AI-Based Predictive Maintenance System and Method for Distributed Energy Storage Devices

Abstract

The embodiments of the present application relate to the field of computer technology. Specifically, it relates to a predictive maintenance system and method for distributed energy storage devices based on AI. This method realizes precise monitoring of the operating status of distributed energy storage devices through the comprehensive collection of multi-dimensional operation data and in-depth analysis of the AI prediction model. Specifically, this method can effectively identify potential abnormal characteristics of the device, combine the quantitative evaluation of the maintenance urgency index, generate targeted predictive maintenance instructions, significantly improve the timeliness and accuracy of device maintenance, reduce the risk of unplanned downtime, extend the service life of the device, and at the same time optimize the efficiency of maintenance resource allocation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular, to an AI-based predictive maintenance system and method for distributed energy storage devices. Background Art

[0002] In the field of state monitoring technology for existing distributed energy storage devices, traditional methods mainly rely on single-dimensional data analysis and simple alarm mechanisms based on thresholds. For example, discrete temperature monitoring points are often used in combination with voltage overlimit detection, and alarms are triggered by setting fixed thresholds. This method is difficult to capture weak correlation features during the device degradation process.

[0003] In addition, the input feature dimensions of traditional fault classification methods based on support vector machines are limited, and they cannot effectively fuse temperature gradients and high-frequency current waveform features with spatio-temporal distributions, resulting in insufficient sensitivity for detecting early subtle anomalies. Moreover, traditional maintenance decision-making systems mostly use static maintenance strategy tables. For example, planned maintenance is carried out by accumulating the device operation time. This time-driven mode is difficult to adapt to the dynamic degradation process under complex working conditions.

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

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

[0006] An embodiment of this application provides an AI-based predictive maintenance method for distributed energy storage devices, which is applied to a predictive maintenance system for distributed energy storage devices. The method includes:

[0007] Collect multi-dimensional operation data of the distributed energy storage device, where the multi-dimensional operation data includes device temperature sequences, charge and discharge current waveforms, and voltage fluctuation characteristics;

[0008] Input the multi-dimensional operation data into a pre-trained AI prediction model to generate a prediction result of the operation state of the distributed energy storage device. The prediction result of the operation state includes a device health score and a set of potential abnormal features;

[0009] Based on the prediction result of the operation state, analyze the relevance between the set of potential abnormal features and a preset fault mode library to determine the maintenance urgency index of the distributed energy storage device;

[0010] Generate a predictive maintenance instruction for the distributed energy storage device according to the comparison result between the maintenance urgency index and a preset maintenance threshold. The predictive maintenance instruction includes a maintenance time window and a maintenance operation type.

[0011] Applying the embodiments of the present application, the method realizes precise monitoring of the operating status of distributed energy storage devices through the comprehensive collection of multi-dimensional operation data and the in-depth analysis of the AI prediction model. This method can effectively identify potential abnormal features of the device, generate targeted predictive maintenance instructions by combining the quantitative evaluation of the maintenance urgency index, significantly improve the timeliness and accuracy of device maintenance, reduce the risk of unplanned downtime, extend the service life of the device, and at the same time optimize the efficiency of maintenance resource allocation.

[0012] As an alternative embodiment, the multi-dimensional operation data further includes an environmental parameter set, and the environmental parameter set includes the humidity change curve of the device deployment location, the environmental temperature gradient, and the vibration frequency spectrum; the collection of the multi-dimensional operation data of the distributed energy storage device includes:

[0013] Through the sensor array deployed on the distributed energy storage device, the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics are obtained in real time;

[0014] Simultaneously collect the humidity change curve, the environmental 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 the time stamp alignment manner to obtain initial fusion data; wherein, the device operation data at least includes the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics obtained in real time;

[0015] Perform noise filtering and data normalization processing on the initial fusion data to generate target multi-dimensional operation data.

[0016] Applying the embodiments of the present application, the method generates a multi-dimensional feature collaborative analysis system through the spatio-temporal fusion of the environmental parameter set and the device operation data. In this way, the influence characteristics of environmental factors on device operation can be effectively captured, the context awareness ability of anomaly detection can be enhanced, the sensor noise interference can be eliminated through data standardization processing, the quality of the input data for subsequent model analysis can be improved, and a data basis for device status evaluation in a complex environment can be provided.

[0017] As an alternative embodiment, the pre-trained AI prediction model generates the operation status prediction result in the following manner:

[0018] Divide the target multi-dimensional operation data into data segments within continuous time windows, and extract the temporal dependence relationship and cross-dimensional association features in each data segment;

[0019] Through the deep convolutional network layer in the AI prediction model, locally enhance the temporal dependence relationship to generate an enhanced spatio-temporal feature vector;

[0020] Using the attention mechanism layer in the AI prediction model, weight distribution is performed on the cross-dimensional correlation features to screen out a key feature subset related to the device health;

[0021] Based on the comparative analysis of the key feature subset and historical fault data, the device health score and the set of potential abnormal features are output.

[0022] Applying the embodiments of the present application, the method adopts a hybrid model architecture combining a deep convolutional network and an attention mechanism, realizes the collaborative feature mining of multi-source heterogeneous data, effectively extracts weak correlation features in the device degradation process through spatio-temporal feature enhancement and key feature screening, improves the physical interpretability of the health score calculation, provides a multi-dimensional decision-making basis for device state prediction, and significantly improves the early fault identification ability.

[0023] As an optional embodiment, the parsing of the relevance between the set of potential abnormal features and a preset fault mode library to determine the maintenance urgency index of the distributed energy storage device includes:

[0024] Extracting abnormal waveform patterns, temperature mutation intervals, and current distortion features from the set of potential abnormal features;

[0025] Performing morphological matching between the abnormal waveform patterns and known fault waveforms in the preset fault mode library, and calculating a waveform similarity index;

[0026] Determining a temperature correlation risk value according to the overlapping ratio between the temperature mutation interval and the historical fault temperature threshold;

[0027] Combining the distribution density of the current distortion features in the time dimension to generate a comprehensive abnormal confidence level;

[0028] Based on the weighted fusion result of the waveform similarity index, temperature correlation risk value, and comprehensive abnormal confidence level, the maintenance urgency index is determined.

[0029] Applying the embodiments of the present application, the method establishes a multi-dimensional abnormal feature fusion evaluation system, realizes the accurate quantification of the maintenance urgency through a triple verification mechanism of waveform morphological 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 fault occurrence and the severity of consequences, provides a scientific basis for maintenance priority decision-making, and avoids the situations of over-maintenance or under-maintenance.

[0030] As an optional embodiment, the generating of a predictive maintenance instruction for the distributed energy storage device according to the comparison result between the maintenance urgency index and a preset maintenance threshold includes:

[0031] Match the corresponding maintenance operation type from a preset maintenance strategy table according to the numerical range where the maintenance urgency index is located. The maintenance operation types include battery cell replacement, connector fastening, and heat dissipation system cleaning.

[0032] Predict the remaining safe operation duration of the distributed energy storage device based on the rate of decline of the device health score output by the AI prediction model.

[0033] Dynamically adjust the start time and duration of the maintenance time window according to the remaining safe operation duration and the standard working hours required for device maintenance.

[0034] Package the maintenance operation type and the maintenance time window as the predictive maintenance instruction and send it to the target maintenance terminal.

[0035] Applying the embodiment of the present application, the method innovatively combines the device health decline trend with the operation and maintenance resource scheduling to generate a dynamic maintenance time planning model. Through the predictive calculation of the remaining safe operation duration, the method realizes the intelligent optimization and arrangement of the maintenance window, effectively coordinates the device operation requirements and the maintenance operation requirements, minimizes the impact of maintenance operations on system availability, and improves the comprehensive operation efficiency of the energy storage system.

[0036] As an optional embodiment, the method further includes:

[0037] After generating the predictive maintenance instruction, real-time monitor the actual operation response data of the distributed energy storage device.

[0038] Extract the maintenance effect features from the actual operation response data. The maintenance effect features include the device temperature decline rate, the smoothness of the current waveform, and the voltage stability improvement index.

[0039] Feed back the maintenance effect features to the AI prediction model and update the parameter weights of the AI prediction model through the incremental learning algorithm.

[0040] Based on the updated AI prediction model, perform iterative prediction on the multi-dimensional operation data collected during the target period to optimize the accuracy of the maintenance urgency index.

[0041] Applying the embodiment of the present application, the method generates a complete model closed-loop optimization mechanism. Through the synergistic effect of maintenance effect feature feedback and incremental learning, the continuous evolution of the prediction model is realized. The method effectively solves the model drift problem caused by device aging, maintains the adaptability of the prediction system to device state changes, forms a virtuous cycle of "monitoring - maintenance - optimization", and significantly improves the long-term operation reliability of the system.

[0042] As an alternative embodiment, updating the parameter weights of the AI prediction model through the incremental learning algorithm includes:

[0043] Extract positive feedback samples and negative feedback samples from the maintenance effect features. The positive feedback samples represent that the device state conforms to the set conditions after the maintenance operation, and the negative feedback samples represent that the device state does not conform to the set conditions after the maintenance operation;

[0044] Calculate the feature distribution difference between the positive feedback samples and the negative feedback samples to generate a model weight adjustment direction vector;

[0045] Adopt the online gradient descent algorithm to finely tune the convolution kernel parameters and attention weights of the AI prediction model according to the weight adjustment direction vector;

[0046] Add an elastic weight consolidation mechanism during the fine-tuning process to lock the model parameters associated with historical failure modes and update the local parameters associated with the newly added maintenance effect features.

[0047] Applying the embodiment of the present application, the method adopts an elastic weight consolidation mechanism to achieve a balance between knowledge protection and adaptive learning during the model incremental update process. This method effectively prevents the forgetting of important historical failure features through feature distribution difference analysis and parameter selective update, while quickly adapting to new failure modes, ensuring the stability of the model iteration process, and providing technical support for the sustainable optimization of the prediction system.

[0048] As an alternative embodiment, the method further includes:

[0049] Deploy edge computing nodes in the distributed energy storage device cluster. The edge computing nodes are used to perform local data preprocessing and model inference;

[0050] Perform preliminary anomaly detection on the collected multi-dimensional operation data through the edge computing nodes to screen out active anomaly data segments;

[0051] Upload the active anomaly data segments to the cloud server, and adopt a federated learning strategy in the cloud server to aggregate the local model updates of multiple edge computing nodes to generate a globally optimized AI prediction model;

[0052] Distribute the globally optimized AI prediction model to each edge computing node to achieve distributed model co-evolution.

[0053] Applying the embodiments of the present application, the method generates a distributed intelligent analysis network through an edge-cloud collaborative computing architecture. This method uses edge computing to achieve real-time preprocessing and preliminary diagnosis of local data, combines federated learning to achieve global knowledge sharing, improves the generalization ability of the model while ensuring data privacy, significantly reduces the communication load of centralized computing, and enhances the adaptability of the system to large-scale deployment.

[0054] As an alternative embodiment, the execution process of the federated learning strategy includes:

[0055] Allocate independent model training copies to each edge computing node and perform training locally using active abnormal data segments;

[0056] Collect the model gradient update amounts from each edge computing node and add random noise to the gradient update amounts using differential privacy technology;

[0057] Perform weighted averaging on the noise-added gradient update amounts in the cloud server to generate a global gradient update direction label;

[0058] 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;

[0059] During the synchronization process, use model distillation technology to compress the knowledge of the cloud backbone model into the lightweight model on the edge side to maintain the inference efficiency under the constraints of edge computing resources.

[0060] Applying the embodiments of the present application, the method innovatively integrates differential privacy and model distillation technologies to generate a secure and efficient federated learning framework. This method ensures data privacy through gradient noise injection, combines knowledge distillation to achieve model lightweighting, while maintaining the inference efficiency on the edge side, ensures the knowledge fusion quality of the global model, and provides reliable technical support for the swarm intelligence evolution of distributed energy storage systems.

[0061] As an alternative embodiment, the method further includes:

[0062] Generate a digital twin of the distributed energy storage device, which is used to map the operating status and environmental parameters of the physical device in real time;

[0063] Configure simulated fault scenario features in the digital twin to generate a training data augmentation set containing multiple fault modes;

[0064] Use the training data augmentation 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 to unknown fault modes;

[0065] 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 state of the physical device in real time to determine the model prediction deviation.

[0066] When the model prediction deviation exceeds the fault tolerance threshold, trigger the model rollback mechanism and switch to the AI prediction model in the previous iteration stage.

[0067] Applying the embodiments of the present application, the method generates a virtual-real fusion model training environment through the collaborative application of digital twins and adversarial training. This method uses high-fidelity simulation data to enhance the model's ability to identify unknown faults, combines real-time deviation detection and model rollback mechanisms, effectively improves the robustness of the prediction system, provides a safety redundancy guarantee for equipment status monitoring under extreme working conditions, and significantly reduces the risk of operation and maintenance decisions.

[0068] The embodiments of the present application also provide a predictive maintenance system for distributed energy storage devices, including a processor, a memory, and a bus connected to the processor; wherein, the processor and the memory complete communication with each other through the bus; the processor is used to call program instructions in the memory to execute the above-mentioned AI-based predictive maintenance method for distributed energy storage devices.

[0069] The embodiments of the present application also provide a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, it implements the above-mentioned AI-based predictive maintenance method for distributed energy storage devices.

[0070] A predictive maintenance system and method for distributed energy storage devices based on AI provided by the embodiments of the present application realizes precise monitoring of the operating state of distributed energy storage devices through the comprehensive collection of multi-dimensional operation data and the in-depth analysis of the AI prediction model. Specifically, this method can effectively identify potential abnormal characteristics of the device, combine the quantitative evaluation of the maintenance urgency index, generate targeted predictive maintenance instructions, significantly improve the timeliness and accuracy of device maintenance, reduce the risk of unplanned downtime, extend the service life of the device, and at the same time optimize the allocation efficiency of maintenance resources. Description of the Drawings

[0071] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0072] Figure 1 It is a flowchart of an AI-based predictive maintenance method for distributed energy storage devices provided by the embodiments of the present application.

[0073] Figure 2 This is a block diagram of a predictive maintenance system for a distributed energy storage device provided by an embodiment of the present application.

[0074] Icon:

[0075] 100 - Predictive maintenance system for distributed energy storage device;

[0076] 101 - Processor; 102 - Memory; 103 - Bus. Detailed implementation manners

[0077] 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 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 so that the present application can be more thoroughly understood and the scope of the present application can be fully conveyed to those skilled in the art.

[0078] To better understand the above technical solutions, the technical solutions of the present application will be 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 solutions of the present application, rather than limitations on the technical solutions of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0079] Figure 1 This is a flowchart of an AI-based predictive maintenance method for a distributed energy storage device according to an embodiment of the present application, which is applied to a predictive maintenance system for a distributed energy storage device and includes steps 110 - 140.

[0080] Step 110: Collect multi-dimensional operation data of the distributed energy storage device, where the multi-dimensional operation data includes a device temperature sequence, a charge and discharge current waveform, and a voltage fluctuation characteristic.

[0081] In an embodiment of the present application, a temperature sensor array can be deployed at key nodes of the distributed energy storage device. The temperature sensor array is composed of twelve PT1000 platinum resistance temperature sensors and is installed on the surface of the battery module in a three-dimensional space grid layout manner. 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 100 kHz is configured on the DC bus side to capture the current waveform details during the charge and discharge process at a sampling frequency of 100,000 times per second. The current waveform details include complete harmonic components and transient response characteristics.

[0082] In terms of voltage monitoring, an analog-to-digital converter with 16-bit precision is used to synchronously collect the voltages of each branch of the energy storage system. Voltage fluctuation characteristic parameters are extracted through the calculation of the sliding window variance, including the peak-to-peak fluctuation range and the standard deviation statistic 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 a high-frequency ripple component with an amplitude reaching 15% of the rated value at the end of charging, and the voltage monitoring system captures an abnormal fluctuation with a cell voltage range exceeding 350 mV during the corresponding period.

[0083] Step 120: Input the multi-dimensional operation data into a pre-trained AI prediction model to generate a prediction result of the operation state of the distributed energy storage device. The operation state prediction result includes a device health score and a set of potential abnormal features.

[0084] Exemplarily, when inputting the above multi-dimensional operation data into the pre-trained AI prediction model, a deep neural network architecture with the ability to extract spatio-temporal features needs to be generated. This 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 spectrogram, and fuses the voltage fluctuation feature vectors through an attention mechanism.

[0085] For another example, during the training process of the AI prediction model, a sample set containing 100,000 groups of historical operation data is used, covering 23 exemplary state modes such as normal operating conditions, battery aging, and loose connectors. When inputting the current operation data, the output layer of the AI prediction model will generate a prediction result including a device health score and a set of potential abnormal features. For example, in a certain actual case, after the model extracts the features of the input data of a certain energy storage container, it identifies features such as an abnormal temperature gradient of the positive electrode post (the temperature difference between adjacent sensors reaches 7.2°C) and an excessive content of the third harmonic of the charging current (THD = 8.7%). Finally, it outputs a device health score of 62 points (on a 100-point scale) and marks two potential abnormal features: electrolyte crystallization risk and current collector corrosion.

[0086] Step 130: Based on the operation state prediction result, analyze the relevance between the set of potential abnormal features and a preset fault mode library to determine the maintenance urgency index of the distributed energy storage device.

[0087] Specifically, when analyzing the relevance between potential abnormal features and preset fault modes, an expert knowledge base containing 48 exemplary fault modes needs to be established. This expert knowledge base uses an ontology modeling method to decompose each fault mode into a multi-dimensional parameter combination in the feature vector space and sets a dynamic matching threshold.

[0088] For example, by calculating the cosine similarity between the current abnormal feature set and the feature vectors of each fault mode, and combining with the analysis of the time-series evolution trend, the maintenance urgency index is determined. Another example is that when it is detected that a certain battery module simultaneously has an abnormal temperature rise rate (0.8 °C / min), an early charging end voltage platform (appearing 120 seconds earlier than the standard curve), and an AC impedance spectrum phase angle shift (a 12% shift at the 45-degree frequency point), it is matched with the "abnormal proliferation of the negative SEI film" mode in the knowledge base, and the maintenance urgency index 0.83 (range 0-1) is calculated. This index comprehensively considers multi-dimensional evaluation results such as the probability of failure occurrence (72%), the severity level of consequences (second level), and the prediction of the remaining service life of the equipment (estimated remaining 83 cycles).

[0089] Step 140: Generate a predictive maintenance instruction for the distributed energy storage device according to the comparison result between the maintenance urgency index and a preset maintenance threshold. The predictive maintenance instruction includes a maintenance time window and a maintenance operation type.

[0090] For example, when generating a predictive maintenance instruction, the system maintenance decision module can compare and analyze the maintenance urgency index with a preset three-level threshold system. This three-level threshold system is dynamically adjusted according to the device type, operating environment, and availability of maintenance resources. Exemplarily, it is set as follows: a red warning is triggered when the urgency is above 0.7 (to be disposed of within 24 hours), an orange warning when it is between 0.5 and 0.7 (to be disposed of within 72 hours), and a yellow warning when it is between 0.3 and 0.5 (to be arranged in the weekly maintenance plan).

[0091] Another example is that the optimal maintenance time window can be calculated using a constraint satisfaction algorithm by combining the device operation calendar, grid load prediction 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 operation plan of the power station for the next three days, avoids the morning peak load period (09:00-11:00), selects the low-load window from 02:00 to 04:00 the next day, specifies the battery module number to be replaced (B3-12-07) and the list of required special tools (including torque wrenches, thermal imagers, etc.), and at the same time generates a detailed work order instruction including operation steps, safety precautions, and technical parameters.

[0092] In summary, the above method realizes the accurate monitoring of the operation status of distributed energy storage devices through the comprehensive collection of multi-dimensional operation data and the in-depth analysis of the AI prediction model. Specifically, this method can effectively identify potential abnormal features of the device, generate targeted predictive maintenance instructions by combining the quantitative evaluation of the maintenance urgency index, significantly improve the timeliness and accuracy of device maintenance, reduce the risk of unplanned downtime, extend the service life of the device, and at the same time optimize the allocation efficiency of maintenance resources.

[0093] In the embodiments of the present application, 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. On this basis, the acquisition of the multi-dimensional operation data of the distributed energy storage device described in step 110 includes:

[0094] Step 111: Real-time obtain the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics through a sensor array deployed on the distributed energy storage device.

[0095] 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, and the temperature values at each point are recorded at a sampling frequency of twice per second to generate a device temperature sequence with spatial resolution. A closed-loop current sensor based on the fluxgate principle is installed at the DC side busbar, and its measurement bandwidth is extended to 200 kHz to capture the nanosecond-level transient distortion in the charge and discharge current waveform at a sampling rate of two hundred thousand times per second, and the rising edge overshoot and ripple attenuation characteristics of the current waveform are completely recorded.

[0096] In the voltage monitoring link, a twenty-four-bit high-precision differential ADC module is used to synchronously sample the voltages of each branch of the energy storage system, and the voltage fluctuation spectrum characteristics are extracted through a sliding window fast Fourier transform, specifically including the fundamental wave amplitude offset and the third harmonic content change rate with a five-minute period. For example, in the actual operation of a certain distributed energy storage station, when performing fast frequency modulation response, the temperature sensor array detects an abnormal distribution with a maximum temperature difference of 9.3 °C between adjacent modules in the fourth battery compartment, the current sensor synchronously records a current spike pulse with a duration of 120 ms during the discharge process, and the voltage monitoring system finds that the total harmonic distortion rate of the AC side voltage increases to 6.8% during the corresponding period.

[0097] Step 112: Synchronously collect the humidity change curve, the environmental 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 fusion data; where the device operation data at least includes the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics obtained in real time.

[0098] For example, deploy a humidity sensor array based on the principle of polymer capacitance on the six azimuth planes inside the equipment cabin, record the relative humidity values at a sampling interval of once per minute, and generate a humidity change curve with spatial distribution characteristics. Arrange eight groups of PT100 platinum resistance ambient temperature probes within a two-meter range around the equipment to form a three-dimensional temperature gradient monitoring network, and update the ambient temperature distribution matrix every 30 seconds. Vibration monitoring uses a three-axis MEMS accelerometer array to obtain the equipment foundation vibration signal at a sampling rate of 5,000 times per second, and extract the vibration frequency spectrum characteristics in the 0-2 kHz frequency band through wavelet packet decomposition.

[0099] Exemplarily, when performing data fusion, the IEEE 1588 Precision Time Protocol is used to align the timestamps of each sensor, 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 linearly increased 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 reached 15°C, and the vibration frequency spectrum analysis found that the fundamental resonance frequency shifted to 23.5 Hz and was accompanied by an abnormal aggregation of the second harmonic energy. These environmental parameters were spatio-temporally correlated with the abnormal phenomenon that the terminal voltage of the battery module dropped by 0.15V during the same period.

[0100] Step 113: Perform noise filtering and data normalization processing on the initial fusion data to generate target multi-dimensional operation data.

[0101] In step 113, an improved Savitzky-Golay filter can be used to smooth the temperature sequence, with the window width set to 21 sampling points and the polynomial order to third order, effectively suppressing the ±0.3°C random fluctuations caused by sensor thermal noise. Implement a wavelet threshold-based denoising algorithm for the current waveform data, select the sym8 wavelet basis for five-layer decomposition, and determine the thresholds at each scale through Stein's unbiased risk estimator to eliminate the pseudo-zero-crossing distortion caused by high-frequency electromagnetic interference.

[0102] Exemplarily, the voltage fluctuation feature processing uses the Z-score normalization method to calculate the mean and standard deviation of each feature parameter based on the data of the past 24 hours, and normalize the five-minute voltage range to the [-1, 1] interval. For example, in a data processing example of a frequency modulation type energy storage system, after the original current waveform is denoised, the high-frequency noise above 200 kHz caused by the converter switching action is effectively filtered, the substantial harmonic components with amplitudes exceeding 8% of the nominal value are retained, 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 series alignment of multi-dimensional data.

[0103] As an alternative embodiment, the pre-trained AI prediction model generates the operating state prediction result in the following manner:

[0104] Step 210: Divide the target multi-dimensional operating data into data segments within consecutive time windows, and extract the temporal dependence relationship and cross-dimensional correlation features in each data segment.

[0105] When dividing the target multi-dimensional operating data into data segments within consecutive time windows in Step 210, a sliding window mechanism is adopted to set a basic analysis window of sixty minutes, with an overlapping rate of fifty percent between adjacent windows. Each data segment contains 3,600 temperature sampling points, 7.2 million current waveform sampling points, and 360 groups of voltage characteristic parameters. The extraction of the temporal dependence relationship uses the dynamic time warping algorithm to calculate the phase delay between different sensor signals, and establish a time-delay correlation model between the temperature gradient change rate and the current harmonic content.

[0106] For example, the cross-dimensional correlation feature mining is realized through exemplary correlation analysis, and the coupling coefficient between the environmental vibration spectrum energy distribution and the battery internal resistance change is calculated. For example, in an analysis case of a certain energy storage system for cascade utilization, it is found through time window division that when the environmental temperature gradient at the end of charging increases by 1 °C / m, the temperature difference expansion coefficient of the battery module increases by 0.15 correspondingly, and the Pearson correlation coefficient between the third harmonic content of the current waveform and the fundamental frequency energy of the environmental vibration is 0.82.

[0107] Step 220: Through the deep convolutional network layer in the AI prediction model, locally enhance the temporal dependence relationship to generate an enhanced spatio-temporal feature vector.

[0108] When locally enhancing the temporal dependence relationship through the deep convolutional network layer in Step 220, a dilated convolutional structure with residual connections is generated. The convolution kernel size is set to seven, and the dilation coefficient is three, expanding the receptive field to 300 sampling points while maintaining the temporal resolution. The generation of the spatio-temporal feature vector adopts three-dimensional convolution operation, stacking the temperature spatial distribution matrix, the current waveform time-frequency spectrum diagram, and the voltage fluctuation feature vector as multi-channel inputs, and extracting cross-scale spatio-temporal features through sixteen layers of deep convolution.

[0109] For example, in a certain battery aging identification task, the convolutional network layer successfully captures the non-linear relationship between the rising rate of the positive electrode temperature and the relaxation time of the negative electrode voltage during charging, and compresses the original feature dimension from 256 dimensions to a 64-dimensional feature vector with physical significance.

[0110] Step 230: Use the attention mechanism layer in the AI prediction model to assign weights to the cross-dimensional correlation features, and screen out the key feature subset related to the device health.

[0111] When performing weight allocation using the attention mechanism layer in step 230, an eight-head self-attention structure is adopted to set up eight parallel attention heads, and each head is responsible for processing the feature interaction relationships at different time scales. The screening of the key feature subset is achieved by calculating the feature importance scores, and a threshold is set to retain the feature dimensions with scores higher than the 75th percentile. For example, in a certain over-temperature warning case, the attention mechanism layer assigns a weight coefficient of 0.62 to the feature of the environmental humidity change rate, which is significantly higher than the weight of the environmental temperature gradient of 0.34, accurately identifying the key association that a high-humidity environment accelerates the risk of thermal runaway.

[0112] Step 240: Based on the comparative analysis of the key feature subset and the historical fault data, output the device health score and the set of potential abnormal features.

[0113] When performing the comparative analysis based on the key feature subset and the historical fault data in step 240, a comparison 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 features and the historical fault patterns. The calculation of the device health score adopts a fuzzy logic inference system, and 72 fuzzy rules are set to map the feature deviation degree to the score range of 0 - 100 points. For example, in the prediction of a certain lithium plating fault in a battery, the system compares the current features with the spectral features of 300 historical lithium plating faults, calculates the Mahalanobis distance to be 2.35, triggers the health score to drop to 55 points, and marks two potential abnormal features: the abnormal negative electrode voltage platform and the shortening of the temperature relaxation time.

[0114] As an optional embodiment, the analysis of the relevance between the set of potential abnormal features and the preset fault mode library in step 130 to determine the maintenance urgency index of the distributed energy storage device includes:

[0115] Step 131: Extract the abnormal waveform pattern, the temperature mutation interval, and the current distortion feature from the set of potential abnormal features.

[0116] For example, an improved dynamic time warping algorithm can be used to perform morphological analysis on the charge and discharge current waveforms, and abnormal pulse segments with an amplitude exceeding 12% of the nominal value and a duration greater than 50 ms are identified by calculating the waveform curvature change rate. At the same time, the sliding window second-order difference method is used to detect the mutation interval in the temperature sequence with a change rate exceeding 3 °C / min. For the current distortion feature, the original signal is decomposed into eight intrinsic mode functions through the Hilbert-Huang transform, and the energy ratio of the modal components in the frequency range of 150 - 250 Hz is extracted as the distortion feature quantity.

[0117] For another example, in a fault diagnosis case of an energy storage power station, the system detected periodic spike pulses with an amplitude reaching 18% of the rated current at the end of charging. The surface temperature of the corresponding battery module jumped from 32°C to 49°C within five minutes. At the same time, the analysis of the current distortion characteristics showed that the proportion of high-frequency modal energy exceeded 2.3 times the reference value, forming a complete set of abnormal characteristics.

[0118] Step 132: Perform morphological matching between the abnormal waveform pattern and the known fault waveforms in the preset fault pattern library, and calculate the waveform similarity index.

[0119] When performing morphological matching between the abnormal waveform pattern and the preset fault pattern library in Step 132, a feature database containing 56 standard fault waveforms is generated. The dynamic time warping distance is used as the similarity metric, 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 can be calculated. For example, when an oscillation mode with an amplitude modulation depth of 14% and a duration of 80 ms is detected in the charging current waveform, the system calculates that the similarity index between it and the "increased contact resistance" template in the fault library reaches 0.82, exceeding the threshold to trigger the matching confirmation of the corresponding fault type.

[0120] Step 133: Determine the temperature-related risk value according to the overlapping ratio between the temperature mutation interval and the historical fault temperature threshold.

[0121] When determining the temperature-related risk value in Step 133, a historical fault temperature distribution model based on kernel density estimation is established, and the probability density overlapping area between the current temperature mutation interval and the historical fault temperature curve is calculated. For example, the temperature time series can be divided into sampling points at one-minute intervals, and the occurrence frequency of each temperature value within the historical fault temperature range is counted. After smoothing through the Gaussian kernel function, the cumulative probability value is calculated. For another example, a battery module experiences a non-linear temperature rise from 28°C to 53°C within 15 minutes. The system calculates that the overlapping area ratio between this temperature trajectory and the historical data of thermal runaway faults reaches 68%, corresponding to a temperature-related risk value of 0.73.

[0122] Step 134: Generate a comprehensive anomaly confidence level in combination with the distribution density of the current distortion characteristics in the time dimension.

[0123] When generating the comprehensive anomaly confidence level in Step 134, a spatio-temporal density clustering algorithm is used to analyze the time distribution characteristics of the current distortion characteristics, and the occurrence frequency of abnormal events per unit time and the spatial distribution entropy value are calculated. In specific implementation, the 24-hour operation period is divided into windows at five-minute intervals, and the number of times the current distortion characteristics exceed the threshold in each window is counted. The occurrence intensity of abnormal events is estimated through the Poisson distribution model.

[0124] For example, during the eight-hour operation of a certain energy storage system, nineteen events were detected where the third harmonic distortion rate of the current exceeded 7%. After calculation, the comprehensive anomaly confidence level corresponding to the time distribution density was 0.67. Combining this value with the spatial distribution entropy value of 0.54, the final comprehensive anomaly confidence level was 0.62.

[0125] Step 135: Determine the maintenance urgency index based on the weighted fusion result of the waveform similarity index, temperature-related risk value, and comprehensive anomaly confidence level.

[0126] When determining the maintenance urgency index in step 135, the analytic hierarchy process is used to generate a third-order judgment matrix. The weight of the waveform similarity index is set to 0.42, the weight of the temperature-related risk value is 0.38, and the weight of the comprehensive anomaly confidence level is 0.2. The weighted fusion formula is: Maintenance urgency index = waveform similarity index * 0.42 + temperature-related risk value * 0.38 + comprehensive anomaly confidence level * 0.2.

[0127] 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 = 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 an emergency maintenance response mechanism.

[0128] As an optional embodiment, the generating of the predictive maintenance instruction for the distributed energy storage device according to the comparison result between the maintenance urgency index and the preset maintenance threshold in step 140 includes:

[0129] Step 141: According to the numerical interval where the maintenance urgency index is located, match the corresponding maintenance operation type from the preset maintenance strategy table. The maintenance operation types include battery unit replacement, connector tightening, and heat dissipation system cleaning.

[0130] When matching the maintenance operation type in step 141, the preset maintenance strategy table includes a three-level response mechanism: when the maintenance urgency index is between 0.7 and 1, match the battery unit replacement operation; when it is between 0.5 and 0.7, match the connector tightening operation; when it is between 0.3 and 0.5, match the heat dissipation system cleaning operation. In specific implementation, the system maps the numerical interval to specific maintenance items through a lookup table. For example, a maintenance urgency index of 0.76 corresponds to the battery unit replacement operation, and at the same time, the device topology structure is retrieved to determine the module number to be replaced. In a certain frequency modulation type energy storage system case, the system automatically generates an operation instruction to replace the seventh module in the third battery compartment according to the maintenance urgency index of 0.78, and associates the factory batch and performance parameters of this module.

[0131] Step 142: Predict the remaining safe operation duration of the distributed energy storage device based on the rate of decline of the device health score output by the AI prediction model.

[0132] When predicting the remaining safe operation duration in Step 142, an exponential smoothing method is used to establish a time series prediction model for the device health score. Based on the score data of the most recent twenty-four hours, the moving average of the rate of decline of the score is calculated. For example, the remaining safe operation duration is equal to the absolute value of the current health score minus the safety threshold divided by the rate of decline of the score. For example, if the current health score of a certain energy storage container is 58 points, the rate of decline of the score in the past six hours is 0.75 points per hour, and the set safety threshold is 45 points, then the remaining safe operation duration is calculated as (58 - 45) / 0.75 ≈ 17.3 hours.

[0133] Step 143: Dynamically adjust the start time and duration of the maintenance time window according to the remaining safe operation duration and the standard working hours required for device maintenance.

[0134] 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, and the constraint conditions include the working hour standards of maintenance personnel, the spare part inventory status, and the power grid load prediction data. The genetic algorithm can be used to solve the optimal time window, with a crossover probability of 0.85 and a mutation probability of 0.01, and it converges after fifty generations of iteration. For example, a certain energy storage power station needs to perform a two-hour battery module replacement operation. The system combines the future twenty-four-hour load prediction, avoids the frequency modulation peak period from 9:00 am to 11:00 am, selects the maintenance window from 1:00 am to 3:15 am the next day, and reserves a 15-minute safety buffer time.

[0135] Step 144: Package the maintenance operation type and the maintenance time window into the predictive maintenance instruction and send it to the target maintenance terminal.

[0136] When packaging the predictive maintenance instruction in Step 144, a data structure defined by the ISO 13374 standard is used. The maintenance operation type is encoded as a four-digit operation code, the maintenance time window is accurate to the minute-level time stamp, and the hash value of the additional safety operation specification document is attached. The instruction transmission uses the MQTT protocol to send through a TLS encrypted channel to the maintenance terminal, and each data packet contains a CRC-32 check code. For example, the generated maintenance instruction contains the operation code 0401 (battery unit replacement), the time window from 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, ensuring the integrity and traceability of the instruction.

[0137] As an optional embodiment, the method further includes:

[0138] Step 310: After generating the predictive maintenance instruction, monitor the actual operation response data of the distributed energy storage device in real time.

[0139] In step 310, within the first charge-discharge cycle after the maintenance operation is completed, enable the high-density data acquisition mode, increase the sampling frequency of the temperature sensor to four times per second, expand the bandwidth of the current sensor to 500 kHz, and switch the voltage monitoring system to the twelve-bit precision continuous sampling mode. Through the triaxial vibration sensor array deployed on the device body, capture the mechanical vibration spectrum change after maintenance at a sampling rate of ten thousand times per second.

[0140] For example, after performing a battery unit replacement operation in a certain energy storage power station, the system continuously monitors the surface temperature change of the replaced module B2-07-13, and finds that the highest temperature during charging drops from 67 °C before maintenance to 51 °C. At the same time, the current sensor records that the amplitude of the original 120 Hz ripple component drops from 9.3% of the nominal value to 2.1%. The voltage monitoring data shows that the voltage difference between modules is reduced from 230 mV before maintenance to 45 mV, forming a complete set of actual operation response data.

[0141] Step 320: Extract the maintenance effect features from the actual operation response data, where the maintenance effect features include the device temperature drop rate, current waveform smoothness, and voltage stability improvement index.

[0142] When extracting the maintenance effect features in step 320, use the sliding window analysis method to calculate the device temperature drop rate, set a ten-minute time window, and use the slope value of the temperature change curve fitted by linear regression as a quantitative index. The evaluation of current waveform smoothness uses the discrete Fourier transform to extract the total harmonic distortion rate in the 0-1 kHz 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 extreme value within a five-minute time scale.

[0143] For example, in a case of fastening maintenance of a certain connector, the monitoring data for 24 hours after maintenance shows that the temperature drop rate reaches 0.8 °C / min, the total harmonic distortion rate of the current waveform drops from 6.7% to 1.9%, and the voltage standard deviation improves from 12.3 mV to 3.1 mV, forming a feature vector containing three dimensions and twelve parameters.

[0144] Step 330: Feed the maintenance effect features back to the AI prediction model, and update the parameter weights of the AI prediction model through the incremental learning algorithm; based on the updated AI prediction model, perform iterative prediction on the multi-dimensional operation data collected during the target period to optimize the accuracy of the maintenance urgency index.

[0145] When the maintenance effect features are fed back to the AI prediction model in step 330, an incremental learning framework is generated, the model update period is set to 24 hours, and the operation data within 72 hours after maintenance is injected each time for update. The incremental learning algorithm adopts an elastic weight consolidation mechanism to impose L2 regularization constraints on the parameters of the bidirectional long short-term memory network layer in the model, restricting the parameter update amplitude not to exceed 15% of the original value.

[0146] For example, in a certain model update, the system injects 300 sets of post-maintenance data samples, adjusts the weights of 16 convolutional kernels in the two-dimensional convolutional neural network layer through the online gradient descent algorithm, and simultaneously locks the eight attention mechanism parameters associated with historical thermal runaway faults, increasing the prediction accuracy of the model for connector loosening faults from 83.2% to 89.7%.

[0147] As an optional embodiment, updating the parameter weights of the AI prediction model through the incremental learning algorithm in step 330 includes:

[0148] Step 331: Extract positive feedback samples and negative feedback samples from the maintenance effect features. The positive feedback samples represent that the device state meets the set conditions after the maintenance operation, and the negative feedback samples represent that the device state does not meet the set conditions after the maintenance operation.

[0149] When extracting positive feedback samples and negative feedback samples in step 331, set the device temperature drop rate threshold of 0.5 °C / min, the current waveform smoothness improvement rate of 30%, and the voltage stability improvement index of 40% as the determination conditions. The positive feedback samples need to meet all three conditions simultaneously, while the negative feedback samples include any situation where one of the conditions is not met. For example, after a certain battery unit replacement, the monitored temperature drop rate is 0.63 °C / min, the current smoothness is improved by 40%, and the voltage stability is improved by 55%. This sample is marked as positive feedback; while after another heat dissipation system cleaning, only a temperature drop rate of 0.38 °C / min is achieved, which is classified as a negative feedback sample.

[0150] Step 332: Calculate the feature distribution difference between the positive feedback samples and the negative feedback samples to generate a model weight adjustment direction vector; adopt the online gradient descent algorithm to finely adjust the convolutional kernel parameters and attention weights of the AI prediction model according to the weight adjustment direction vector; add an elastic weight consolidation mechanism during the fine-tuning process to lock the model parameters associated with the historical fault modes and update the local parameters associated with the newly added maintenance effect features.

[0151] When adjusting the model weights in step 332, first calculate the KL divergence of the positive and negative feedback samples in the feature space to generate a weight adjustment direction vector with a dimension of 256. Use the stochastic gradient descent algorithm with momentum, set the learning rate to 0.0001 and the momentum coefficient to 0.9, and fine-tune the parameters of the three-dimensional convolutional layer in the AI prediction model. The elastic weight consolidation 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.

[0152] For example, in a certain model update, the system detects that the Mahalanobis distance between the current waveform smoothness feature in the newly added maintenance data and the historical data reaches 3.2, and then adjusts the relevant convolutional kernel parameters by 12%, while keeping the attention weights associated with the electrolyte crystallization fault unchanged, so that the model's recognition rate for the newly emerged contact resistance increase fault is improved by 11.3%.

[0153] As an optional embodiment, the method further includes:

[0154] Step 410: Deploy edge computing nodes in the distributed energy storage device cluster, where the edge computing nodes are used to perform local data preprocessing and model inference; perform preliminary anomaly detection on the collected multi-dimensional operation data through the edge computing nodes, and filter out active anomaly data segments.

[0155] In step 410, install an edge computing hardware unit based on the ARM architecture on top of each energy storage container. This unit is equipped with a quad-core Cortex-A72 processor and 8GB of LPDDR4 memory, and establishes a communication connection with the local sensor array through the CAN bus protocol to ensure that the data acquisition delay is less than twenty milliseconds. The edge computing hardware unit has a real-time operating system built-in and performs the first-layer processing on the multi-dimensional operation data from the temperature sensor array, current sensor module, and voltage monitoring device.

[0156] For example, in a wind farm supporting energy storage system, twelve edge computing nodes are respectively deployed in six energy storage containers. Each node is responsible for processing the operation data of twenty-four battery modules in its own container, eliminates sensor noise through the sliding window mean filtering algorithm, and aligns the timestamps of temperature data (twice per second) with different sampling rates and current waveform data (two hundred thousand times per second) to generate a time-synchronized preprocessed data stream. When detecting abnormal fluctuations in the temperature of the seventh module in the third container, the edge computing node completes data normalization processing locally and filters out a ten-minute data segment containing abnormal features.

[0157] Step 420: Upload the active abnormal data segments to the cloud server, and in the cloud server, adopt a federated learning strategy to aggregate the 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 co-evolution.

[0158] 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 data packet 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 perform homomorphic encryption processing on the long short-term memory network layer gradient matrices 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 ratio and data quality score of each node.

[0159] For example, in a certain-level energy storage monitoring system, the cloud server aggregates the model update data of six regional substations. Among them, the eastern substation contributes 40,000 groups of battery aging samples, and the western substation provides 20,000 groups of environmental mutation samples. After calculation by the federated average algorithm, the accuracy of the global model for temperature gradient anomaly detection is increased to 92.7%. The updated globally optimized AI prediction model generates a model parameter difference packet through a differential privacy protection mechanism, signs it using the elliptic curve encryption algorithm, and is synchronized to all edge computing nodes via a content delivery network to ensure that the full-node model iteration is completed within fifteen minutes. In a specific implementation case, after the federated learning update of a certain energy storage cluster, the detection response time of the edge node for the increased contact resistance fault is shortened from eight seconds to three seconds, and the false alarm rate is reduced by 40%.

[0160] As an independently implementable technical solution, the deployment of edge computing nodes in the distributed energy storage device cluster in Step 410 includes:

[0161] Step 411: Deploy edge computing hardware units at the neighborhood positions of each physical node in the distributed energy storage device cluster. The edge computing hardware units are connected to the local sensor array through a target latency communication protocol; receive the multi-dimensional operation data through the edge computing hardware units, and perform format standardization and timestamp alignment processing on the multi-dimensional operation data to generate a preprocessed data stream.

[0162] The specific method for deploying the edge computing hardware unit in step 411 is as follows: Install a waterproof and explosion-proof edge computing hardware unit within 0.5 meters of the electrical interface of the battery module string. This unit is equipped with dual gigabit Ethernet interfaces 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, performing dimensional normalization on the voltage signal output by the Hall current sensor, and resampling the voltage fluctuation characteristic data with different sampling rates to a unified time base.

[0163] For example, in a certain industrial and commercial energy storage project, the edge computing hardware unit receives data streams from thirty-two temperature sensors, converts them into the floating-point format of the IEEE 754 standard, and synchronizes the timestamps with the twelve-bit ADC sampling values of the current sensor to generate a fusion data packet with a time resolution of one millisecond.

[0164] Step 412: Load a lightweight AI inference model in the edge computing hardware unit, input the preprocessed data stream into the lightweight AI inference model to generate a preliminary anomaly detection result; the preliminary anomaly detection result includes an anomaly type label and a confidence score; according to the comparison result between the confidence score and a preset anomaly threshold, filter out the initial anomaly data segments that meet the confidence requirement from the preprocessed data stream.

[0165] 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. This model includes a combined structure of two layers of one-dimensional convolutional neural networks and long short-term memory networks. 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 including three anomaly labels: over-temperature warning, current distortion, and voltage imbalance.

[0166] For example, when the input data window shows that the temperature rise rate of a certain battery module reaches 1.8 °C / s within thirty seconds, and at the same time, the content of the third harmonic of the charging current exceeds 12% of the nominal value, the model outputs an over-temperature warning confidence of 0.87 and a current distortion confidence of 0.68, exceeding the preset anomaly threshold of 0.7, triggering this data segment to be marked as an active anomaly data segment. The system automatically extracts the complete data window five minutes before and after the anomaly occurs and generates metadata labels including timestamps, device numbers, and anomaly types.

[0167] Step 413: Perform feature compression encoding on the initial anomaly data segments that meet the confidence requirement to generate a bandwidth transmission data packet; upload the bandwidth transmission data packet to the cloud server through an encrypted channel, triggering the model aggregation process of the federated learning strategy.

[0168] When performing feature compression encoding in step 413, the principal component analysis algorithm is used to reduce the dimension of the initial abnormal data segment from 256 dimensions to 32 dimensions, while retaining 95% of the original information volume. The compressed data packet is appended with an SHA-256 hash check code and encapsulated through the AES-256 encryption algorithm, and then transmitted to the cloud server via the MQTT protocol. For example, after an edge computing node detects a voltage fluctuation anomaly lasting for two minutes, a data stream with 12 million sampling points of original data is compressed into a 12-megabyte feature vector packet, metadata such as the timestamp 2023-11-05T14:23:00 and the device number B3-07-12 is added, and it is uploaded to the cloud storage pool through a TLS 1.3 encrypted channel.

[0169] Step 414: After the globally optimized AI prediction model is generated on the cloud server, receive the downloaded updated model parameter file; load the updated model parameter file into the lightweight AI inference model through the hot update mechanism to overwrite the original model weights.

[0170] 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 inference model. The edge computing hardware unit downloads an 800-kilobyte model update package through the HTTPS protocol and uses the hot patching technology to replace the weight matrix of the bidirectional long short-term memory network layer without interrupting the service. For example, after a certain federated learning aggregation, the updated package sent by the cloud adjusts 16 weight parameters of the current waveform analysis convolution kernel, and the edge computing node completes the model reload within 20 milliseconds. The detection accuracy of the new inference model for the connector loosening fault is increased from 78% to 85%.

[0171] 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 optimized preliminary anomaly detection results; cross-validate the optimized preliminary anomaly detection results with the real-time data stream of the local sensor array, eliminate false alarm anomaly segments and correct the confidence score; dynamically adjust the screening threshold of the initial abnormal data segment that meets the confidence requirement according to the corrected confidence score to achieve adaptive anomaly detection on the edge side.

[0172] When performing iterative inference optimization in step 415, the updated lightweight AI inference model re-analyzes the real-time data stream and performs spatio-temporal correlation verification on the preliminary anomaly detection results and the local vibration sensor data. For example, when the model detects an abnormal amplitude modulation in the charging current waveform, the system synchronously checks the vibration spectrum data for the corresponding period. If no mechanical resonance characteristics are detected, this anomaly is marked as a false alarm and the confidence score is reduced. The dynamic threshold adjustment algorithm adaptively adjusts the confidence threshold for over-temperature warning from 0.7 to 0.65 based on the detection results statistics in the past twenty-four hours to address the problem of increased false positive rate caused by sudden changes in ambient temperature and ensure that the anomaly detection recall rate remains above 95%.

[0173] As an optional embodiment, the execution process of the federated learning strategy includes:

[0174] Step 510: Allocate independent model training copies to each edge computing node and perform training locally using active anomaly data segments.

[0175] In step 510, the specific implementation of allocating independent model training copies to each edge computing node is as follows: Deploy initialized long short-term memory network model copies on twelve edge computing nodes of the distributed energy storage device cluster. Each copy has the same network architecture but an independent parameter space. During local training, use the active anomaly data segments stored by the edge nodes in the past seventy-two hours as the training set. These data segments contain anomaly waveform patterns, temperature mutation intervals, and current distortion characteristics encoded by feature compression.

[0176] For example, in the seventh edge node of a provincial energy storage monitoring system, 325 anomaly data segments are stored. Each segment contains multi-dimensional operation data with a duration of five minutes. During training, the mini-batch gradient descent algorithm is used, with a batch size of 32, a learning rate of 0.001, and iterative training for fifty epochs, enabling the model to accurately identify the abnormal contact resistance pattern caused by salt spray corrosion unique to this area. The accuracy of the trained model copy for local anomaly detection tasks is increased to 89.3%.

[0177] Step 520: Collect the model gradient update amounts from each edge computing node and add random noise to the gradient update amounts using differential privacy technology.

[0178] The specific method of collecting the model gradient update amounts and adding differential privacy noise in step 520 is as follows: After each edge computing node completes local training, extract the gradient matrix of the last layer of the long short-term memory network. The dimension of this matrix is 128×256, which contains the weight parameter update direction information. Using the (ε, δ)-differential privacy protection mechanism, 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.

[0179] 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 values is controlled within the range of ±0.15. While ensuring the effectiveness of the gradient direction, sensitive information that may expose the specific operation details of the device is eliminated. The noise injection process is implemented through a hardware security module to ensure the unpredictability and non-repeatability of the random number generator.

[0180] Step 530: Perform weighted averaging on the gradient update amount after adding noise in the cloud server to generate a global gradient update direction label.

[0181] When generating the global gradient update direction label through weighted averaging in Step 530, the cloud server establishes an aggregation module based on the federated averaging algorithm and assigns weight coefficients according to the number and quality scores of the abnormal data segments uploaded by each edge computing node. In a specific implementation, the weight of the computing node is the proportion 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 segments.

[0182] For example, in an energy storage cluster containing nine edge nodes, the third node contributes 420 abnormal segments (accounting for 18% of the total), with a data quality factor of 1.05, and its weight coefficient is 0.18 × 1.05 = 0.189; the eighth node contributes 300 segments (13%) but the quality factor is 0.9, and the weight coefficient is 0.117. The cloud server performs weighted summation on the nine gradient matrices after adding noise to generate a global gradient matrix with the same dimension. The maximum element value of this matrix converges to 0.39, and the direction consistency index reaches 0.87, indicating that the gradient update directions of each node have significant correlations.

[0183] 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.

[0184] When updating the cloud backbone model parameters in Step 540, the stochastic gradient descent algorithm with momentum is adopted, and the momentum coefficient is set to 0.9, and the learning rate decays to one-tenth of the initial value. During the update process, 28 core parameters strongly associated with historical failure modes in the backbone model are kept unchanged, and only the remaining 1200 parameters are adjusted.

[0185] For example, a certain global update increased the detection sensitivity of the long short-term memory network to temperature mutation intervals by 12%, but the recognition weights for known electrolyte leakage patterns remained stable. Parameter synchronization uses a dual-channel mechanism. The main channel transmits the encrypted model parameter file via HTTPS, and the backup channel uses LoRaWAN to broadcast differential update packets to ensure that parameter synchronization for all edge nodes is completed within fifteen minutes. After synchronization was implemented in a mountain microgrid energy storage system, the model alignment error of twelve edge nodes decreased from 7.3% to 1.8%.

[0186] Step 550: During the synchronization process, use model distillation technology to compress the knowledge of the cloud backbone model into the lightweight model on the edge side to maintain the inference efficiency under the constraints of edge computing resources.

[0187] When performing model distillation in step 550, the cloud backbone model serves as the teacher model to output the unpruned complete probability distribution, and the lightweight model on the edge side serves as the student model to learn the soft target of 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.

[0188] For example, when compressing the 256-dimensional output vector of the cloud backbone model to 64 dimensions in the edge model, feature mapping is achieved through a three-layer fully connected network. On the premise of maintaining an abnormal detection accuracy of contact resistance of 98%, the model calculation amount is reduced by 62%. After distillation, the single inference time of the edge computing node of an industrial and commercial energy storage project decreased from 58 ms to 22 ms, and the memory occupancy was compressed from 1.2 GB to 380 MB, while maintaining a fault recall rate above 95%. During the distillation process, a progressive knowledge transfer strategy is adopted. In the first round, 80% of the parameter knowledge of the teacher model is transferred, and in subsequent rounds, it is gradually refined to local features to ensure a smooth transition during the model compression process.

[0189] As an optional embodiment, the method further includes:

[0190] Step 610: Generate a digital twin of the distributed energy storage device, which is used to map the operating status and environmental parameters of the physical device in real time.

[0191] In step 610, the specific implementation of generating the digital twin of the distributed energy storage device is as follows: Based on the physical topology of the distributed energy storage device, use a parametric modeling tool to generate a three-dimensional dynamic structure model, which includes the precise geometric dimensions and material property parameters of 1200 battery cells, and each unit node is associated with dynamic attributes such as heat conduction coefficient, internal resistance value, and aging factor.

[0192] For example, the temperature sequence of the device, the charge and discharge current waveform, the voltage fluctuation characteristics, and the environmental temperature and humidity data stream can be accessed in real time through the OPC UA protocol to establish a digital mapping relationship that is synchronized and updated with the physical device. Another example is that in the digital transformation of a 2MWh energy storage power station, the digital twin accurately reproduces the three-dimensional spatial layout of forty-eight battery clusters. Each cluster contains twenty-four series-connected modules, and the error of the spacing of the heat dissipation channels between the modules is controlled within ±1mm, and the deviation between the impedance value of the electrical connection path and 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 to update the temperature field distribution cloud map and the current density vector map in the three-dimensional dynamic structure model.

[0193] Step 620: Configure the characteristics of the simulated fault scenario in the digital twin to generate an enhanced training data set containing multiple fault modes.

[0194] 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. For example, the attenuation rate of the loose contact area is set to 0.1 - 0.5mm² / h, and the electrolyte leakage rate is set to 0.5 - 2ml / min. The enhanced training data set contains mixed samples of normal and fault conditions, and one hundred thousand groups of enhanced data are generated through random parameter combinations. For example, when simulating the loose connector fault, the digital twin gradually increases the contact resistance value (increasing by 0.2mΩ per minute), and simultaneously generates the corresponding temperature gradient change curve and voltage fluctuation spectrum to form a fault sample containing four hundred time steps.

[0195] Step 630: Use the enhanced training data 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 to unknown fault modes.

[0196] When performing adversarial training in step 630, first construct a generative adversarial network architecture. The generator receives normal condition data and injects perturbations to generate adversarial samples, and the discriminator distinguishes between real faults and generated faults. During the training process, the generator and the discriminator are alternately optimized to improve the sensitivity of the AI prediction model to hidden fault characteristics. For example, the generator successfully creates an adversarial sample with a subtle temperature difference abnormality of 0.5℃ that the original model fails to detect. After adversarial training, the recognition accuracy of the model for this type of sample is increased from 32% to 87%. On the test set containing unknown fault types, the F1 score of the trained AI prediction model is increased by 15 percentage points.

[0197] Step 640: Deploy the adversarially trained AI prediction model to the actual operation and maintenance environment, and compare the prediction results of the digital twin with the actual state of the physical device 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.

[0198] When deploying the adversarially trained model in Step 640, establish an online verification mechanism. The digital twin generates a state prediction every five minutes and compares it 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 multi-dimensional 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 the temperature of a certain module to be 45°C while the measured value reaches 51°C, the model prediction deviation index rises to 4.7, exceeding the threshold of 3.0 and triggering 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.

[0199] As an independently implementable technical solution, generating the digital twin of the distributed energy storage device in Step 610 includes:

[0200] Step 611: Obtain the real-time operation data stream of the distributed energy storage device, where the real-time operation data stream includes the device temperature sequence, charge and discharge current waveforms, voltage fluctuation characteristics, and environmental parameter set; based on the real-time operation data stream, create a three-dimensional dynamic structure model of the distributed energy storage device, and the three-dimensional dynamic structure model includes a battery unit topology network, a heat dissipation channel distribution map, and an electrical connection path.

[0201] On the one hand, when creating the three-dimensional dynamic structure model in Step 611, use the finite element analysis method to divide hexahedral mesh elements. Each battery unit is divided into eight layers of meshes, and the total number of meshes reaches 3 million. The battery unit topology network stores the connection relationship using a graph database, 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, marking the air flow velocity gradient and temperature decay curve of the forced air cooling system. For example, in the digital twin of a certain liquid-cooled energy storage system, the three-dimensional dynamic structure 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, with a matching degree of 97% with the infrared thermal imaging measured data of the physical device.

[0202] On the other hand, in the real-time running data stream processing of step 611, the data acquisition system captures the charge and discharge current waveforms at a sampling rate of 20,000 times per second, and extracts the harmonic features in the frequency band of 0 - 500 Hz through 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 sixteen sub-bands in the frequency band of 0 - 2 kHz. For example, in the three-dimensional dynamic structure model of a certain coastal energy storage station, the phenomenon of increased contact resistance caused by salt spray corrosion is accurately modeled. The digital twin simulation shows that when the contact area of the connector decreases by 10%, the corresponding temperature rise rate increases by 0.8 °C / min, which is consistent with the trend of the measured data of the physical device under the same working conditions.

[0203] Step 612: Configure a virtual sensor interface in the three-dimensional dynamic structure model, and the virtual sensor interface is used to receive the real-time running data stream and map it to the corresponding position of the three-dimensional dynamic structure model.

[0204] For example, when configuring the virtual sensor interface in step 612, twenty-four virtual temperature probe points are set on the surface of each battery unit in the three-dimensional dynamic structure model, and the position distribution is exactly the same as that of the actual PT1000 sensor array. The virtual sensor interface receives the real-time running data stream of the physical device through the Apache Kafka message queue, maps the device temperature sequence to the corresponding probe points, and decomposes the charge and discharge current waveforms to each branch electrical connection path. For example, when the physical device detects that the temperature of the third module in the fifth battery cluster is abnormal, the digital twin synchronously renders a red warning area at the corresponding position and highlights thirty-two adjacent units affected along the electrical connection path.

[0205] Another example is that 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, electricity, environment) of the physical device into an 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 harmonic content of current. For example, when the physical device detects a current ripple with an amplitude reaching 12% of the rated 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 of the ripple amplitude in real time.

[0206] Step 613: Drive the three-dimensional dynamic structure model through the physical module to simulate the dynamic response behavior of the distributed energy storage device under the action of the real-time running data stream, and generate a device state simulation sequence; align the time sequence of the device state simulation sequence with the real-time running data stream to establish a two-way data mapping channel between the digital twin and the physical device.

[0207] Exemplarily, when driving the three-dimensional dynamic structure model in step 613, a multi-physics coupling simulation module is adopted, and the electro-thermal-mechanical coupling calculation is performed every thirty milliseconds. The charge and discharge current waveforms in the real-time operation data stream are input, the dynamic balance between the Joule heat distribution and the heat dissipation efficiency of the cooling system is solved, and the surface temperature simulation sequence of the battery unit is output. The time series alignment module uses the dynamic time warping algorithm to synchronize the simulation temperature sequence with the measured data of 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 measured value of 8.1 °C by the infrared thermal imager is controlled within 2.4%, verifying the accuracy of the two-way data mapping channel.

[0208] In addition, in the physical module simulation of step 613, an 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 effect of the start and stop of the cooling system. For example, when the forced air cooling system is delayed by 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, and the error of this prediction result verified by actual tests is only 0.7 °C.

[0209] Step 614: Based on the two-way data mapping channel, real-time detect the deviation signal between the device state simulation sequence and the actual operating state of the physical device, and generate a dynamic calibration instruction set; adjust the parameterized components of the three-dimensional dynamic structure model according to the dynamic calibration instruction set, so that the simulation output of the digital twin is synchronized with the measured data of the physical device.

[0210] In the embodiment of the present application, when performing dynamic calibration in step 614, the deviation detection module calculates the root mean square error between the simulation temperature sequence and the measured data, and generates a parameter calibration instruction when the error exceeds 1.5 °C. The calibration algorithm uses the Bayesian optimization method to adjust the thermal conductivity and contact resistance values in the three-dimensional dynamic structure model. For example, during a certain calibration process, it is detected that the simulation temperature of the seventh module is 2.8 °C higher than the measured value, and the system automatically corrects the interfacial contact thermal resistance in the model from 0.8 K·m² / W to 0.92 K·m² / W, making the subsequent three-hour simulation error stable within 0.5 °C.

[0211] In addition, during the dynamic calibration process of step 614, an adaptive parameter adjustment algorithm is developed. When it is detected that the phase deviation between the simulation temperature sequence and the measured data exceeds five seconds, the thermal time constant of the three-dimensional dynamic structure model is automatically corrected. The calibration instruction set contains twelve types of adjustable parameters, and each type of parameter is set with an adjustment range of ±15%. For example, during a certain calibration, the air convection coefficient between modules is corrected from 8.7 W / (m²·K) to 9.3 W / (m²·K), reducing the average error of the subsequent two-hour simulation results to 0.4 °C.

[0212] Step 615: Embed an abnormal propagation path prediction module in the digital twin body. Based on the abnormal characteristics of the device state simulation sequence, simulate the abnormal diffusion path of the fault in the battery unit topology network; superimpose the abnormal diffusion path on the electrical connection path of the three-dimensional dynamic structure model to generate a visualized fault impact map; feedback the visualized fault impact map to the physical device operation and maintenance terminal through the bidirectional data mapping channel to achieve self-update of the digital twin body.

[0213] For example, when implementing abnormal propagation prediction in Step 615, a graph neural network is used to analyze the fault diffusion path in the battery unit topology network. When it is detected that the internal resistance of a certain unit increases abnormally by 5%, the system simulates the propagation process of the fault along the electrical connection path and predicts that the internal resistance values of twelve adjacent units will rise by more than 3% within the next thirty minutes. The visualized fault impact map adopts a gradient coloring scheme, where the red area represents the directly affected unit, and the orange represents the secondary affected unit. The generated heat map is superimposed on the three-dimensional structure model. For example, a certain simulation shows that a poor contact fault will cause the temperature rise rate of six upstream modules to increase by 40%. The accuracy of this prediction result is verified by the actual operation data of the physical device in the subsequent thirty minutes and reaches 89%.

[0214] Another example, in the abnormal propagation prediction of Step 615, an improved Dijkstra algorithm is used to search for the shortest fault propagation path in the battery unit topology network, and the heat runaway reaction kinetics model is combined to calculate the fault diffusion rate. The visualized fault impact map marks four levels of impact areas. The red core area represents the directly faulty unit, the orange primary impact area includes six adjacent units, the yellow secondary impact area extends to twenty-four units, and the blue tertiary impact area covers the entire battery cluster. For example, a certain simulation shows that an electrolyte leakage fault will cause internal short circuits in three adjacent modules within eight minutes. This prediction guides the operation and maintenance personnel to complete the isolation of the faulty module fifteen minutes in advance.

[0215] The embodiment of the present application provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the predictive maintenance method for distributed energy storage devices based on AI is implemented.

[0216] The embodiment of the present application provides a processor, which is used to run a program. When the program runs, the predictive maintenance method for distributed energy storage devices based on AI is executed.

[0217] In the embodiment of the present application, as Figure 2As shown, the predictive maintenance system 100 for distributed energy storage devices includes at least one processor 101, at least one memory 102 connected to the processor 101, and a bus 103. Among them, 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 predictive maintenance method for distributed energy storage devices based on AI.

[0218] This application is described with reference to the flowcharts and / or block diagrams of methods, predictive maintenance systems (systems) for distributed energy storage devices, and computer program products according to embodiments of this application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0219] In a typical configuration, the predictive maintenance system for distributed energy storage devices includes one or more processors (CPUs), a memory, and a bus. The predictive maintenance system for distributed energy storage devices may also include an input / output interface, a network interface, etc.

[0220] The memory may include non-permanent memory in computer-readable media, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory includes at least one storage chip. The memory is an example of computer-readable media.

[0221] A computer-readable medium includes permanent and non-permanent, removable and non-removable media and can implement information storage by any method or technology. The 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 technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, disk storage or other magnetic storage computer-readable storage media or any other non-transmission medium that can be used to store information that can be accessed by a distributed energy storage device predictive maintenance system. As defined herein, a computer-readable medium does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0222] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or computer-readable storage medium comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or computer-readable storage medium. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, article or computer-readable storage medium comprising the element.

[0223] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can 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 code.

[0224] The above are only the embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A predictive maintenance method for AI-based distributed energy storage devices, characterized in that, Applied to a predictive maintenance system for distributed energy storage devices, the method includes: Collect multi-dimensional operation data of the distributed energy storage device, where the multi-dimensional operation data includes device temperature sequences, charge and discharge current waveforms, and voltage fluctuation characteristics; Input the multi-dimensional operation data into a pre-trained AI prediction model to generate a prediction result of the operation state of the distributed energy storage device, where the operation state prediction result includes a device health score and a set of potential abnormal characteristics; Based on the operation state prediction result, analyze the relevance between the set of potential abnormal characteristics and a preset fault mode library to determine the maintenance urgency index of the distributed energy storage device; According to the comparison result between the maintenance urgency index and a preset maintenance threshold, generate a predictive maintenance instruction for the distributed energy storage device, where the predictive maintenance instruction includes a maintenance time window and a maintenance operation type; The method further includes: deploying edge computing nodes in the distributed energy storage device cluster, where the edge computing nodes are used to perform local data preprocessing and model inference; performing preliminary anomaly detection on the collected multi-dimensional operation data through the edge computing nodes to filter out active anomaly data segments; uploading the active anomaly data segments 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 collaborative evolution of distributed models; The execution process of the federated learning strategy includes: allocating independent model training copies to each edge computing node and training locally using active anomaly data segments; collecting model gradient update amounts from each edge computing node and adding random noise to the gradient update amounts using differential privacy technology; performing weighted averaging on the noise-added gradient update amounts in the cloud server to generate a global gradient update direction label; updating the cloud backbone model parameters according to the global gradient update direction label and synchronizing the updated cloud backbone model parameters to each edge computing node; adopting model distillation technology during the synchronization process to compress the knowledge of the cloud backbone model into a lightweight model on the edge side to maintain the inference efficiency under the constraints of edge computing resources; The deploying edge computing nodes in the distributed energy storage device cluster includes: Deploying edge computing hardware units at the neighborhood positions of each physical node in the distributed energy storage device cluster, where the edge computing hardware units are connected to the local sensor array through a target latency communication protocol; receiving the multi-dimensional operation data through the edge computing hardware units, performing format standardization and timestamp alignment processing on the multi-dimensional operation data to generate a preprocessed data stream; Load a lightweight AI inference model in the edge computing hardware unit, input the preprocessed data stream into the lightweight AI inference model, and generate a preliminary anomaly detection result; the preliminary anomaly detection result includes an anomaly type label and a confidence score; according to the comparison result between the confidence score and a preset anomaly threshold, filter out the initial anomaly data segments that meet the confidence requirement from the preprocessed data stream; Perform feature compression encoding on the initial anomaly data segments that meet the confidence requirement to generate bandwidth transmission data packets; upload the bandwidth transmission data packets to the cloud server through an encrypted channel to trigger the model aggregation process of the federated learning strategy; After the globally optimized AI prediction model is generated in the cloud server, receive the downloaded updated model parameter file; load the updated model parameter file into the lightweight AI inference model through a hot update mechanism to overwrite the original model weights; 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; perform cross-verification on the optimized preliminary anomaly detection result and the real-time data stream of the local sensor array, eliminate false alarm anomaly segments and correct the confidence score; dynamically adjust the screening threshold of the initial anomaly data segments that meet the confidence requirement according to the corrected confidence score to achieve adaptive anomaly detection on the edge side.

2. The method according to claim 1, wherein The multi-dimensional operation data further includes an environmental parameter set, and the environmental parameter set includes the humidity change curve, the environmental temperature gradient, and the vibration frequency spectrum of the device deployment location; the collection of the multi-dimensional operation data of the distributed energy storage device includes: Real-time obtain the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics through a sensor array deployed on the distributed energy storage device; Synchronously collect the humidity change curve, the environmental 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 time stamp alignment manner to obtain initial fusion data; wherein, the device operation data at least includes the device temperature sequence, the charge and discharge current waveform, and the voltage fluctuation characteristics obtained in real time; Perform noise filtering and data normalization processing on the initial fusion data to generate target multi-dimensional operation data.

3. The method according to claim 2, wherein The pre-trained AI prediction model generates the operation state prediction result through the following method: Divide the target multi-dimensional operation data into data segments within a continuous time window, and extract the temporal dependence relationship and cross-dimensional correlation features in each data segment; Through the deep convolutional network layer in the AI prediction model, perform local feature enhancement on the temporal dependence relationship to generate an enhanced spatio-temporal feature vector; Use the attention mechanism layer in the AI prediction model to perform weight assignment on the cross-dimensional correlation features, and filter out the key feature subset related to the device health; Based on the comparative analysis of the key feature subset and the historical fault data, output the device health score and the potential anomaly feature set.

4. The method according to claim 3, characterized in that, Parsing the relevance between the set of potential abnormal features and a preset fault mode library to determine a maintenance urgency index for the distributed energy storage device, including: Extracting an abnormal waveform pattern, a temperature mutation interval, and a current distortion feature from the set of potential abnormal features; Performing morphological matching between the abnormal waveform pattern and known fault waveforms in the preset fault mode library to calculate a waveform similarity index; Determining a temperature correlation risk value according to an overlapping ratio between the temperature mutation interval and a historical fault temperature threshold; Generating a comprehensive abnormality confidence level in combination with the distribution density of the current distortion feature in the time dimension; Determining the maintenance urgency index based on a weighted fusion result of the waveform similarity index, the temperature correlation risk value, and the comprehensive abnormality confidence level.

5. The method according to claim 4, wherein 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, including: Matching a corresponding maintenance operation type from a preset maintenance strategy table according to a numerical interval where the maintenance urgency index is located, and the maintenance operation types include battery unit replacement, connector tightening, and heat dissipation system cleaning; Predicting a remaining safe operation duration of the distributed energy storage device based on a decline rate of a device health score output by the AI prediction model; Dynamically adjusting a start time and a duration of the maintenance time window according to the remaining safe operation duration and a standard man-hour required for device maintenance; Encapsulating the maintenance operation type and the maintenance time window into the predictive maintenance instruction and sending it to a target maintenance terminal.

6. The method according to claim 1, wherein The method further includes: After generating the predictive maintenance instruction, monitoring actual operation response data of the distributed energy storage device in real time; Extracting maintenance effect features from the actual operation response data, and the maintenance effect features include a device temperature decline rate, a current waveform smoothness, and a voltage stability improvement index; Feeding back the maintenance effect features to the AI prediction model and updating parameter weights of the AI prediction model through an incremental learning algorithm; Performing iterative prediction on multi-dimensional operation data collected in a target period based on the updated AI prediction model to optimize the accuracy of the maintenance urgency index; Updating the parameter weights of the AI prediction model through the incremental learning algorithm, including: Extracting positive feedback samples and negative feedback samples from the maintenance effect features, where the positive feedback samples represent that the device state conforms to set conditions after a maintenance operation, and the negative feedback samples represent that the device state does not conform to the set conditions after the maintenance operation; Calculating a feature distribution difference between the positive feedback samples and the negative feedback samples to generate a model weight adjustment direction vector; Adopting an online gradient descent algorithm to finely adjust convolution kernel parameters and attention weights of the AI prediction model according to the weight adjustment direction vector; Adding an elastic weight consolidation mechanism during the fine adjustment process to lock model parameters having an associated relationship with historical fault modes and update local parameters associated with newly added maintenance effect features.

7. The method according to claim 1, characterized in that The method further includes: Generate a digital twin of the distributed energy storage device, which is used to map the operating status and environmental parameters of the physical device in real time; Configure simulated fault scenario features in the digital twin to generate an enhanced training data set containing multiple fault modes; Use the enhanced training data 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 state of the physical device 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 in the previous iteration stage.

8. A predictive maintenance system for a distributed energy storage device, characterized in that, It includes a processor, a memory and a bus connected to the processor; wherein, the processor and the memory complete communication with each other through the bus; the processor is used to call program instructions in the memory to execute the AI-based predictive maintenance method for distributed energy storage devices according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, A program is stored thereon, and when the program is executed by the processor, the AI-based predictive maintenance method for distributed energy storage devices according to any one of claims 1-7 is implemented.

Images