integrating comprehensive collection of multi-dimensional operating data with in-depth analysis of AI prediction models. This method can effectively identify potential abnormal characteristics of equipment, combine quantitative evaluation of maintenance urgency indicators, generate targeted predictive maintenance instructions, significantly improve the timeliness and accuracy of equipment maintenance, reduce the risk of unplanned downtime, extend the service life of equipment, and optimize the efficiency of maintenance resource allocation.

[0008] As an optional embodiment, the multi-dimensional operation data further includes an environmental parameter set, and the environmental parameter set includes a humidity change curve, an environmental temperature gradient, and a vibration frequency spectrum at the device deployment location; the multi-dimensional operation data of the distributed energy storage device is collected, including: The device temperature sequence, the charge and discharge current waveform and the voltage fluctuation characteristics are acquired in real time by a sensor array deployed on the distributed energy storage device; Synchronously collecting the humidity change curve, the ambient temperature gradient and the vibration frequency spectrum in the environment where the distributed energy storage device is located, and fusing them with the device operation data in a timestamp alignment manner to obtain initial fused data; wherein the device operation data at least includes the device temperature sequence, the charge and discharge current waveform and the voltage fluctuation characteristics acquired in real time; The initial fusion data is subjected to noise filtering and data normalization processing to generate target multi-dimensional operation data.

[0009] By applying the embodiment of the present application, the method generates a multi-dimensional feature collaborative analysis system by integrating the spatial and temporal fusion of the environmental parameter set and the equipment operation data. In this way, it is possible to effectively capture the impact of environmental factors on equipment operation, enhance the context perception capability of anomaly detection, eliminate sensor noise interference through data standardization, improve the input data quality of subsequent model analysis, and provide a data basis for equipment status evaluation in complex environments.

[0010] As an optional embodiment, the pre-trained AI prediction model generates the running status prediction result in the following manner: Dividing the target multi-dimensional operation data into data segments within a continuous time window, and extracting the temporal dependency and cross-dimensional correlation features in each data segment; By using the deep convolutional network layer in the AI ​​prediction model, the temporal dependency is locally enhanced to generate an enhanced spatiotemporal feature vector; Using the attention mechanism layer in the AI ​​prediction model, weights are assigned to the cross-dimensional correlation features to screen out a subset of key features related to device health; Based on the comparative analysis of the key feature subset and the historical fault data, the equipment health score and the potential abnormal feature set are output.

[0011] By applying the embodiments of the present application, the method adopts a hybrid model architecture that combines a deep convolutional network with an attention mechanism to achieve collaborative feature mining of multi-source heterogeneous data. Through spatiotemporal feature enhancement and key feature screening, it effectively extracts weakly correlated features in the equipment degradation process, improves the physical interpretability of health score calculation, provides a multi-dimensional decision-making basis for equipment status prediction, and significantly improves the ability to identify early faults.

[0012] As an optional embodiment, the analysis of the correlation between the potential abnormal feature set and the preset fault mode library to determine the maintenance urgency index of the distributed energy storage device includes: Extracting abnormal waveform patterns, temperature mutation intervals and current distortion features from the potential abnormal feature set; Performing morphological matching between the abnormal waveform pattern and the known fault waveforms in the preset fault pattern library, and calculating a waveform similarity index; Determining a temperature-associated risk value according to an overlap ratio between the temperature mutation interval and a historical fault temperature threshold; Combining the distribution density of the current distortion characteristics in the time dimension, generating a comprehensive anomaly confidence level; The maintenance urgency index is determined based on a weighted fusion result of the waveform similarity index, the temperature-associated risk value and the comprehensive anomaly confidence.

[0013] By applying the embodiments of the present application, the method establishes a multi-dimensional abnormal feature fusion evaluation system, and realizes the accurate quantification of maintenance urgency through the triple verification mechanism of waveform morphology matching, temperature risk quantification and current distortion analysis. The method breaks through the limitations of traditional single threshold judgment, generates a dynamic weighted evaluation model, effectively balances the probability of failure and the severity of consequences, provides a scientific basis for maintenance priority decision-making, and avoids excessive or insufficient maintenance.

[0014] As an optional embodiment, generating a predictive maintenance instruction for the distributed energy storage device according to a comparison result between the maintenance urgency index and a preset maintenance threshold includes: According to the numerical range of the maintenance urgency indicator, a corresponding maintenance operation type is matched from a preset maintenance strategy table, wherein the maintenance operation type includes battery unit replacement, connector tightening, and cooling system cleaning; Predicting the remaining safe operating time of the distributed energy storage device based on the device health score decrease rate output by the AI ​​prediction model; Dynamically adjust the start time and duration of the maintenance time window according to the remaining safe operation time and the standard working hours required for equipment maintenance; The maintenance operation type and the maintenance time window are encapsulated as the predictive maintenance instruction and sent to the target maintenance terminal.

[0015] By applying the embodiment of the present application, the method innovatively combines the equipment health decline trend with the operation and maintenance resource scheduling to generate a dynamic maintenance time planning model. The method realizes intelligent optimization and scheduling of maintenance windows through predictive calculation of the remaining safe operation time, effectively coordinates equipment operation requirements and maintenance operation requirements, minimizes the impact of maintenance operations on system availability, and improves the comprehensive operation efficiency of the energy storage system.

[0016] As an optional embodiment, the method further includes: After generating the predictive maintenance instruction, monitoring the actual operation response data of the distributed energy storage device in real time; Extracting maintenance effect features from the actual operation response data, the maintenance effect features including equipment temperature drop rate, current waveform smoothness, and voltage stability improvement index; Feeding back the maintenance effect characteristics to the AI ​​prediction model, and updating the parameter weights of the AI ​​prediction model through an incremental learning algorithm; Based on the updated AI prediction model, iterative prediction is performed on the multi-dimensional operating data collected during the target period to optimize the accuracy of the maintenance urgency indicator.

[0017] By applying the embodiment of the present application, the method generates a complete model closed-loop optimization mechanism, and realizes the continuous evolution of the prediction model through the synergy of maintenance effect feature feedback and incremental learning. This method effectively solves the model drift problem caused by equipment aging, maintains the adaptability of the prediction system to changes in equipment status, forms a virtuous cycle of "monitoring-maintenance-optimization", and significantly improves the reliability of the long-term operation of the system.

[0018] As an optional embodiment, updating the parameter weights of the AI ​​prediction model by an incremental learning algorithm includes: Extracting positive feedback samples and negative feedback samples from the maintenance effect features, the positive feedback samples representing that the device state after the maintenance operation meets the set conditions, and the negative feedback samples representing that the device state after the maintenance operation does not meet the set conditions; Calculating the difference in feature distribution between the positive feedback sample and the negative feedback sample to generate a model weight adjustment direction vector; Using an online gradient descent algorithm, fine-tuning the convolution kernel parameters and attention weights of the AI ​​prediction model according to the weight adjustment direction vector; During the fine-tuning process, an elastic weight solidification mechanism is added to lock the model parameters associated with historical failure modes and update the local parameters associated with the newly added maintenance effect characteristics.

[0019] By applying the embodiment of the present application, the method adopts an elastic weight solidification mechanism to achieve a balance between knowledge protection and adaptive learning in the incremental model update process. This method effectively prevents the forgetting of important historical fault features through feature distribution difference analysis and parameter selective update, while quickly adapting to new fault modes, ensuring the stability of the model iteration process, and providing technical support for the sustainable optimization of the prediction system.

[0020] As an optional embodiment, the method further includes: Deploy edge computing nodes in the distributed energy storage device cluster, wherein the edge computing nodes are used to perform local data preprocessing and model reasoning; Performing preliminary anomaly detection on the collected multi-dimensional operation data through the edge computing node to filter out active abnormal data fragments; Uploading the active abnormal data fragments to a cloud server, and adopting a federated learning strategy in the cloud server to aggregate local model updates of multiple edge computing nodes to generate a globally optimized AI prediction model; The globally optimized AI prediction model is distributed to each edge computing node to achieve distributed model collaborative evolution.

[0021] By applying the embodiment of the present application, the method generates a distributed intelligent analysis network through an edge-cloud collaborative computing architecture. The method uses edge computing to realize real-time preprocessing and preliminary diagnosis of local data, combines federated learning to realize global knowledge sharing, improves the model generalization ability under the premise of ensuring data privacy, significantly reduces the communication load of centralized computing, and enhances the system's adaptability to large-scale deployment.

[0022] As an optional embodiment, the execution process of the federated learning strategy includes: Allocate an independent model training copy to each edge computing node and train it locally using active anomaly data fragments; Collect model gradient updates from each edge computing node, and use differential privacy technology to add random noise to the gradient updates; In the cloud server, weighted averaging is performed on the noisy gradient update amount to generate a global gradient update direction label; Update the cloud backbone model parameters according to the global gradient update direction label, and synchronize the updated cloud backbone model parameters to each edge computing node; Model distillation technology is used in the synchronization process to compress the knowledge of the cloud backbone model into a lightweight model on the edge side to maintain the reasoning efficiency under the constraints of edge computing resources.

[0023] By applying the embodiments of this application, the method innovatively integrates differential privacy and model distillation technology to generate a secure and efficient federated learning framework. This method protects data privacy by injecting gradient noise, and combines knowledge distillation to achieve model lightweighting. While maintaining the efficiency of edge-side reasoning, it ensures the knowledge fusion quality of the global model, providing reliable technical support for the evolution of swarm intelligence in distributed energy storage systems.

[0024] As an optional embodiment, the method further includes: Generating a digital twin of the distributed energy storage device, wherein the digital twin is used to map the operating status and environmental parameters of the physical device in real time; Configuring simulated fault scenario features in the digital twin to generate an enhanced set of training data containing multiple fault modes; Using the training data enhancement set to perform adversarial training on the AI ​​prediction model; wherein the adversarial training is used to improve the generalization ability of the AI ​​prediction model for unknown fault modes; Deploy the AI ​​prediction model after adversarial training to the actual operation and maintenance environment, and compare the prediction results of the digital twin with the actual status of the physical equipment in real time to determine the model prediction deviation; When the model prediction deviation exceeds the fault tolerance threshold, the model rollback mechanism is triggered to switch to the AI ​​prediction model of the previous iteration stage.

[0025] By using the embodiment of the present application, the method generates a virtual-real fusion model training environment through the collaborative application of digital twins and adversarial training. The method uses high-fidelity simulation data to enhance the model's ability to identify unknown faults, combines real-time deviation detection with model rollback mechanisms, effectively improves the robustness of the prediction system, provides safety redundancy for equipment status monitoring under extreme working conditions, and significantly reduces the risk of operation and maintenance decision-making.

[0026] An embodiment of the present application also provides a distributed energy storage device predictive maintenance system, comprising a processor and a memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the above-mentioned AI-based distributed energy storage device predictive maintenance method.

[0027] An embodiment of the present application also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-mentioned AI-based distributed energy storage device predictive maintenance method.

[0028] The embodiment of the present application provides a predictive maintenance system and method for distributed energy storage equipment based on AI. This method realizes accurate monitoring of the operating status of distributed energy storage equipment by integrating comprehensive collection of multi-dimensional operating data and in-depth analysis of AI prediction models. In detail, this method can effectively identify potential abnormal characteristics of equipment, combine quantitative evaluation of maintenance urgency indicators, generate targeted predictive maintenance instructions, significantly improve the timeliness and accuracy of equipment maintenance, reduce the risk of unplanned downtime, extend the service life of equipment, and optimize the efficiency of maintenance resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 A flowchart of an AI-based predictive maintenance method for distributed energy storage equipment provided in an embodiment of the present application.

[0031] Figure 2 A block diagram of a distributed energy storage device predictive maintenance system provided in an embodiment of the present application.

[0032] icon: 100-Distributed energy storage equipment predictive maintenance system; 101 - processor; 102 - memory; 103 - bus. DETAILED DESCRIPTION

[0033] The exemplary embodiments disclosed in the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0034] In order to better understand the above technical scheme, the technical scheme of the present application is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical scheme of the present application, rather than limitations on the technical scheme of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0035] Figure 1This is a flowchart of an AI-based distributed energy storage device predictive maintenance method provided according to an embodiment of the present application, which is applied to a distributed energy storage device predictive maintenance system, including steps 110 to 140.

[0036] Step 110: Collect multi-dimensional operating data of the distributed energy storage device, wherein the multi-dimensional operating data includes device temperature sequence, charging and discharging current waveform and voltage fluctuation characteristics.

[0037] In the embodiment of the present application, a temperature sensor array can be deployed at the key nodes of the distributed energy storage device. The temperature sensor array is composed of twelve PT1000 platinum resistance temperature sensors, which are installed on the surface of the battery module in a three-dimensional grid layout. The temperature measurement value is recorded every five seconds to form a device temperature sequence with time continuity. At the same time, a Hall current sensor with a bandwidth of 100kHz is configured on the DC bus side to capture the current waveform details during the charging and discharging process at a sampling frequency of 100,000 times per second. The current waveform details contain complete harmonic components and transient response characteristics.

[0038] In terms of voltage monitoring, a 16-bit precision analog-to-digital converter is used to synchronously collect the voltage of each branch of the energy storage system, and the voltage fluctuation characteristic parameters are extracted through sliding window variance calculation, including the peak-to-peak fluctuation range and standard deviation statistics within a five-minute time scale. For example, in the actual application of a certain energy storage power station, when the energy storage power station performs a two-hour constant current charging operation, the temperature sensor array detects that the temperature of the middle module of the third battery cluster rises exponentially from 25°C to 48°C. At the same time, the current sensor records the high-frequency ripple component with an amplitude of 15% of the nominal value at the end of charging, and the voltage monitoring system captures the abnormal fluctuation of the single cell voltage range exceeding 350mV in the corresponding period.

[0039] Step 120: Input the multi-dimensional operating data into a pre-trained AI prediction model to generate an operating status prediction result of the distributed energy storage device, wherein the operating status prediction result includes an equipment health score and a set of potential abnormal features.

[0040] For example, when the above multi-dimensional operating data is input into the pre-trained AI prediction model, a deep neural network architecture with spatiotemporal feature extraction capability needs to be generated. The deep neural network architecture uses a three-layer bidirectional long short-term memory network to process the temperature time series, uses a two-dimensional convolutional neural network to analyze the current waveform spectrum, and fuses the voltage fluctuation feature vector through an attention mechanism.

[0041] For another example, the training process of the AI ​​prediction model uses a sample set containing 100,000 sets of historical operating data, covering 23 exemplary status modes such as normal operating conditions, battery aging, and loose connectors. When the current operating data is input, the output layer of the AI ​​prediction model will generate a prediction result containing the equipment health score and a set of potential abnormal features. For example, in an actual case, after the model extracts features from the input data of a certain energy storage container, it identifies features such as abnormal temperature gradient of the positive pole (the temperature difference between adjacent sensors is 7.2°C) and excessive third harmonic content of the charging current (THD=8.7%). The final output equipment health score is 62 points (out of 100), and two potential abnormal features of electrolyte crystallization risk and current collector corrosion are marked.

[0042] Step 130: Based on the operation status prediction result, the correlation between the potential abnormal feature set and the preset fault mode library is analyzed to determine the maintenance urgency index of the distributed energy storage device.

[0043] In detail, when analyzing the correlation between potential abnormal features and preset fault modes, it is necessary to establish an expert knowledge base containing 48 exemplary fault modes. The expert knowledge base adopts an ontological modeling method to decompose each fault mode into a multidimensional parameter combination in the feature vector space and set a dynamic matching threshold.

[0044] For example, by calculating the cosine similarity between the current abnormal feature set and the feature vector of each fault mode, combined with the time series evolution trend analysis, the maintenance urgency index is determined. For another example, when a battery module is detected to have an abnormal temperature rise rate (0.8℃ / min), an early charging terminal voltage platform (120 seconds earlier than the standard curve), and an AC impedance spectrum phase angle shift (45-degree frequency point shift of 12%), it is matched with the "abnormal proliferation of negative electrode SEI film" mode in the knowledge base, and the maintenance urgency index of 0.83 (range 0-1) is calculated. This index combines multi-dimensional evaluation results such as the probability of failure (72%), the severity level of consequences (secondary), and the remaining service life prediction of the equipment (estimated number of cycles remaining 83 times).

[0045] Step 140: Generate a predictive maintenance instruction for the distributed energy storage device according to a comparison result between the maintenance urgency index and a preset maintenance threshold, wherein the predictive maintenance instruction includes a maintenance time window and a maintenance operation type.

[0046] For example, when generating predictive maintenance instructions, the system maintenance decision module can compare and analyze the maintenance urgency index with the preset three-level threshold system. The three-level threshold system is dynamically adjusted according to the equipment type, operating environment and maintenance resource availability. The exemplary setting is: urgency above 0.7 triggers a red warning (needs to be handled within 24 hours), 0.5-0.7 is an orange warning (needs to be handled within 72 hours), and 0.3-0.5 is a yellow warning (needs weekly maintenance plan arrangement).

[0047] For another example, the constraint satisfaction algorithm can be used to calculate the optimal maintenance time window by combining the equipment operation calendar, grid load forecast data, and maintenance personnel scheduling information. For example, when the maintenance urgency index of a certain energy storage power station reaches 0.78, the system automatically retrieves the power station operation plan for the next three days, avoids the morning load peak period (09:00-11:00), selects the low load window of 02:00-04:00 the next morning, and specifies the battery module number (B3-12-07) to be replaced and the required list of special tools (including torque wrenches, thermal imagers, etc.), and generates detailed work order instructions containing operation steps, safety precautions, and technical parameters.

[0048] In summary, the above method realizes accurate monitoring of the operating status of distributed energy storage equipment by integrating comprehensive collection of multi-dimensional operating data and in-depth analysis of AI prediction models. In detail, this method can effectively identify potential abnormal characteristics of equipment, combine quantitative evaluation of maintenance urgency indicators, generate targeted predictive maintenance instructions, significantly improve the timeliness and accuracy of equipment maintenance, reduce the risk of unplanned downtime, extend the service life of equipment, and optimize the efficiency of maintenance resource allocation.

[0049] In an embodiment of the present application, the multi-dimensional operation data also includes an environmental parameter set, and the environmental parameter set includes a humidity change curve, an environmental temperature gradient, and a vibration frequency spectrum at the device deployment location. On this basis, the multi-dimensional operation data of the distributed energy storage device described in step 110 includes: Step 111: The device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics are acquired in real time through a sensor array deployed on the distributed energy storage device.

[0050] In step 111, twenty-four K-type thermocouple temperature sensors can be installed on the surface of the battery module in an equidistant matrix arrangement, with each sensor spaced ten centimeters apart to form a three-dimensional monitoring network, recording the temperature value of each point at a sampling frequency of twice per second, and generating a device temperature sequence with spatial resolution. A closed-loop current sensor based on the fluxgate principle is installed at the DC side bus, with a measurement bandwidth extended to 200kHz, capturing nanosecond transient distortion in the charging and discharging current waveform at a sampling rate of 200,000 times per second, and completely recording the rising edge overshoot and ripple attenuation characteristics of the current waveform.

[0051] In the voltage monitoring link, a 24-bit high-precision differential ADC module is used to synchronously sample the voltage of each branch of the energy storage system, and the voltage fluctuation spectrum characteristics are extracted through the sliding window fast Fourier transform, including the fundamental amplitude offset and the change rate of the third harmonic content with a period of five minutes. For example, in the actual operation of a distributed energy storage station, when executing a fast frequency modulation response, the temperature sensor array detected an abnormal distribution with a maximum temperature difference of 9.3°C between adjacent modules in the fourth battery compartment. The current sensor synchronously recorded a current spike pulse lasting 120ms during the discharge process, and the voltage monitoring system found that the total harmonic distortion rate of the AC side voltage increased to 6.8% during the corresponding period.

[0052] Step 112: synchronously collect the humidity change curve, the ambient temperature gradient and the vibration frequency spectrum in the environment where the distributed energy storage device is located, and fuse them with the device operation data in a timestamp alignment manner to obtain initial fused data; wherein the device operation data at least includes the device temperature sequence, the charge and discharge current waveform and the voltage fluctuation characteristics acquired in real time.

[0053] For example, a humidity sensor array based on the principle of polymer capacitance is deployed on six azimuth surfaces inside the equipment cabin, and the relative humidity value is recorded at a sampling interval of once per minute to generate a humidity change curve with spatial distribution characteristics. Eight groups of PT100 platinum resistance ambient temperature probes are arranged within two meters of the periphery of the equipment to form a three-dimensional temperature gradient monitoring network, and the ambient temperature distribution matrix is ​​updated every thirty seconds. Vibration monitoring uses a three-axis MEMS accelerometer array to obtain the basic vibration signal of the equipment at a sampling rate of 5,000 times per second, and the vibration frequency spectrum characteristics of the 0-2kHz frequency band are extracted through wavelet packet decomposition.

[0054] For example, the IEEE 1588 precision time protocol is used to align the timestamps of each sensor during data fusion, and the environmental parameters and equipment operation data are integrated into initial fusion data with a unified time reference. For example, during a typhoon passing through a coastal energy storage power station, the humidity sensor detected that the humidity value in the southeast corner of the cabin rose linearly from 45%RH to 82%RH within two hours. The environmental temperature gradient monitoring showed that the temperature difference between the north side of the equipment and the environment was 15°C. The vibration frequency spectrum analysis found that the basic resonance frequency shifted to 23.5Hz and was accompanied by abnormal accumulation of second-order harmonic energy. These environmental parameters formed a temporal and spatial correlation with the abnormal phenomenon of a 0.15V drop in the terminal voltage of the battery module during the same period.

[0055] Step 113: Perform noise filtering and data normalization processing on the initial fusion data to generate target multi-dimensional operation data.

[0056] In step 113, the improved Savitzky-Golay filter can be used to smooth the temperature series, the window width is set to 21 sampling points, the polynomial order is third order, and the ±0.3°C random fluctuation caused by the thermal noise of the sensor is effectively suppressed. The current waveform data is subjected to a denoising algorithm based on wavelet threshold, and the sym8 wavelet base is selected for five-layer decomposition. The thresholds of each scale are determined by the stein unbiased risk estimation to eliminate the pseudo zero-crossing distortion caused by high-frequency electromagnetic interference.

[0057] Exemplarily, the voltage fluctuation feature processing adopts the Z-score standardization method, and the mean and standard deviation of each characteristic parameter are calculated based on the data of the past 24 hours, and the five-minute voltage range is normalized to the interval [-1, 1]. For example, in the data processing example of a frequency-modulated energy storage system, the original current waveform is denoised to effectively filter out the high-frequency noise above 200kHz caused by the switching action of the converter, retaining the substantial harmonic components with an amplitude exceeding 8% of the nominal value, and at the same time, the temperature and environmental data with different sampling rates are uniformly resampled to one data point per minute to ensure the time alignment of multi-dimensional data.

[0058] As an optional embodiment, the pre-trained AI prediction model generates the running status prediction result in the following manner: Step 210: Divide the target multi-dimensional operation data into data segments within a continuous time window, and extract the temporal dependency and cross-dimensional correlation features in each data segment.

[0059] When the target multi-dimensional operation data is divided into data segments within a continuous time window in step 210, a sliding window mechanism is used to set a basic analysis window of 60 minutes, and the overlap rate of adjacent windows is 50%. Each data segment contains 3,600 temperature sampling points, 7.2 million current waveform sampling points, and 360 sets of voltage characteristic parameters. The dynamic time warping algorithm is used to calculate the phase delay between different sensor signals to establish a time-delay correlation model between the temperature gradient change rate and the current harmonic content.

[0060] For example, cross-dimensional correlation feature mining is achieved through exemplary correlation analysis to calculate the coupling coefficient between the energy distribution of the environmental vibration spectrum and the change in the battery internal resistance. For example, in an analysis case of a cascade utilization energy storage system, it was found through time window division that for every 1°C / m increase in the ambient temperature gradient at the end of charging, the corresponding battery module temperature difference expansion coefficient increased by 0.15, and at the same time, the third harmonic content of the current waveform and the fundamental frequency energy of the environmental vibration showed a Pearson correlation coefficient of 0.82.

[0061] Step 220: Perform local feature enhancement on the temporal dependency through the deep convolutional network layer in the AI ​​prediction model to generate an enhanced spatiotemporal feature vector.

[0062] In step 220, when the local features of the temporal dependency are enhanced through the deep convolutional network layer, an expanded convolution structure with residual connection is generated, the convolution kernel size is set to seven, and the expansion coefficient is three, which expands the receptive field to three hundred sampling points while maintaining the temporal resolution. The spatiotemporal feature vector is generated by using a three-dimensional convolution operation, stacking the temperature spatial distribution matrix, the current waveform time-frequency spectrum, and the voltage fluctuation feature vector as a multi-channel input, and extracting cross-scale spatiotemporal features through sixteen layers of deep convolution.

[0063] For example, in a battery aging identification task, the convolutional network layer successfully captured the nonlinear relationship between the temperature rise rate of the positive electrode and the voltage relaxation time of the negative electrode during the charging process, compressing the original feature dimension from 256 dimensions to a 64-dimensional feature vector with physical significance.

[0064] Step 230: Utilize the attention mechanism layer in the AI ​​prediction model to assign weights to the cross-dimensional correlation features and screen out a subset of key features related to device health.

[0065] When using the attention mechanism layer to allocate weights in step 230, a multi-head self-attention structure is used to set eight parallel attention heads, each of which is responsible for processing feature interactions at different time scales. The key feature subset screening is achieved by calculating the feature importance score, and a threshold is set to retain feature dimensions with scores higher than the 75th percentile. For example, in a certain overtemperature warning case, the attention mechanism layer assigns a weight coefficient of 0.62 to the ambient humidity change rate feature, which is significantly higher than the weight of 0.34 for the ambient temperature gradient, accurately identifying the key association of high humidity environments that accelerate the risk of thermal runaway.

[0066] Step 240: Based on the comparative analysis of the key feature subset and the historical fault data, output the equipment health score and the potential abnormal feature set.

[0067] In step 240, based on the comparative analysis of the key feature subset and the historical fault data, a comparative database containing 50,000 groups of fault samples can be generated, and the dynamic time warping distance is used to measure the similarity between the current feature and the historical fault mode. The equipment health score calculation adopts a fuzzy logic reasoning system, sets 72 fuzzy rules, and maps the feature deviation to a score range of 0-100 points. For example, in the prediction of a battery lithium deposition fault, the system compares the current feature with the spectral features of 300 historical lithium deposition faults, and calculates the Mahalanobis distance to be 2.35, triggering the health score to drop to 55 points, and marking two potential abnormal features: abnormal negative electrode voltage platform and shortened temperature relaxation time.

[0068] As an optional embodiment, the step 130 of analyzing the correlation between the potential abnormal feature set and the preset fault mode library to determine the maintenance urgency index of the distributed energy storage device includes: Step 131: extracting abnormal waveform patterns, temperature mutation intervals and current distortion features from the potential abnormal feature set.

[0069] For example, the improved dynamic time warping algorithm can be used to perform morphological analysis on the charging and discharging current waveforms. By calculating the rate of change of the waveform curvature, the abnormal pulse segments with amplitudes exceeding 12% of the nominal value and durations greater than 50ms can be identified. At the same time, the sliding window second-order difference method can be used to detect the mutation intervals in the temperature series with a rate of change exceeding 3°C / min. For the current distortion characteristics, the original signal is decomposed into eight intrinsic mode functions through the Hilbert-Huang transform, and the energy proportion of the modal component in the frequency range of 150-250Hz is extracted as the distortion feature.

[0070] For example, in a fault diagnosis case of a certain energy storage power station, the system detected a periodic spike pulse with an amplitude of 18% of the rated current at the end of charging. The corresponding battery module surface temperature jumped from 32°C to 49°C within five minutes. At the same time, the current distortion feature analysis showed that the proportion of high-frequency modal energy exceeded the baseline value by 2.3 times, forming a complete set of abnormal features.

[0071] Step 132: performing morphological matching between the abnormal waveform pattern and the known fault waveforms in the preset fault pattern library, and calculating a waveform similarity index.

[0072] When the abnormal waveform pattern is morphologically matched with the preset fault pattern library in step 132, a feature database containing fifty-six standard fault waveforms is generated, and the dynamic time bending distance is used as the similarity measurement index, and the matching threshold is set to 0.7. For example, the abnormal waveform can be resampled to the standard length by cubic spline interpolation, and the minimum bending path cost between it and each template in the fault library is calculated. For example, when it is detected that the charging current waveform has an oscillation mode with an amplitude modulation depth of 14% and a duration of 80ms, the system calculates that its similarity index with the "contact resistance increase" template in the fault library is 0.82, and exceeding the threshold triggers the matching confirmation of the corresponding fault type.

[0073] Step 133: Determine a temperature-associated risk value according to an overlap ratio between the temperature mutation interval and a historical fault temperature threshold.

[0074] When determining the temperature-associated risk value in step 133, a historical fault temperature distribution model based on kernel density estimation is established to calculate the probability density overlap area between the current temperature mutation interval and the historical fault temperature curve. For example, the temperature time series can be divided into sampling points at one-minute intervals, and the frequency of occurrence of the temperature value of each point in the historical fault temperature interval is counted, and the cumulative probability value is calculated after smoothing with a Gaussian kernel function. For another example, a battery module experiences a nonlinear temperature rise from 28°C to 53°C within fifteen minutes. The system calculates that the overlapping area of ​​the temperature trajectory and the historical data of thermal runaway faults accounts for 68%, corresponding to a temperature-associated risk value of 0.73.

[0075] Step 134: Generate a comprehensive anomaly confidence level based on the distribution density of the current distortion feature in the time dimension.

[0076] When generating the comprehensive abnormal confidence in step 134, the time-space density clustering algorithm is used to analyze the time distribution characteristics of the current distortion characteristics, and the frequency of occurrence and spatial distribution entropy of abnormal events per unit time are calculated. In the specific implementation, the 24-hour operation period is divided into windows with five-minute intervals, and the number of times the current distortion characteristics exceed the threshold in each window is counted, and the intensity of abnormal events is estimated through the Poisson distribution model.

[0077] For example, during an eight-hour operation period, a storage system detected nineteen events in which the third harmonic distortion rate of current exceeded 7%. After calculation, the comprehensive abnormality confidence corresponding to the time distribution density was 0.67. This value was combined with the spatial distribution entropy value of 0.54 to finally generate a comprehensive abnormality confidence of 0.62.

[0078] Step 135: Determine the maintenance urgency index based on the weighted fusion result of the waveform similarity index, the temperature-related risk value and the comprehensive anomaly confidence.

[0079] When determining the maintenance urgency index in step 135, the hierarchical analysis method is used to generate a third-order judgment matrix, and the waveform similarity index weight is set to 0.42, the temperature-related risk value weight is set to 0.38, and the comprehensive abnormality confidence weight is set to 0.2. The weighted fusion formula is: maintenance urgency index = waveform similarity index * 0.42 + temperature-related risk value * 0.38 + comprehensive abnormality confidence * 0.2.

[0080] For example, in a certain fault case, the three indicators are 0.82, 0.73 and 0.62 respectively. After weighted calculation, the maintenance urgency index is = 0.82*0.42+0.73*0.38+0.62*0.2=0.7458≈0.75. This value exceeds the preset red warning threshold of 0.7, triggering the emergency maintenance response mechanism.

[0081] As an optional embodiment, the step 140 of generating a predictive maintenance instruction for the distributed energy storage device according to the comparison result between the maintenance urgency index and the preset maintenance threshold includes: Step 141: According to the numerical range of the maintenance urgency indicator, a corresponding maintenance operation type is matched from a preset maintenance strategy table, wherein the maintenance operation type includes battery unit replacement, connector tightening, and cooling system cleaning.

[0082] When matching the maintenance operation type in step 141, the preset maintenance strategy table contains a three-level response mechanism: when the maintenance urgency index is between 0.7 and 1, the battery cell replacement operation is matched; 0.5 to 0.7 matches the connector tightening operation; 0.3 to 0.5 matches the cooling system cleaning operation. In the specific implementation, the system maps the numerical interval to a specific maintenance item by means of a lookup table. For example, the maintenance urgency index of 0.76 corresponds to the battery cell replacement operation, and at the same time retrieves the equipment topology structure to determine the module number to be replaced. In the case of a frequency-modulated energy storage system, the system automatically generates an operation instruction for replacing the seventh module of the third battery compartment based on the maintenance urgency index of 0.78, and associates the factory batch and performance parameters of the module.

[0083] Step 142: Based on the equipment health score decrease rate output by the AI ​​prediction model, predict the remaining safe operating time of the distributed energy storage device.

[0084] When predicting the remaining safe operation time in step 142, the exponential smoothing method is used to establish a time series prediction model for the equipment health score, and the moving average of the score decline rate is calculated based on the score data of the last 24 hours. For example, the remaining safe operation time is equal to the absolute value of the current health score minus the safety threshold divided by the score decline rate. For example, the current health score of a certain energy storage container is 58 points, the score decline rate in the past six hours is 0.75 points per hour, and the safety threshold is set to 45 points. The remaining safe operation time is calculated as (58-45) / 0.75≈17.3 hours.

[0085] Step 143: Dynamically adjust the start time and duration of the maintenance time window according to the remaining safe operating time and the standard working hours required for equipment maintenance.

[0086] When dynamically adjusting the maintenance time window in step 143, a multi-constraint optimization model is established. The objective function is to minimize the impact of maintenance operations on system availability. The constraints include maintenance personnel working hours, spare parts inventory status, and grid load forecast data. Yes, a genetic algorithm is used to solve the optimal time window, set the crossover probability to 0.85, the mutation probability to 0.01, and converge after fifty iterations. For example, a two-hour battery module replacement operation is required for a certain energy storage power station. The system combines the load forecast for the next 24 hours, avoids the frequency regulation peak period from 9:00 to 11:00 in the morning, selects the maintenance window from 1:00 to 3:00 in the morning of the next day, and reserves 15 minutes of safety buffer time.

[0087] Step 144: Encapsulate the maintenance operation type and the maintenance time window into the predictive maintenance instruction, and send it to the target maintenance terminal.

[0088] When encapsulating the predictive maintenance instructions in step 144, the data structure defined by the ISO 13374 standard is used to encode the maintenance operation type into a four-bit opcode, the maintenance time window is accurate to the minute-level timestamp, and the hash value of the safety operation specification document is attached. The instruction transmission uses the MQTT protocol to send it to the maintenance terminal through a TLS encrypted channel, and each data packet contains a CRC-32 checksum. For example, the generated maintenance instruction contains the opcode 0401 (battery cell replacement), the time window 2023-11-05T01:00:00 to 2023-11-05T03:15:00, the target module number B2-07-13, and the SHA-256 digest value of the required tool list to ensure the integrity and traceability of the instruction.

[0089] As an optional embodiment, the method further includes: Step 310: After generating the predictive maintenance instruction, monitor the actual operation response data of the distributed energy storage device in real time.

[0090] In step 310, during the first charge and discharge cycle after the maintenance operation is completed, the high-density data acquisition mode is enabled, the temperature sensor sampling frequency is increased to four times per second, the current sensor bandwidth is extended to 500kHz, and the voltage monitoring system is switched to a 12-bit precision continuous sampling mode. The three-axis vibration sensor array deployed on the equipment body is used to capture the mechanical vibration spectrum changes after maintenance at a sampling rate of 10,000 times per second.

[0091] For example, after a battery cell replacement operation was performed in a certain energy storage power station, the system continuously monitored the surface temperature changes of the replacement module B2-07-13 and found that the maximum temperature during the charging process dropped from 67°C before maintenance to 51°C. At the same time, the current sensor recorded that the amplitude of the original 120Hz ripple component decreased from 9.3% of the nominal value to 2.1%. The voltage monitoring data showed that the voltage difference between modules was reduced from 230mV before maintenance to 45mV, forming a complete actual operation response data set.

[0092] Step 320: extracting maintenance effect features from the actual operation response data, wherein the maintenance effect features include equipment temperature drop rate, current waveform smoothness, and voltage stability improvement index.

[0093] When extracting the maintenance effect characteristics in step 320, the sliding window analysis method is used to calculate the equipment temperature drop rate, a ten-minute time window is set, and the slope value of the temperature change curve is fitted by linear regression as a quantitative indicator. The current waveform smoothness evaluation uses discrete Fourier transform to extract the total harmonic distortion rate of the 0-1kHz frequency band, and combines the waveform peak-to-peak fluctuation coefficient for comprehensive evaluation. The voltage stability improvement index is obtained by calculating the weighted combination of the voltage standard deviation and the range value within a five-minute time scale.

[0094] For example, in a case of fastening maintenance of a certain connector, monitoring data twenty-four hours after maintenance showed that the temperature drop rate reached 0.8°C / min, the total harmonic distortion rate of the current waveform dropped from 6.7% to 1.9%, and the voltage standard deviation improved from 12.3mV to 3.1mV, forming a feature vector containing twelve parameters in three dimensions.

[0095] Step 330: Feedback the maintenance effect characteristics to the AI ​​prediction model, and update the parameter weights of the AI ​​prediction model through an incremental learning algorithm; based on the updated AI prediction model, iteratively predict the multi-dimensional operating data collected during the target period to optimize the accuracy of the maintenance urgency indicator.

[0096] When the maintenance effect characteristics are fed back to the AI ​​prediction model in step 330, an incremental learning framework is generated, and the model update cycle is set to 24 hours. Each update injects the operating data within 72 hours after the maintenance. The incremental learning algorithm adopts an elastic weight solidification mechanism, and applies L2 regularization constraints to the parameters of the bidirectional long short-term memory network layer in the model, limiting the parameter update amplitude to no more than 15% of the original value.

[0097] For example, in a certain model update, the system injected three hundred sets of post-maintenance data samples, adjusted the weights of the sixteen convolution kernels of the two-dimensional convolutional neural network layer through the online gradient descent algorithm, and locked the eight attention mechanism parameters associated with historical thermal runaway failures, thereby increasing the model's prediction accuracy for loose connector failures from 83.2% to 89.7%.

[0098] As an optional embodiment, the updating of the parameter weights of the AI ​​prediction model by the incremental learning algorithm in step 330 includes: Step 331: extracting positive feedback samples and negative feedback samples from the maintenance effect characteristics, wherein the positive feedback samples represent that the device state after the maintenance operation meets the set conditions, and the negative feedback samples represent that the device state after the maintenance operation does not meet the set conditions.

[0099] When extracting positive feedback samples and negative feedback samples in step 331, set the device temperature drop rate threshold of 0.5℃ / min, the current waveform smoothness improvement rate of 30%, and the voltage stability improvement index of 40% as judgment conditions. Positive feedback samples must meet all three conditions at the same time, while negative feedback samples include any one of the conditions that does not meet the standard. For example, after a battery cell is replaced, the temperature drop rate is monitored to be 0.63℃ / min, the current smoothness is improved by 40%, and the voltage stability is improved by 55%. The sample is marked as positive feedback; and after another cooling system is cleaned, only the temperature drop rate is 0.38℃ / min, which is classified as a negative feedback sample.

[0100] Step 332: Calculate the difference in feature distribution between the positive feedback sample and the negative feedback sample to generate a model weight adjustment direction vector; use an online gradient descent algorithm to fine-tune the convolution kernel parameters and attention weights of the AI ​​prediction model according to the weight adjustment direction vector; add an elastic weight solidification mechanism during the fine-tuning process to lock the model parameters that are associated with historical failure modes, and update local parameters associated with the newly added maintenance effect features.

[0101] When adjusting the model weight in step 332, the KL divergence of the positive and negative feedback samples in the feature space is first calculated to generate a weight adjustment direction vector with a dimension of 256. The stochastic gradient descent algorithm with momentum is used, and the learning rate is set to 0.0001 and the momentum coefficient is set to 0.9 to fine-tune the parameters of the three-dimensional convolution layer in the AI ​​prediction model. The elastic weight solidification mechanism calculates the parameter importance score through the Fisher information matrix and implements freeze protection for 52 parameters with an importance higher than 0.8.

[0102] For example, in a certain model update, the system detected that the Mahalanobis distance between the smoothness characteristics of the current waveform in the newly added maintenance data and the historical data was 3.2, so it adjusted the relevant convolution kernel parameters by 12%, while keeping the attention weight associated with the electrolyte crystallization fault unchanged, which increased the model's recognition rate for the newly emerging contact resistance increase fault by 11.3%.

[0103] As an optional embodiment, the method further includes: Step 410: deploy edge computing nodes in the distributed energy storage device cluster, and the edge computing nodes are used to perform local data preprocessing and model reasoning; perform preliminary anomaly detection on the collected multi-dimensional operating data through the edge computing nodes to screen out active abnormal data fragments.

[0104] In step 410, an ARM-based edge computing hardware unit is installed on top of each energy storage container. The unit is equipped with a quad-core Cortex-A72 processor and 8GB LPDDR4 memory, and establishes a communication connection with the local sensor array through the CAN bus protocol to ensure that the data collection delay is less than 20 milliseconds. The edge computing hardware unit has a built-in real-time operating system to perform the first-level processing of the multi-dimensional operating data from the temperature sensor array, current sensor module and voltage monitoring device.

[0105] For example, in a wind farm supporting energy storage system, twelve edge computing nodes are deployed in six energy storage containers. Each node is responsible for processing the operating data of the twenty-four battery modules in the container, eliminating sensor noise through the sliding window mean filtering algorithm, and aligning the timestamps of temperature data (twice per second) with current waveform data (200,000 times per second) with different sampling rates to generate a time-synchronized pre-processed data stream. When an abnormal temperature fluctuation is detected in the seventh module of the third container, the edge computing node completes the data standardization processing locally and selects a ten-minute data segment containing abnormal features.

[0106] Step 420: Upload the active abnormal data segment to the cloud server, and adopt a federated learning strategy in the cloud server to aggregate local model updates of multiple edge computing nodes to generate a globally optimized AI prediction model; distribute the globally optimized AI prediction model to each edge computing node to achieve distributed model collaborative evolution.

[0107] For example, the cloud server establishes a secure access channel based on OAuth 2.0 authentication to receive encrypted data packets from thirty-six edge computing nodes, each of which contains compressed abnormal feature vectors, local model parameter increments, and device operating environment metadata. The federated learning aggregation module uses a secure multi-party computing protocol to homomorphically encrypt the long-short-term memory network layer gradient matrix of twelve edge nodes in the distributed energy storage device cluster, and generates global model update parameters through a weighted average algorithm. The weight distribution coefficient is dynamically adjusted according to the data volume proportion of each node and the data quality score.

[0108] For example, in a certain level of energy storage monitoring system, the cloud server aggregates the model update data of six regional substations, of which the eastern substation contributes 40,000 sets of battery aging samples and the western substation provides 20,000 sets of environmental mutation samples. After calculation by the federated average algorithm, the accuracy of the global model in detecting temperature gradient anomalies is increased to 92.7%. The updated globally optimized AI prediction model generates a model parameter differential package through a differential privacy protection mechanism, which is signed with an elliptic curve encryption algorithm and synchronized to all edge computing nodes via a content distribution network to ensure that the full-node model iteration is completed within fifteen minutes. In a specific implementation case, after a certain energy storage cluster completes the federated learning update, the edge node's detection response time for contact resistance increase faults is shortened from eight seconds to three seconds, and the false alarm rate is reduced by 40%.

[0109] As an independently implementable technical solution, the step 410 of deploying edge computing nodes in the distributed energy storage device cluster includes: Step 411: deploy an edge computing hardware unit in the neighborhood of each physical node of the distributed energy storage device cluster, and connect the edge computing hardware unit to the local sensor array through a target delay communication protocol; receive the multi-dimensional operation data through the edge computing hardware unit, perform format standardization and timestamp alignment on the multi-dimensional operation data, and generate a preprocessed data stream.

[0110] The specific method of deploying the edge computing hardware unit in step 411 is: installing a waterproof and explosion-proof edge computing hardware unit within 0.5 meters of the battery module string electrical interface, the unit is equipped with a dual Gigabit Ethernet interface and a LoRa wireless communication module, and establishes a connection with the local sensor array through the Modbus-TCP protocol. The data preprocessing process includes converting the original resistance value of the PT1000 platinum resistance temperature sensor into a Celsius temperature value, normalizing the voltage signal output by the Hall current sensor, and resampling the voltage fluctuation characteristic data of different sampling rates to a unified time base.

[0111] For example, in a commercial and industrial energy storage project, the edge computing hardware unit receives data streams from thirty-two temperature sensors, converts them into the IEEE 754 standard floating-point format, and synchronizes the timestamp with the twelve-bit ADC sampling value of the current sensor to generate a fused data packet with a time resolution of one millisecond.

[0112] Step 412: Load a lightweight AI reasoning model in the edge computing hardware unit, input the preprocessed data stream into the lightweight AI reasoning model, and generate a preliminary anomaly detection result; the preliminary anomaly detection result includes an anomaly type label and a confidence score; based on the comparison result of the confidence score with a preset anomaly threshold, filter out the initial anomaly data fragment that meets the confidence requirement from the preprocessed data stream.

[0113] When performing preliminary anomaly detection in step 412, the edge computing hardware unit loads a lightweight AI inference model developed based on the TensorFlow Lite framework. The model includes a two-layer one-dimensional convolutional neural network and a long short-term memory network combination structure. The input layer receives a multi-dimensional data window with a length of three hundred sampling points, and the output layer generates a probability distribution of three types of abnormal labels including over-temperature warning, current distortion, and voltage imbalance.

[0114] For example, when the input data window shows that the temperature rise rate of a battery module reaches 1.8℃ / s within 30 seconds, and the third harmonic content of the charging current exceeds the nominal value by 12%, the model outputs an overtemperature warning confidence of 0.87 and a current distortion confidence of 0.68, which exceeds the preset abnormal threshold of 0.7, triggering the data segment to be marked as an active abnormal data segment. The system automatically extracts the complete data window of five minutes before and after the abnormality occurs, and generates a metadata tag containing timestamp, device number and abnormality type.

[0115] Step 413: Perform feature compression encoding on the initial abnormal data fragment that meets the confidence requirement to generate a bandwidth transmission data packet; upload the bandwidth transmission data packet to the cloud server through an encrypted channel to trigger the model aggregation process of the federated learning strategy.

[0116] When feature compression encoding is performed in step 413, the principal component analysis algorithm is used to reduce the dimension of the initial abnormal data fragment from 256 dimensions to 32 dimensions, while retaining 95% of the original information volume. The compressed data packet is attached with a SHA-256 hash checksum, encapsulated by the AES-256 encryption algorithm, and transmitted to the cloud server via the MQTT protocol. For example, after an edge computing node detects a voltage fluctuation anomaly that lasts for two minutes, it compresses the original data stream of 12 million sampling points into a twelve-megabyte feature vector packet, adds the metadata of the timestamp 2023-11-05T14:23:00 and the device number B3-07-12, and uploads it to the cloud storage pool through the TLS 1.3 encrypted channel.

[0117] Step 414: After the cloud server completes the generation of the globally optimized AI prediction model, it receives the updated model parameter file issued; loads the updated model parameter file into the lightweight AI reasoning model through a hot update mechanism, overwriting the original model weights.

[0118] When updating the model parameters in step 414, the cloud server generates a differential update file, which only contains 15% of the key parameter changes in the lightweight AI reasoning model. The edge computing hardware unit downloads an 800-kilobyte model update package through the HTTPS protocol, and uses hot patching technology to replace the weight matrix of the bidirectional long short-term memory network layer without interrupting service. For example, after a certain federated learning aggregation, the update package sent by the cloud adjusted the sixteen weight parameters of the current waveform analysis convolution kernel. The edge computing node completed the model reload within twenty milliseconds, and the new reasoning model’s detection accuracy for loose connector failures increased from 78% to 85%.

[0119] Step 415: Based on the updated lightweight AI inference model, perform iterative inference on the preprocessed data stream input after the update time node to generate an optimized preliminary anomaly detection result; cross-validate the optimized preliminary anomaly detection result with the real-time data stream of the local sensor array, eliminate false positive anomaly fragments and correct the confidence score; dynamically adjust the screening threshold of the initial abnormal data fragments that meet the confidence requirement according to the corrected confidence score to achieve adaptive anomaly detection on the edge side.

[0120] When iterative reasoning optimization is performed in step 415, the updated lightweight AI reasoning model re-analyzes the real-time data stream and verifies the temporal and spatial correlation of the preliminary anomaly detection results with the local vibration sensor data. For example, when the model detects that the charging current waveform has an amplitude modulation anomaly, the system synchronously checks the vibration spectrum data of the corresponding time period. If no mechanical resonance feature is detected, the anomaly is marked as a false alarm and the confidence score is lowered. The dynamic threshold adjustment algorithm adaptively adjusts the confidence threshold of the overtemperature warning from 0.7 to 0.65 based on the detection results of the past 24 hours to cope with the problem of increased misjudgment rate caused by sudden changes in ambient temperature, ensuring that the anomaly detection recall rate remains above 95%.

[0121] As an optional embodiment, the execution process of the federated learning strategy includes: Step 510: Allocate an independent model training copy to each edge computing node and train it locally using the active anomaly data fragment.

[0122] In step 510, a specific implementation method of allocating an independent model training copy to each edge computing node is as follows: deploying initialized long short-term memory network model copies on twelve edge computing nodes of the distributed energy storage device cluster, each copy having the same network architecture but independent parameter space. During the local training process, 72 hours of active abnormal data fragments stored in the edge node are used as training sets, and these data fragments contain abnormal waveform patterns, temperature mutation intervals, and current distortion features that are feature compressed and encoded.

[0123] For example, in the seventh edge node of a provincial energy storage monitoring system, 325 abnormal data fragments are stored, each of which contains five minutes of multi-dimensional operating data. The small batch gradient descent algorithm is used for training, with the batch size set to 32, the learning rate set to 0.001, and fifty iterative training cycles. This allows the model to accurately identify the abnormal contact resistance pattern caused by salt spray corrosion that is unique to the area. The accuracy of the trained model copy for local anomaly detection tasks is increased to 89.3%.

[0124] Step 520: Collect model gradient updates from each edge computing node, and use differential privacy technology to add random noise to the gradient updates.

[0125] The specific method of collecting the model gradient update amount and adding differential privacy noise in step 520 is as follows: after completing local training, each edge computing node extracts the gradient matrix of the last layer of the long short-term memory network, which has a dimension of 128×256 and contains the update direction information of the weight parameters. The (ε, δ)-differential privacy protection mechanism is adopted to inject Gaussian noise with a mean of zero and a standard deviation of 0.1 into the gradient matrix, where the privacy budget ε is set to 2.0 and δ is set to 1e-5.

[0126] For example, the maximum element value of the original gradient matrix generated by the fifth edge node of the energy storage station in the coastal area is 0.47. After adding noise, the fluctuation range of the gradient matrix element value is controlled within the range of ±0.15, which ensures the effectiveness of the gradient direction while eliminating sensitive information that may expose the details of the specific equipment operation. The noise injection process is implemented through the hardware security module to ensure the unpredictability and non-repeatability of the random number generator.

[0127] Step 530: In the cloud server, weighted average is performed on the noisy gradient update amount to generate a global gradient update direction label.

[0128] When performing weighted averaging to generate the global gradient update direction label in step 530, the cloud server establishes an aggregation module based on the federated average algorithm to allocate weight coefficients according to the number and quality scores of abnormal data segments uploaded by each edge computing node. In a specific implementation, the computing node weight is the ratio of the local data volume to the total data volume multiplied by the data quality factor (in the range of 0.8-1.2), where the data quality factor is determined by the signal-to-noise ratio and feature integrity of the abnormal segment.

[0129] For example, in an energy storage cluster with nine edge nodes, the third node contributes 420 abnormal fragments (accounting for 18% of the total), with a data quality factor of 1.05 and a weight coefficient of 0.18×1.05=0.189; the eighth node contributes 300 fragments (13%) but with a quality factor of 0.9 and a weight coefficient of 0.117. The cloud server performs a weighted summation of the nine noisy gradient matrices to generate a global gradient matrix with unchanged dimensions. The maximum element value of the matrix converges to 0.39, and the directional consistency index reaches 0.87, indicating that the gradient update directions of each node are significantly correlated.

[0130] Step 540: Update the cloud backbone model parameters according to the global gradient update direction label, and synchronize the updated cloud backbone model parameters to each edge computing node.

[0131] When updating the cloud backbone model parameters in step 540, a stochastic gradient descent algorithm with momentum is used, the momentum coefficient is set to 0.9, and the learning rate decays to one tenth of the initial value. During the update process, the twenty-eight core parameters in the backbone model that are strongly associated with the historical failure mode are kept unchanged, and only the remaining 1200 parameters are adjusted.

[0132] For example, a global update increased the detection sensitivity of the long short-term memory network to the temperature mutation range by 12%, but the recognition weight of the known electrolyte leakage mode remained stable. The parameter synchronization adopts a dual-channel mechanism. The main channel transmits the encrypted model parameter file through HTTPS, and the backup channel uses LoRaWAN to broadcast the differential update package to ensure that the parameter synchronization of all edge nodes is completed within fifteen minutes. After the synchronization of a microgrid energy storage system in a mountainous area, the model alignment error of twelve edge nodes was reduced from 7.3% to 1.8%.

[0133] Step 550: During the synchronization process, model distillation technology is used to compress the knowledge of the cloud backbone model into a lightweight model on the edge side to maintain the reasoning efficiency under the constraints of edge computing resources.

[0134] When performing model distillation in step 550, the cloud-side backbone model outputs the unpruned complete probability distribution as the teacher model, and the edge-side lightweight model serves as the soft target for the student model to learn the teacher model. The distillation loss function uses KL divergence to measure the difference between the two distributions, combined with the cross entropy loss of the student model, and the temperature parameter is set to 5.

[0135] For example, when compressing the 256-dimensional output vector of the cloud backbone model to the 64-dimensional one of the edge model, feature mapping is achieved through a three-layer fully connected network. While maintaining a 98% accuracy rate for contact resistance anomaly detection, the model calculation amount is reduced by 62%. After distillation, the edge computing node of a commercial and industrial energy storage project reduced the single inference time from 58ms to 22ms, and the memory usage was compressed from 1.2GB to 380MB, while maintaining a fault recall rate of more than 95%. A progressive knowledge transfer strategy is adopted during the distillation process. In the first round, 80% of the parameter knowledge of the teacher model is transferred, and the subsequent rounds are gradually refined to local features to ensure a smooth transition of the model compression process.

[0136] As an optional embodiment, the method further includes: Step 610: Generate a digital twin of the distributed energy storage device, where the digital twin is used to map the operating status and environmental parameters of the physical device in real time.

[0137] In step 610, the specific implementation method of generating the digital twin of the distributed energy storage device is as follows: based on the physical topological structure of the distributed energy storage device, a parametric modeling tool is used to generate a three-dimensional dynamic structural model, which includes the precise geometric dimensions and material property parameters of 1,200 battery cells, and each cell node is associated with dynamic properties such as thermal conductivity, internal resistance and aging factor.

[0138] For example, the OPC UA protocol can be used to access the device temperature sequence, charge and discharge current waveform, voltage fluctuation characteristics, and ambient temperature and humidity data streams in real time to establish a digital mapping relationship that is updated synchronously with the physical device. For another example, in the digital transformation of a 2MWh energy storage power station, the digital twin accurately reproduced the three-dimensional spatial layout of forty-eight battery clusters, each cluster containing twenty-four series modules, the spacing error of the heat dissipation channels between modules is controlled at ±1mm, and the deviation of the electrical connection path impedance value from the measured value is less than 0.5mΩ. The digital twin receives the real-time operation data stream of the physical device every five seconds, and updates the temperature field distribution cloud map and current density vector map in the three-dimensional dynamic structure model.

[0139] Step 620: Configure simulated fault scenario features in the digital twin to generate an enhanced set of training data containing multiple fault modes.

[0140] When configuring the simulated fault scenario in step 620, twenty-eight exemplary fault modes are injected into the digital twin, including loose connectors, electrolyte leakage, SEI film proliferation, etc. Each fault mode is associated with twelve adjustable parameters, such as the loose contact area decay rate is set to 0.1-0.5mm² / h, and the electrolyte leakage rate is set to 0.5-2ml / min. The training data enhancement set contains mixed samples of normal working conditions and fault conditions, and one hundred thousand sets of enhanced data are generated by random combination of parameters. For example, when simulating a loose connector failure, the digital twin gradually increases the contact resistance value (increases by 0.2mΩ per minute), and simultaneously generates the corresponding temperature gradient change curve and voltage fluctuation spectrum, forming a fault sample containing four hundred time steps.

[0141] Step 630: Use the training data enhancement set to perform adversarial training on the AI ​​prediction model; wherein the adversarial training is used to improve the generalization ability of the AI ​​prediction model for unknown fault modes.

[0142] When adversarial training is performed in step 630, a generative adversarial network architecture is first constructed. The generator receives normal operating data and injects disturbances to generate adversarial samples, and the discriminator distinguishes between real faults and generated faults. During the training process, the generator and the discriminator are optimized alternately to improve the sensitivity of the AI ​​prediction model to hidden fault characteristics. For example, the generator successfully created an adversarial sample with a subtle temperature difference of 0.5°C, which the original model failed to detect. After adversarial training, the model's recognition accuracy for this type of sample increased from 32% to 87%. The F1 score of the trained AI prediction model on the test set containing unknown fault types increased by 15 percentage points.

[0143] Step 640: Deploy the AI ​​prediction model after adversarial training to the actual operation and maintenance environment, and compare the prediction results of the digital twin with the actual status of the physical equipment in real time to determine the model prediction deviation; when the model prediction deviation exceeds the fault tolerance threshold, trigger the model rollback mechanism and switch to the AI ​​prediction model of the previous iteration stage.

[0144] When the adversarial training model is deployed in step 640, an online verification mechanism is established, and the digital twin generates a state prediction every five minutes, which is compared item by item with the actual sensor readings of the physical device. The deviation calculation module uses the Mahalanobis distance to measure the degree of deviation in the multidimensional feature space, and the fault tolerance threshold is set to three times the standard deviation of the historical data distribution. For example, when the digital twin predicts that the temperature of a module is 45°C and the measured value is 51°C, the model prediction deviation index rises to 4.7, exceeding the threshold of 3.0 to trigger the rollback mechanism. The system automatically switches to the stable model version two weeks ago and uploads the abnormal data packet to the cloud analysis platform. The rollback process is completed within 300 milliseconds to ensure the continuity of operation and maintenance decisions.

[0145] As an independently implementable technical solution, the generating of the digital twin of the distributed energy storage device in step 610 includes: Step 611: Acquire the real-time operation data stream of the distributed energy storage device, wherein the real-time operation data stream includes a device temperature sequence, a charge and discharge current waveform, a voltage fluctuation characteristic, and an environmental parameter set; based on the real-time operation data stream, create a three-dimensional dynamic structural model of the distributed energy storage device, wherein the three-dimensional dynamic structural model includes a battery cell topology network, a heat dissipation channel distribution diagram, and an electrical connection path.

[0146] On the one hand, when creating a three-dimensional dynamic structural model in step 611, the finite element analysis method is used to divide the hexahedral grid units, and each battery cell is divided into eight layers of grids, with a total number of grids of 3 million. The battery cell topology network uses a graph database to store the connection relationship, recording the positive and negative connection paths of each unit and the heat conduction paths of adjacent units. The heat dissipation channel distribution map is generated based on the computational fluid dynamics simulation results, and the air flow velocity gradient and temperature attenuation curve of the forced air cooling system are marked. For example, in the digital twin of a liquid-cooled energy storage system, the three-dimensional dynamic structural model accurately reproduces the three-dimensional spiral structure of the cooling pipeline, including the pressure drop coefficients of 360 elbow nodes and the flow distribution parameters of 240 branches, and the matching degree with the infrared thermal imaging measured data of the physical equipment is 97%.

[0147] On the other hand, in the real-time data stream processing of step 611, the data acquisition system captures the charging and discharging current waveform at a sampling rate of 20,000 times per second, and extracts the harmonic characteristics of the 0-500Hz frequency band through the sliding window Fourier transform. The environmental parameter set contains the vibration spectrum data of the deployment location, and wavelet packet decomposition is used to extract the energy values ​​of the sixteen sub-bands in the 0-2kHz frequency band. For example, in the three-dimensional dynamic structure model of a coastal energy storage station, the increase in contact resistance caused by salt spray corrosion is accurately modeled, and the digital twin simulation shows that for every 10% reduction in the contact area of ​​the connector, the corresponding temperature rise rate increases by 0.8℃ / min, which is consistent with the trend of the measured data of the physical equipment under the same working conditions.

[0148] Step 612: configuring a virtual sensor interface in the three-dimensional dynamic structure model, wherein the virtual sensor interface is used to receive the real-time operation data stream and map it to a corresponding position of the three-dimensional dynamic structure model.

[0149] For example, when configuring the virtual sensor interface in step 612, twenty-four virtual temperature probes are set on the surface of each battery cell in the three-dimensional dynamic structure model, and the position distribution is exactly the same as the actual PT1000 sensor array. The virtual sensor interface receives the real-time operation data stream of the physical device through the Apache Kafka message queue, maps the device temperature sequence to the corresponding probe, and decomposes the charge and discharge current waveform into the electrical connection path of each branch. For example, when the physical device detects that the temperature of the third module of the fifth battery cluster is abnormal, the digital twin synchronously renders a red warning area at the corresponding position and highlights the affected thirty-two adjacent cells along the electrical connection path.

[0150] For another example, in the implementation of the virtual sensor interface in step 612, a dedicated data mapping module is developed to convert the four-dimensional data stream (time, space, electrical, and environment) of the physical device into the attribute matrix of the three-dimensional dynamic structure model. Each virtual sensor interface is associated with sixteen data channels, including parameters such as temperature, voltage, and current harmonic content. For example, when the physical device detects a current ripple with an amplitude of 12% of the nominal value at the end of charging, the digital twin renders a pulsating current vector arrow on the corresponding electrical connection path, and its length and color depth reflect the change in ripple amplitude in real time.

[0151] Step 613: Drive the three-dimensional dynamic structure model through the physical module, simulate the dynamic response behavior of the distributed energy storage device under the action of the real-time operation data stream, and generate a device state simulation sequence; align the device state simulation sequence with the real-time operation data stream in time sequence, and establish a bidirectional data mapping channel between the digital twin and the physical device.

[0152] Exemplarily, when driving the three-dimensional dynamic structure model in step 613, a multi-physics field coupling simulation module is used to perform an electric-thermal-mechanical coupling calculation every thirty milliseconds. The charging and discharging current waveforms in the real-time running data stream are input to solve the dynamic balance between the Joule heat distribution and the heat dissipation efficiency of the cooling system, and output the battery cell surface temperature simulation sequence. The timing alignment module uses a dynamic time warping algorithm to synchronize the simulated temperature sequence with the actual data measured by the physical sensor at the millisecond level. For example, when simulating a two-hour constant power discharge condition, the maximum temperature difference between modules calculated by the digital twin is 8.3°C, and the deviation from the actual measured value of 8.1°C by the infrared thermal imager is controlled within 2.4%, verifying the accuracy of the bidirectional data mapping channel.

[0153] In addition, in the physical module simulation of step 613, the explicit time integration algorithm is used to solve the partial differential equation of heat conduction, and the temperature field distribution of all battery cells is calculated at each time step. The dynamic response behavior simulation includes the transient effects of the start and stop of the cooling system. For example, when the forced air cooling system is delayed for three seconds to start, the digital twin accurately predicts that the temperature at the top of the battery compartment will exceed the limit by 2.3°C within the next five minutes. The prediction result has been verified by actual testing and the error is only 0.7°C.

[0154] Step 614: Based on the bidirectional data mapping channel, the deviation signal between the device state simulation sequence and the actual operating state of the physical device is detected in real time to generate a dynamic calibration instruction set; the parameterized components of the three-dimensional dynamic structure model are adjusted according to the dynamic calibration instruction set to synchronize the simulation output of the digital twin with the measured data of the physical device.

[0155] In an embodiment of the present application, when dynamic calibration is performed in step 614, the deviation detection module calculates the root mean square error between the simulated temperature sequence and the measured data, and generates a parameter calibration instruction when the error exceeds 1.5°C. The calibration algorithm uses a Bayesian optimization method to adjust the thermal conductivity and contact resistance values ​​in the three-dimensional dynamic structure model. For example, during a calibration process, it was detected that the simulated temperature of the seventh module was 2.8°C higher than the measured value. The system automatically corrected the interface contact thermal resistance in the model from 0.8K·m² / W to 0.92K·m² / W, so that the simulation error in the next three hours was stabilized within 0.5°C.

[0156] In addition, during the dynamic calibration process of step 614, an adaptive parameter adjustment algorithm is developed to automatically correct the thermal time constant of the three-dimensional dynamic structure model when the phase deviation between the simulated temperature sequence and the measured data exceeds five seconds. The calibration instruction set contains twelve types of adjustable parameters, each of which has an adjustment range of ±15%. For example, a calibration corrected the air convection coefficient between modules from 8.7W / (m²·K) to 9.3W / (m²·K), reducing the average error of the simulation results for the next two hours to 0.4℃.

[0157] Step 615: embed an abnormal propagation path prediction module in the digital twin, simulate the abnormal diffusion path of the fault in the battery unit topology network based on the abnormal characteristics of the equipment state simulation sequence; superimpose the abnormal diffusion path with the electrical connection path of the three-dimensional dynamic structure model to generate a visualized fault impact map; and feed back the visualized fault impact map to the physical equipment operation and maintenance terminal through the bidirectional data mapping channel to achieve self-update of the digital twin.

[0158] For example, when implementing the abnormal propagation prediction in step 615, a graph neural network is used to analyze the fault diffusion path in the battery cell topology network. When it is detected that the internal resistance of a certain unit has increased abnormally by 5%, the system simulates the propagation process of the fault along the electrical connection path, and predicts that the internal resistance of twelve adjacent units will increase by more than 3% in the next thirty minutes. The visual fault impact map uses a gradient coloring scheme, with red areas representing directly affected units and orange representing secondary affected units. The generated thermal map is superimposed on the three-dimensional structural model. For example, a simulation showed that a poor contact fault would cause the temperature rise rate of the six upstream modules to increase by 40%. The prediction result was verified by the actual operation data of the physical equipment for the next thirty minutes, with an accuracy rate of 89%.

[0159] For another example, in the abnormal propagation prediction of step 615, the improved Dijkstra algorithm is used to search for the shortest fault propagation path in the battery cell topology network, and the fault diffusion rate is calculated in combination with the thermal runaway reaction kinetic model. The visual fault impact map marks the four-level impact area. The red core area represents the direct fault unit, the orange first-level impact area includes six adjacent units, the yellow second-level impact area extends to twenty-four units, and the blue third-level impact area covers the entire battery cluster. For example, a simulation shows that an electrolyte leakage fault will cause an internal short circuit in three adjacent modules within eight minutes. The prediction guides the operation and maintenance personnel to complete the isolation of the faulty module fifteen minutes in advance.

[0160] An embodiment of the present application provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the AI-based distributed energy storage device predictive maintenance method.

[0161] An embodiment of the present application provides a processor, which is used to run a program, wherein the program executes the AI-based distributed energy storage device predictive maintenance method when running.

[0162] In the present application embodiment, Figure 2As shown, the distributed energy storage device predictive maintenance system 100 includes at least one processor 101, and at least one memory 102 and a bus 103 connected to the processor 101; wherein the processor 101 and the memory 102 communicate with each other through the bus 103; the processor 101 is used to call the program instructions in the memory 102 to execute the above-mentioned AI-based distributed energy storage device predictive maintenance method.

[0163] The present application is described with reference to the flowcharts and / or block diagrams of the methods, distributed energy storage device predictive maintenance systems (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the instructions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0164] In a typical configuration, the distributed energy storage device predictive maintenance system includes one or more processors (CPUs), a memory, and a bus. The distributed energy storage device predictive maintenance system may also include an input / output interface, a network interface, and the like.

[0165] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip. The memory is an example of a computer-readable medium.

[0166] Computer-readable media include permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage computer-readable storage media or any other non-transmission media that can be used to store information that can be accessed by the distributed energy storage device predictive maintenance system. As defined in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carriers.

[0167] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or computer-readable storage medium that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such process, method, commodity or computer-readable storage medium. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the presence of other identical elements in the process, method, commodity or computer-readable storage medium that includes the elements.

[0168] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0169] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A predictive maintenance method for distributed energy storage equipment based on AI, characterized in that: Applied to a predictive maintenance system for distributed energy storage equipment, the method comprises: Collecting multi-dimensional operating data of distributed energy storage devices, wherein the multi-dimensional operating data includes device temperature series, charge and discharge current waveforms, and voltage fluctuation characteristics; Input the multi-dimensional operation data into a pre-trained AI prediction model to generate an operation status prediction result of the distributed energy storage device, wherein the operation status prediction result includes a device health score and a set of potential abnormal features; Based on the operation status prediction result, the correlation between the potential abnormal feature set and the preset fault mode library is analyzed to determine the maintenance urgency index of the distributed energy storage device; According to the comparison result between the maintenance urgency index and the preset maintenance threshold, a predictive maintenance instruction for the distributed energy storage device is generated, and the predictive maintenance instruction includes a maintenance time window and a maintenance operation type.

2. The method according to claim 1, characterized in that The multi-dimensional operation data also includes an environmental parameter set, which includes a humidity change curve, an environmental temperature gradient, and a vibration frequency spectrum at the device deployment location; the multi-dimensional operation data of the distributed energy storage device is collected, including: The device temperature sequence, the charge and discharge current waveform and the voltage fluctuation characteristics are acquired in real time by a sensor array deployed on the distributed energy storage device; Synchronously collecting the humidity change curve, the ambient temperature gradient and the vibration frequency spectrum in the environment where the distributed energy storage device is located, and fusing them with the device operation data in a timestamp alignment manner to obtain initial fused data; wherein the device operation data at least includes the device temperature sequence, the charge and discharge current waveform and the voltage fluctuation characteristics acquired in real time; The initial fusion data is subjected to noise filtering and data normalization processing to generate target multi-dimensional operation data.

3. The method according to claim 2, characterized in that The pre-trained AI prediction model generates the running status prediction result in the following manner: Dividing the target multi-dimensional operation data into data segments within a continuous time window, and extracting the temporal dependency and cross-dimensional correlation features in each data segment; By using the deep convolutional network layer in the AI ​​prediction model, the temporal dependency is locally enhanced to generate an enhanced spatiotemporal feature vector; Using the attention mechanism layer in the AI ​​prediction model, weights are assigned to the cross-dimensional correlation features to screen out a subset of key features related to device health; Based on the comparative analysis of the key feature subset and the historical fault data, the equipment health score and the potential abnormal feature set are output.

4. The method according to claim 3, characterized in that The analyzing the correlation between the potential abnormal feature set and the preset fault mode library to determine the maintenance urgency index of the distributed energy storage device includes: Extracting abnormal waveform patterns, temperature mutation intervals and current distortion features from the potential abnormal feature set; Performing morphological matching between the abnormal waveform pattern and the known fault waveforms in the preset fault pattern library, and calculating a waveform similarity index; Determining a temperature-associated risk value according to an overlap ratio between the temperature mutation interval and a historical fault temperature threshold; Combining the distribution density of the current distortion characteristics in the time dimension, generating a comprehensive anomaly confidence; The maintenance urgency index is determined based on a weighted fusion result of the waveform similarity index, the temperature-associated risk value and the comprehensive anomaly confidence.

5. The method according to claim 4, characterized in that The generating of a predictive maintenance instruction for the distributed energy storage device according to a comparison result between the maintenance urgency index and a preset maintenance threshold comprises: According to the numerical range of the maintenance urgency indicator, a corresponding maintenance operation type is matched from a preset maintenance strategy table, wherein the maintenance operation type includes battery unit replacement, connector tightening, and cooling system cleaning; Predicting the remaining safe operating time of the distributed energy storage device based on the device health score decrease rate output by the AI ​​prediction model; Dynamically adjust the start time and duration of the maintenance time window according to the remaining safe operation time and the standard working hours required for equipment maintenance; The maintenance operation type and the maintenance time window are encapsulated as the predictive maintenance instruction and sent to the target maintenance terminal.

6. The method according to claim 1, characterized in that The method further comprises: After generating the predictive maintenance instruction, monitoring the actual operation response data of the distributed energy storage device in real time; Extracting maintenance effect features from the actual operation response data, the maintenance effect features including equipment temperature drop rate, current waveform smoothness, and voltage stability improvement index; Feeding back the maintenance effect characteristics to the AI ​​prediction model, and updating the parameter weights of the AI ​​prediction model through an incremental learning algorithm; Based on the updated AI prediction model, iterative prediction is performed on the multi-dimensional operation data collected during the target period to optimize the accuracy of the maintenance urgency indicator; The updating of the parameter weights of the AI ​​prediction model by an incremental learning algorithm includes: Extracting positive feedback samples and negative feedback samples from the maintenance effect features, the positive feedback samples representing that the device state after the maintenance operation meets the set conditions, and the negative feedback samples representing that the device state after the maintenance operation does not meet the set conditions; Calculating the difference in feature distribution between the positive feedback sample and the negative feedback sample to generate a model weight adjustment direction vector; Using an online gradient descent algorithm, fine-tuning the convolution kernel parameters and attention weights of the AI ​​prediction model according to the weight adjustment direction vector; During the fine-tuning process, an elastic weight solidification mechanism is added to lock the model parameters associated with historical failure modes and update the local parameters associated with the newly added maintenance effect characteristics.

7. The method according to claim 1, characterized in that The method further comprises: Deploy edge computing nodes in the distributed energy storage device cluster, wherein the edge computing nodes are used to perform local data preprocessing and model reasoning; Performing preliminary anomaly detection on the collected multi-dimensional operation data through the edge computing node to filter out active abnormal data fragments; Uploading the active abnormal data fragments to a cloud server, and adopting a federated learning strategy in the cloud server to aggregate local model updates of multiple edge computing nodes to generate a globally optimized AI prediction model; Distributing the globally optimized AI prediction model to each edge computing node to achieve distributed model co-evolution; The execution process of the federated learning strategy includes: Allocate an independent model training copy to each edge computing node and train it locally using active anomaly data fragments; Collect model gradient updates from each edge computing node, and use differential privacy technology to add random noise to the gradient updates; In the cloud server, weighted averaging is performed on the noisy gradient update amount to generate a global gradient update direction label; Update the cloud backbone model parameters according to the global gradient update direction label, and synchronize the updated cloud backbone model parameters to each edge computing node; Model distillation technology is used in the synchronization process to compress the knowledge of the cloud backbone model into a lightweight model on the edge side to maintain the reasoning efficiency under the constraints of edge computing resources.

8. The method according to claim 1, characterized in that The method further comprises: Generating a digital twin of the distributed energy storage device, wherein the digital twin is used to map the operating status and environmental parameters of the physical device in real time; Configuring simulated fault scenario features in the digital twin to generate an enhanced set of training data containing multiple fault modes; Using the training data enhancement set to perform adversarial training on the AI ​​prediction model; wherein the adversarial training is used to improve the generalization ability of the AI ​​prediction model for unknown fault modes; Deploy the AI ​​prediction model after adversarial training to the actual operation and maintenance environment, and compare the prediction results of the digital twin with the actual status of the physical equipment in real time to determine the model prediction deviation; When the model prediction deviation exceeds the fault tolerance threshold, the model rollback mechanism is triggered to switch to the AI ​​prediction model of the previous iteration stage.

9. A distributed energy storage device predictive maintenance system, characterized in that: It includes a processor and a memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call the program instructions in the memory to execute the AI-based distributed energy storage equipment predictive maintenance method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the AI-based predictive maintenance method for distributed energy storage equipment described in any one of claims 1-8 is implemented.

Images