A method, system, device and medium for energy storage battery safety early warning

Abstract

This invention discloses a method, system, device, and medium for safety early warning of energy storage batteries. The method includes: acquiring multi-dimensional operational data of the energy storage battery within a target time window as target operational data; using a first anomaly detection model deployed at the edge to perform a risk assessment of the energy storage battery based on the target operational data to obtain a target assessment result; the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, which describes the mapping relationship between the operating state of the energy storage battery and anomaly evolution; determining a target early warning method based on the target assessment result; and performing an early warning for the energy storage battery based on the target early warning method; the early warning method includes edge-side local early warning and server-side collaborative early warning, with a second anomaly detection model deployed on the server side, obtained by structured pruning and volume compression of the reference model. This solution achieves high-precision, low-latency safety early warning for energy storage batteries based on a model hierarchical derivation and collaborative reasoning mechanism.

Description

Technical Field

[0001] This invention relates to the field of energy storage system safety monitoring technology, and in particular to a method, system, device and medium for early warning of energy storage battery safety. Background Technology

[0002] With the accelerated pace of global energy structure transformation, the application of electrochemical energy storage systems, represented by lithium-ion batteries, in power systems continues to expand. Energy storage power stations, as supporting infrastructure for new energy power generation, play a crucial role in smoothing power fluctuations, improving power quality, and participating in grid peak shaving and frequency regulation. However, during long-term high-rate charge-discharge operation, energy storage batteries face various safety risks, including thermal runaway, internal short circuits, increased cell inconsistency, and lithium dendrite growth. If these risks are not identified and warned of in a timely manner, they can easily lead to fires or even explosions, causing severe economic losses and casualties.

[0003] In related technologies, a centralized deep learning approach is employed: all collected data is uploaded to the cloud or server, analyzed using deep learning models (such as LSTM and CNN), and then used to issue safety warnings for energy storage batteries based on the analysis results. Although the models have strong expressive power, models with high computational complexity and memory consumption are difficult to deploy directly on resource-constrained embedded edge devices. Furthermore, energy storage power stations are typically located in remote areas, and due to limited network bandwidth, uploading all data to the cloud for inference results in high communication latency, failing to meet the stringent real-time requirements of energy storage battery safety warnings. Summary of the Invention

[0004] This invention provides a method, system, device, and medium for early warning of energy storage batteries. Based on a unified model, a lightweight edge-side model and a deep inference model are derived hierarchically. While making full use of the powerful representation learning capabilities of the model, the early warning is coordinated between the edge side and the server side, adapting to the computing and memory resource constraints of the edge side, thereby achieving high-precision and low-latency early warning of energy storage batteries.

[0005] According to one aspect of the present invention, a method for early warning of safety of energy storage batteries is provided, the method comprising: Acquire multi-dimensional operational data of the energy storage battery within the target time window as the target operational data; The first anomaly detection model deployed at the edge is used to perform a risk assessment of the energy storage battery based on the target operating data to obtain the target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution; Based on the target evaluation results, a target early warning method is determined, and an early warning for energy storage batteries is performed based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and the server side is deployed with a second anomaly detection model, which is obtained by performing structured pruning and volume compression on the reference model.

[0006] According to another aspect of the present invention, a safety early warning system for energy storage batteries is provided, the system comprising: The data acquisition module is used to acquire multi-dimensional operating data of the energy storage battery within the target time window as the target operating data; The risk assessment module is used to perform a risk assessment of the energy storage battery based on the target operating data using a first anomaly detection model deployed on the edge side to obtain the target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution. A battery early warning module is used to determine a target early warning method based on the target evaluation results, and to perform early warning for energy storage batteries based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and the server side is deployed with a second anomaly detection model, which is obtained by performing structured pruning and volume compression on the reference model.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the energy storage battery safety warning method according to any embodiment of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the energy storage battery safety warning method according to any embodiment of the present invention.

[0009] The technical solution of this invention first acquires multi-dimensional operational data of the energy storage battery within a target time window as target operational data; then, it uses a first anomaly detection model deployed on the edge side to perform a risk assessment of the energy storage battery based on the target operational data to obtain a target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution; then, a target early warning method is determined based on the target assessment result, and an early warning for the energy storage battery is performed based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and a second anomaly detection model is deployed on the server side, which is obtained by structured pruning and volume compression of the reference model. This technical solution, based on a unified model, derives a lightweight edge-side model and a deep inference model on the server side in a hierarchical manner. While fully utilizing the powerful representation learning capabilities of the model, it achieves high-precision, low-latency energy storage battery safety early warning through edge-side and server-side collaborative early warning, triggering server-side collaborative inference only when necessary. This adapts to the computational and memory resource constraints of the edge side, avoids communication delays caused by full data upload, and thus achieves high-precision, low-latency energy storage battery safety early warning.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a safety early warning method for energy storage batteries provided according to an embodiment of the present invention; Figure 2 This is a flowchart of another energy storage battery safety early warning method provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall process of a safety early warning method for energy storage batteries according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an energy storage battery safety early warning system according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device that implements a safety early warning method for energy storage batteries according to an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0014] It should be noted that the terms "first," "second," "target," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example 1 Figure 1 This is a flowchart of a method for early warning of energy storage battery safety provided in Embodiment 1 of the present invention. This embodiment can be applied to the situation of hierarchical and coordinated early warning of energy storage battery safety. The method can be executed by an energy storage battery safety early warning system, which can be implemented in hardware and / or software. The energy storage battery safety early warning system can be configured in an electronic device with data processing capabilities.

[0016] The technical solution provided by this invention is applicable to electrochemical energy storage battery safety early warning systems that achieve unified model hierarchical derivation and collaborative reasoning between the edge and server sides. The edge side is configured as an embedded device, such as a slave control unit of a battery management system or an edge computing gateway, with limited computing power and memory resources. The server side can be configured as a server within an Energy Management System (EMS). This invention generates derived models suitable for different computing environments through unified model hierarchical derivation processing. Through collaborative reasoning between the edge and server sides, server-side collaborative reasoning is triggered only when necessary. This adapts to the computing and memory resource constraints of the edge side, meeting the requirements for low-latency, real-time early warning, while leveraging the deep reasoning capabilities of the server-side model to ensure high-precision early warning accuracy. Thus, it achieves high-precision, low-latency energy storage battery safety early warning.

[0017] like Figure 1As shown, the method includes: S110, acquire multi-dimensional operating data of the energy storage battery within the target time window as the target operating data.

[0018] The time window can refer to a pre-set time length based on actual needs. For example, the time window can be set based on the sampling interval and the number of time steps; specifically, the physical time length corresponding to the time window is equal to the product of the sampling interval and the number of time steps. The target time window can refer to the current time window awaiting anomaly detection. For example, multi-dimensional operational data can include individual cell voltage data, battery cluster temperature data (including individual cell temperature and ambient temperature), charge / discharge current data, estimated state of charge (SOC), and estimated state of health (SOH). Target operational data can refer to the multi-dimensional operational data of the energy storage battery within the target time window.

[0019] In this embodiment, multi-dimensional operating data of the energy storage battery can be collected in real time through CAN bus, Modbus or IEC 61850 communication protocol, with a sampling frequency of not less than 1Hz, and the multi-dimensional operating data of the energy storage battery of the most recent N time steps is locally cached on the edge side as target operating data, where the size of N can be set according to actual needs.

[0020] S120 utilizes the first anomaly detection model deployed on the edge side to perform a risk assessment of the energy storage battery based on the target's operating data to obtain the target assessment result.

[0021] The first anomaly detection model is obtained by knowledge distillation and volume compression of the reference model, which describes the mapping relationship between the operating state of the energy storage battery and the anomaly evolution. For example, the reference model adopts a 12-layer Transformer encoder architecture, with a hidden layer dimension of 768, an attention head of 12, a feedforward network dimension of 3072, and a total parameter count of approximately 1.1B.

[0022] In this embodiment, a reference model needs to be trained based on the historical operating data of the energy storage battery to establish a mapping relationship between the operating state of the energy storage battery and its abnormal evolution. The historical operating data can include historical normal operation data and historical fault operation data; the historical normal operation data can refer to the aforementioned multi-dimensional operating data; the historical fault operation data can include fault type, fault event, and pre-fault operating conditions, where the pre-fault operating conditions can refer to the aforementioned multi-dimensional operating data. Optionally, the training process of the reference model includes: acquiring the historical normal operation data and historical fault operation data of the energy storage battery; training the initial model using a self-supervised learning method to obtain a candidate model based on the historical normal operation data; wherein the initial model adopts a multi-layer Transformer encoder architecture; and training the candidate model using a supervised learning method to obtain the reference model based on the historical fault operation data to obtain a classification task.

[0023] In this embodiment, a combination of supervised and self-supervised learning is employed. Historical normal operation data is used to train the reconstruction task, while labeled historical fault operation data is used to train the classification task. This allows the reference model to possess both strong normal pattern modeling and abnormal pattern recognition capabilities. Specifically, historical normal operation data of the energy storage battery, as well as historical fault operation data including fault type, fault event, and pre-fault operating conditions, are first acquired. Both the historical normal operation data and historical fault operation data are then preprocessed to improve data validity and accuracy. For example, preprocessing may include missing value imputation, outlier cleaning, normalization, and time series alignment. For missing value imputation, when some data is missing, historical statistical values ​​(such as mean, median) or interpolation methods can be used to fill in the missing data, ensuring the continuity and completeness of the time series data. For outlier cleaning, obviously unreasonable data (such as extreme voltage or temperature fluctuations exceeding preset thresholds) can be removed. For normalization, values ​​with different dimensions (such as temperature values ​​of 20-60℃, SOC values ​​of 0-100%) can be mapped to a unified range (such as 0-1 or -1-1) to accelerate model convergence.

[0024] The preprocessed data is then organized into multi-dimensional input vectors that the model can recognize. For example, the input sequence length can be set to 256 time steps (corresponding to approximately 4.3 hours of continuous operation data), with each time step containing a 32-dimensional input vector (voltage × 16, temperature × 8, current × 4, SOC × 2, SOH × 2). It should be noted that the dimension of each parameter can be set according to its importance to fault diagnosis, and the parameter dimension should be consistent with the actual measurement points of the parameter (e.g., the number of temperature sensors corresponds to the number of temperature data dimensions). Next, model training is conducted in two stages. In the first stage, multi-dimensional input vectors corresponding to large-scale unlabeled historical normal operation data are used to perform masked autoencoding pre-training on an initial model using a multi-layer (e.g., 12-layer) Transformer encoder architecture to obtain a candidate model, thereby learning a general representation of the battery's normal operation mode. In the second stage, multi-dimensional input vectors corresponding to historical fault operation data containing multiple (e.g., 5000) precisely labeled faults are used to perform supervised fine-tuning on the candidate model to obtain a reference model, thereby optimizing the anomaly detection and classification tasks. Training employed the AdamW optimizer with an initial learning rate of 1e-4 and a batch size of 64, for a total of 100 epochs. In supervised learning, the cross-entropy loss function was used to measure the difference between the model's predicted values ​​and the true values, aiming to enable the model to accurately distinguish between "normal" and "abnormal" states. The reference model's input was a multi-dimensional input vector, and its output consisted of the anomalous probability distribution and the battery state representation vector obtained through model feature extraction.

[0025] This invention, through the powerful representation learning capability of the reference model, can capture complex nonlinear relationships and long-distance temporal dependencies in battery operation data. It has higher recognition sensitivity for early weak anomalies and complex coupled fault modes, which can significantly reduce the false alarm rate and the false alarm rate, and help improve the accuracy of early warning.

[0026] After the reference model is trained, to adapt to the different computing environments on the edge and server sides, a model derivation technique can be used to perform layered derivation of the reference model. Specifically, for the edge side, the reference model is used as the teacher network, and knowledge distillation is used to train the student network. The student network can employ a shallow Transformer or lightweight convolutional architecture, trained by minimizing the differences between the teacher and student networks in the feature and output layers. After knowledge distillation, the student network is quantized with INT8 integers to further compress the model size and accelerate inference, thus obtaining the first anomaly detection model. The loss function used in the training process is constructed based on hard-label loss and soft-label loss. The supervision signal for hard-label loss comes from the true labels, while the supervision signal for soft-label loss comes from the anomaly probability distribution output by the teacher network. For example, hard-label loss can use cross-entropy loss, and soft-label loss can use KL divergence.

[0027] For the server side, the reference model is structurally pruned to remove redundant attention heads and feedforward network neurons, followed by FP16 half-precision quantization. While preserving the model's deep reasoning ability, the number of parameters is compressed to about 10% of the reference model, thus obtaining the second anomaly detection model. For example, the redundancy of attention heads and feedforward network neurons can be judged in the following ways: (1) Importance score: calculate the contribution of each head to the final output, and those with low contribution can be pruned; (2) Similarity analysis: when the attention patterns learned by multiple heads are highly similar, they can be judged as functional duplication and can be pruned; (3) Activation value statistics: when some neurons have an activation value of 0 or close to 0 on most data, they can be judged as "dead neurons" and can be pruned; (4) L1 or L2 regularization: after sparse training, neurons with parameters close to 0 can be removed. After the model is processed in a hierarchical manner, the first anomaly detection model and the second anomaly detection model can be deployed on the edge side and the server side, respectively.

[0028] This invention deploys a lightweight model on the edge side, achieving extreme lightweighting while sacrificing some accuracy, thus solving the deployment challenge of models on resource-constrained edge devices, while simultaneously meeting the real-time inference requirements of embedded devices. On the server side, a deep inference model is deployed, retaining more model capacity and depth, which can be used to handle complex multivariate coupling analysis and refined fault diagnosis. The two derived models share the core knowledge representation learned from the reference model to ensure consistency and synergy of inference results.

[0029] In this embodiment, after the edge-side model deployment is completed, the first anomaly detection model deployed on the edge side can be used to perform a risk assessment of the energy storage battery based on the target operating data to obtain the target assessment result. Optionally, performing a risk assessment of the energy storage battery based on the target operating data using the first anomaly detection model deployed on the edge side to obtain the target assessment result includes: extracting time-domain and frequency-domain features from the target operating data to obtain target feature data; concatenating the target feature data with the target operating data to form a target input vector; inputting the target input vector into the first anomaly detection model to perform anomaly detection of the energy storage battery to obtain a first anomaly probability distribution; and performing a risk assessment of the energy storage battery based on the first anomaly probability distribution to obtain the target assessment result.

[0030] Specifically, firstly, time-domain statistical features (such as mean, variance, slope, and peak value) and frequency-domain features (such as FFT spectral energy) are extracted from the target operating data to obtain target feature data. Then, the target feature data and the target operating data are concatenated sequentially according to their dimensions and subjected to dimensionality reduction to obtain the target input vector. Next, the target input vector is input into a first anomaly detection model to detect anomalies in the energy storage battery, outputting the first anomaly probability distribution corresponding to the target time window data and the first battery state representation vector obtained through feature extraction by the first anomaly detection model. Subsequently, an energy storage battery risk assessment can be performed based on the first anomaly probability distribution to obtain the target assessment result. For example, the risk assessment content may include at least one of risk level, model uncertainty, and model input distribution bias. For instance, the risk level may include four levels: normal, attention, warning, and danger. Model uncertainty can be used to reflect the prediction confidence of the first anomaly detection model regarding the first anomaly probability distribution; model input distribution bias can be used to characterize the difference between the target input vector and the reference input vector, which is the input vector of the first anomaly detection model during training.

[0031] In this embodiment, optionally, the target assessment result is obtained by performing a risk assessment of the energy storage battery based on the first abnormal probability distribution, including: determining the target risk level based on the maximum abnormal probability score in the first abnormal probability distribution; determining the entropy value of the first abnormal probability distribution as the model uncertainty; wherein, the model uncertainty is used to reflect the prediction confidence of the first anomaly detection model on the first abnormal probability distribution; and determining the target assessment result based on the target risk level and the model uncertainty.

[0032] Specifically, based on the maximum anomaly probability score in the first anomaly probability distribution, a preset risk level database is consulted, and the risk level corresponding to the range of anomaly probability scores containing the maximum anomaly probability score is determined as the target risk level. The preset risk level database describes the mapping relationship between anomaly probability score ranges and risk levels. Furthermore, this can be achieved through formulas... Calculate the entropy value of the first abnormal probability distribution. As model uncertainty, it reflects the confidence level of the first anomaly detection model in predicting the first anomaly probability distribution, where, This represents the anomaly probability score in the first anomaly probability distribution. Therefore, the target risk level and model uncertainty can be used as the target assessment results.

[0033] S130 determines the target early warning method based on the target assessment results, and conducts early warning for energy storage batteries based on the target early warning method.

[0034] The early warning methods include edge-side local early warning and server-side collaborative early warning. Target-based early warning methods can refer to those determined based on target assessment results. It should be noted that a second anomaly detection model is deployed on the server side, which is obtained by performing structured pruning and volume compression on the reference model.

[0035] In this embodiment, optionally, determining the target early warning method based on the target assessment result includes: if the target risk level reaches a preset risk level, then the target early warning method is determined to be server-side collaborative early warning; if the model uncertainty is greater than a preset uncertainty, then the target early warning method is determined to be server-side collaborative early warning.

[0036] The preset risk level can be used to characterize the level of safety risk posed by the energy storage battery. For example, taking four risk levels—normal, caution, warning, and danger—as an example, the normal and caution levels indicate that the energy storage battery poses no safety risk, while the warning and danger levels indicate that it does. In this case, the preset risk level can be set to warning, and both warning and danger levels can be considered to have reached the preset risk level. The preset uncertainty can refer to a model uncertainty threshold pre-set according to actual needs, which can serve as an important basis for determining the reliability of the prediction results of the first anomaly detection model. For example, the preset uncertainty can be set to 1.5.

[0037] In this embodiment, after obtaining the target risk level and model uncertainty, the target risk level can be compared with a preset risk level. If the target risk level reaches the preset risk level, it indicates that the energy storage battery has a safety risk. At this time, the target early warning method can be determined as server-side collaborative early warning. Simultaneously, the model uncertainty is compared with a preset uncertainty. If the model uncertainty is greater than the preset uncertainty, it indicates that the prediction result of the first anomaly detection model is unreliable. In this case, the target early warning method can also be determined as server-side collaborative early warning, so as to ensure the accuracy of the early warning through server-side deep inference.

[0038] In this embodiment, optionally, determining the target early warning method based on the target evaluation result further includes: determining the model input distribution deviation according to the difference between the target input vector and the reference input vector; wherein, the reference input vector is the input vector of the first anomaly detection model during the training process; if the model input distribution deviation is greater than the preset distribution deviation, the target early warning method is determined to be server-side collaborative early warning; if the current detection time reaches the preset collaborative inspection cycle, the target early warning method is determined to be server-side collaborative early warning.

[0039] In this embodiment, the difference between the target input vector and the reference input vector can be determined based on KL divergence or Wasserstein distance as the model input distribution deviation. The reference input vector is the input vector of the first anomaly detection model during training. A larger KL divergence or Wasserstein distance indicates a greater difference between the target input vector and the reference input vector. Then, the model input distribution deviation is compared with a preset distribution deviation, and the target warning method is determined based on the comparison result. The preset distribution deviation can refer to a pre-set threshold for model input distribution deviation according to actual needs, serving as an important basis for measuring whether the model input distribution has undergone a significant shift. If the model input distribution deviation is greater than the preset distribution deviation, it indicates a significant shift in the model input distribution. This significant shift may be caused by factors such as equipment aging or seasonal changes. In this case, the target warning method can be determined as server-side collaborative warning, and the shift event is recorded for subsequent model updates. The content of the shift event may include the first battery state representation vector, the first anomaly probability distribution, the model input distribution deviation, a timestamp, and target running data.

[0040] Furthermore, triggering conditions for server-side collaborative alerts can be set based on a preset collaborative inspection cycle. The preset collaborative inspection cycle refers to the time interval for server-side collaborative alerts set according to actual needs. For example, the preset collaborative inspection cycle can be set to 15 minutes. If the current detection time reaches the preset collaborative inspection cycle, the target alert method can be determined as a server-side collaborative alert.

[0041] In general, server-side collaborative early warning can be triggered when any of the following conditions are met: (1) the target risk level reaches the preset risk level; (2) the model uncertainty is greater than the preset uncertainty; (3) the model input distribution deviation is greater than the preset distribution deviation; (4) the current detection time reaches the preset collaborative inspection cycle. Among them, conditions (1)-(3) are abnormal triggering conditions, which can effectively deal with the occurrence of abnormal events; condition (4) is a periodic triggering condition, which ensures continuous monitoring of gradual degradation. It should be noted that after each server-side collaborative early warning is triggered, regardless of whether the preset collaborative inspection cycle has been reached, the timing needs to be reset, thereby reducing the waste of computing resources and the delay in early warning time caused by frequent triggering of server-side collaborative early warning.

[0042] This invention incorporates model uncertainty indicators and model input distribution offset detection into the collaborative inference triggering mechanism, enabling the system to automatically identify its own cognitive boundaries and proactively seek support from the server-side model when confidence is insufficient. This avoids the decrease in early warning accuracy caused by blind decision-making by the edge-side model, thus improving system robustness. Furthermore, the inference latency of the lightweight edge-side model can be controlled within 50 milliseconds (example only), meeting the stringent real-time requirements of energy storage safety early warning. It also triggers server-side collaborative inference only when necessary, avoiding communication delays caused by full data uploads and achieving low-latency real-time response on the edge side.

[0043] In this embodiment, optionally, the energy storage battery early warning based on the target early warning method includes: if the target early warning method is server-side collaborative early warning, then the target operating data and the first battery state representation vector are pushed to the server side; wherein, the first battery state representation vector is obtained by feature extraction of the target operating data through a first anomaly detection model; using a second anomaly detection model deployed on the server side, the energy storage battery fault diagnosis is performed based on the target operating data and the first battery state representation vector to obtain the target diagnosis result; wherein, the target diagnosis result includes the fault type identification result and the risk level determination result; the target control command is generated by querying a preset control strategy library based on the target diagnosis result, and the energy storage battery early warning is performed based on the target diagnosis result and the target control command; wherein, the preset control strategy library is used to describe the mapping relationship between fault type, risk level and control command.

[0044] Specifically, if the target early warning method is server-side collaborative early warning, the target operating data and the first battery state representation vector output by the first anomaly detection model can be pushed to the server side. Using the second anomaly detection model deployed on the server side, energy storage battery fault diagnosis is performed based on the target operating data and the first battery state representation vector, obtaining target diagnostic results including fault type identification results and risk level determination results. The fault type can include at least one of the following: thermal runaway precursor, internal short circuit, electrolyte leakage, loose connection, inconsistent SOC, and accelerated capacity decay.

[0045] In this embodiment, optionally, a second anomaly detection model deployed on the server side is used to perform energy storage battery fault diagnosis based on target operating data and a first battery state representation vector to obtain a target diagnosis result. This includes: inputting the target operating data and the first battery state representation vector into the second anomaly detection model to perform energy storage battery anomaly detection, obtaining a second battery state representation vector and a second anomaly probability distribution; wherein, the second battery state representation vector is obtained by the second anomaly detection model through feature extraction of the target operating data based on the first battery state representation vector; querying a preset fault mode library based on the second battery state representation vector to obtain a fault type identification result; wherein, the preset fault mode library is used to describe the mapping relationship between the battery state representation vector and the fault type; performing an energy storage battery risk assessment based on the second anomaly probability distribution to obtain a risk level determination result; and determining the target diagnosis result based on the fault type identification result and the risk level determination result.

[0046] When diagnosing energy storage battery faults, the target operating data and the first battery state representation vector are first input into the second anomaly detection model. Utilizing the multi-layer attention mechanism of the second anomaly detection model, the first battery state representation vector is used as a reference input to perform cross-module and cross-cluster correlation analysis on the target operating data. This identifies the causal relationships and propagation paths between single-point anomalies and system-level anomalies, thereby achieving deep anomaly detection of the energy storage battery. The final output is the second anomaly probability distribution and the second battery state representation vector. Then, a preset fault mode library is queried based on the second battery state representation vector. This library describes the mapping relationship between battery state representation vectors and fault types. The similarity between the second battery state representation vector and each battery state representation vector in the preset fault mode library is calculated. The battery state representation vector with the highest similarity is selected as the target battery state representation vector. The fault type corresponding to the target battery state representation vector in the preset fault mode library is used as the fault type identification result. Simultaneously, an energy storage battery risk assessment is performed based on the second anomaly probability distribution to obtain the risk level determination result. The specific implementation process can refer to the energy storage battery risk assessment using the first anomaly probability distribution, and will not be elaborated here. Therefore, the fault type identification results and risk level determination results can be used as the final target diagnostic results.

[0047] After determining the target diagnostic result, a preset control strategy library can be queried based on the result, and the control instructions matching the target diagnostic result in the library can be used as the target control instructions. The preset control strategy library describes the mapping relationship between fault type, risk level, and control instructions. Specifically, the fault type determines the control action to take; for example, thermal runaway requires circuit disconnection and cooling activation, while leakage requires battery cluster isolation and fire suppression activation. The risk level determines the extent of the action; for example, a hazard level requires immediate automatic control, a warning level requires personnel to prepare contingency plans, and a caution level only increases the frequency of inspections. For example, control instructions may include at least one of the following: adjusting charging and discharging power limits, initiating active balancing, triggering fire pre-processing, or performing emergency disconnection. The target diagnostic result can then be fed back as early warning information to the edge side. For example, based on the risk level, different warning methods such as pop-up reminders, voice broadcasts, and audible and visual alarms can be used on the edge side to update local displays and log records. Simultaneously, the target diagnostic result is cached as tag data for subsequent model updates. Additionally, the target control instructions can be fed back to the battery management system, which then executes corresponding safety control actions based on the received instructions.

[0048] Furthermore, if the fault type identification results include thermal runaway-related faults (such as thermal runaway precursors), the temperature evolution trend within a future time window T can be predicted based on the time-series extrapolation model deployed on the server side using temperature data from the target operating data. Simultaneously, based on the electrochemical-thermal coupling model deployed on the server side, thermal runaway prediction results are obtained by inputting the target operating data. The electrochemical-thermal coupling model can describe the electrochemical reactions and heat generation during battery charging and discharging, and the thermal runaway prediction results can include the thermal runaway trigger temperature, thermal runaway trigger time, and thermal runaway probability. Correspondingly, the thermal runaway prediction results can be used to generate target control commands, for example, setting the thermal runaway trigger time as the execution time of the target control command, to better address thermal runaway-related faults.

[0049] In this embodiment, optionally, energy storage battery early warning is performed based on the target early warning method, including: if the target early warning method is edge-side local early warning, then the target inspection frequency is determined according to the target evaluation result; and energy storage battery early warning is performed according to the target evaluation result and the target inspection frequency.

[0050] In this embodiment, if the target warning method is edge-side local warning, there is no need to trigger server-side collaborative warning. In this case, the target inspection frequency can be determined based on the target assessment results, and the energy storage battery warning can be performed locally on the edge side based on the target assessment results and the target inspection frequency. Specifically, if the target risk level is normal, the initial inspection frequency (e.g., once per second) can be directly determined as the target inspection frequency, that is, the initial inspection frequency is kept unchanged for continuous inspection, and there is no need to perform energy storage battery warning; if the target risk level is attention level, the inspection frequency needs to be increased based on the initial inspection frequency to obtain the target inspection frequency (e.g., twice per second), thereby performing encrypted inspection, and at the same time, prompts and warnings can be given on the edge side.

[0051] It should be noted that this invention fully considers the actual engineering constraints of energy storage power stations. All modules of the system can be integrated with existing battery management systems and station-level energy management systems through standardized interfaces (IEC 61850, Modbus TCP, etc.), without the need for large-scale modifications to existing infrastructure, thus possessing good engineering practicality.

[0052] The technical solution of this invention first acquires multi-dimensional operational data of the energy storage battery within a target time window as target operational data; then, it uses a first anomaly detection model deployed on the edge side to perform a risk assessment of the energy storage battery based on the target operational data to obtain a target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution; then, a target early warning method is determined based on the target assessment result, and an early warning for the energy storage battery is performed based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and a second anomaly detection model is deployed on the server side, which is obtained by structured pruning and volume compression of the reference model. This technical solution, based on a unified model, derives a lightweight edge-side model and a deep inference model on the server side in a hierarchical manner. While fully utilizing the powerful representation learning capabilities of the model, it achieves high-precision, low-latency energy storage battery safety early warning through edge-side and server-side collaborative early warning, triggering server-side collaborative inference only when necessary. This adapts to the computational and memory resource constraints of the edge side, avoids communication delays caused by full data upload, and thus achieves high-precision, low-latency energy storage battery safety early warning.

[0053] Example 2 Figure 2This is a flowchart of a safety early warning method for energy storage batteries provided in Embodiment 2 of the present invention. This embodiment is based on the above embodiment and optimized. Specifically, the optimization includes: if the cumulative amount of newly added operating data of the energy storage battery reaches a preset amount of data, then a low-rank adaptation method is used to update the reference model based on the newly added operating data; based on the updated reference model, the first anomaly detection model and the second anomaly detection model are updated respectively.

[0054] like Figure 2 As shown, the method in this embodiment specifically includes the following steps: S210, acquire the multi-dimensional operation data of the energy storage battery within the target time window as the target operation data.

[0055] S220 utilizes the first anomaly detection model deployed on the edge side to perform a risk assessment of the energy storage battery based on the target's operating data to obtain the target assessment result.

[0056] The first anomaly detection model is obtained by knowledge distillation and volume compression of the reference model, which is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution.

[0057] S230 determines the target early warning method based on the target assessment results, and conducts early warning for energy storage batteries based on the target early warning method.

[0058] The early warning methods include edge-side local early warning and server-side collaborative early warning. A second anomaly detection model is deployed on the server side, which is obtained by performing structured pruning and volume compression on the reference model. Furthermore, the specific implementation methods of steps S210-S230 can be found in the detailed description in Embodiment 1 above, and will not be repeated here.

[0059] S240, if the cumulative amount of newly added operating data of the energy storage battery reaches the preset amount of data, the reference model is updated based on the newly added operating data using a low-rank adaptation method.

[0060] In this embodiment, to achieve continuous model evolution, the Low-Rank Adaptation (LoRA) method can be used for adaptive fine-tuning and updating of the model. Specifically, when the cumulative amount of newly added running data from the energy storage battery reaches a preset amount, most of the pre-trained weights of the reference model are frozen, and only the injected low-rank adaptation matrix is ​​trained. For example, the adaptation matrix is ​​a low-rank decomposition matrix with rank r=8-16. Then, the fine-tuned low-rank adaptation matrix is ​​merged with the reference model weights to update the reference model. The newly added running data can include target running data, offset event data, and pseudo-label data generated by deep inference using the server-side model. The preset amount of data can be set according to actual needs to avoid wasting computing resources due to frequent model updates using too little data. It should be noted that after each reference model update, the newly added running data needs to be counted again. The low-rank adaptation method can significantly reduce the number of trainable parameters (only 0.1%-1% of full fine-tuning) while maintaining rapid adaptability to new scenarios.

[0061] S250, based on the updated reference model, update the first anomaly detection model and the second anomaly detection model respectively.

[0062] In this embodiment, after the reference model is updated, the updated reference model can be used as a new teacher network for knowledge distillation and volume compression to obtain the updated first anomaly detection model. Simultaneously, the updated reference model undergoes structured pruning and volume compression to obtain the updated second anomaly detection model. For example, the first and second anomaly detection models can be pushed to the edge and server sides respectively via OTA (Over-The-Air) to complete the model deployment and update.

[0063] In the technical solution of this invention, if the cumulative amount of newly added operational data from the energy storage battery reaches a preset amount, a low-rank adaptation method is used to update the reference model based on the newly added operational data; based on the updated reference model, the first anomaly detection model and the second anomaly detection model are updated respectively. This technical solution, through the LoRA fine-tuning mechanism, allows new sites to quickly adapt using only a small amount of local data, significantly reducing the technical threshold and data cost of deploying new sites. It enables continuous model evolution while supporting cross-site migration, which helps to further improve the accuracy of energy storage battery safety early warning.

[0064] Figure 3This paper presents a schematic diagram of the overall process of a safety early warning method for energy storage batteries, involving a total of 6 steps. For example, the energy storage battery safety early warning system proposed in this invention is deployed in a 100MW / 200MWh electrochemical energy storage power station. The edge computing layer adopts an ARM architecture embedded gateway (4GB RAM, 64GB storage, quad-core Cortex-A72 processor), running a lightweight edge-side model quantized with INT8. The station-level EMS server uses dual Xeon processors, 128GB RAM, and an RTX 4090 GPU accelerator, running a deep inference model quantized with FP16. The edge devices and the EMS server are connected via industrial Ethernet, using IEC 61850 GOOSE messages and Modbus TCP as the communication protocols, with a network bandwidth of 100Mbps and an average communication latency of approximately 5ms.

[0065] like Figure 3 As shown, in step S1, historical operating data of the energy storage battery is collected and preprocessed. The preprocessed historical operating data is used to train a Transformer-based unified basic model (i.e., a reference model) to learn the mapping relationship between battery state and anomaly evolution, thus completing the construction of the unified basic model. In step S2, a hierarchical model derivation technique is used. The reference model is structured pruning and FP16 half-precision quantization to generate an EMS-side model (i.e., the second anomaly detection model on the server side) and deployed. Simultaneously, the reference model is knowledge distilled and INT8 integer quantization to generate an edge-side model (i.e., the first anomaly detection model on the edge side) and deployed. In step S3, data acquisition and time-frequency feature extraction are performed using an edge-side embedded device to construct a target input vector. This target input vector is input into a lightweight model for forward inference, ultimately outputting the target risk level and model uncertainty, thus completing real-time edge-side inference. Then, based on the target risk level and model uncertainty, combined with the model input distribution deviation and a preset collaborative inspection cycle, trigger conditions are determined. If the triggering condition is met, the target operation data collected on the edge side and the first battery state representation vector output by the lightweight model on the end side are uploaded to the EMS, and step S4 is entered to perform EMS deep inference; if the triggering condition is not met, step S3 is repeated to perform real-time inference on the end side.

[0066] In step S4, multivariate fusion analysis and fault mode identification are performed based on the target operating data and the first battery state representation vector, and thermal runaway prediction and control strategy (i.e., target control command) are generated, thereby completing EMS deep inference. In step S5, the target control command is fed back to the BMS (Battery Management System) as an EMS result for command execution control, thereby completing result feedback and safety control. In step S6, if the cumulative amount of newly added operating data of the energy storage battery reaches a preset amount, the unified basic model is fine-tuned using low-rank adaptation (LoRA) based on the new data (i.e., newly added operating data), and the end-side model and EMS-side model are synchronously updated using the fine-tuned unified basic model, thereby completing the model adaptive update.

[0067] For example, under normal operating conditions, the edge-side lightweight model performs inference once per second, outputting risk level and uncertainty indicators. Suppose that at a certain moment, the temperature of cell number 15 in battery cluster 2 rapidly rises from 35°C to 52°C within 30 seconds, while the cell's voltage experiences an abnormal voltage drop of 50mV. The edge-side model detects that the maximum anomaly probability score increases from 0.02 to 0.78, the risk level jumps from "normal" to "warning," and the output entropy value increases from 0.3 to 1.8, exceeding the preset uncertainty. According to the collaborative inference triggering mechanism, the edge device automatically uploads the complete data (256 steps × 32 dimensions) for this time window, along with the first battery state representation vector (output of the 6th hidden layer of the edge-side model, dimension 256), to the EMS server. After receiving the data, the EMS deep model performed cross-cluster correlation analysis and found that the temperature gradient of the module containing this cell was significantly higher than that of other modules. By querying the preset fault mode library, the fault type was identified as "early thermal runaway precursor with internal short circuit," and the predicted thermal runaway trigger time window was 15-25 minutes with a confidence level of 92%. The EMS automatically generated control commands: limiting the charging and discharging power of the module to 30% of the rated power, initiating module-level active balancing, and sending a warning notification to maintenance personnel. The BMS completed the power adjustment within 2 seconds of receiving the control command, effectively curbing the risk of thermal runaway.

[0068] For example, after a complete charge-discharge cycle season, the system detected a significant shift in the target input vector relative to the training baseline (i.e., the reference input vector), such as the Wasserstein distance increasing from 0.05 to 0.18. This triggered the model adaptive update process: the system selected 20,000 high-quality samples (including pseudo-labels generated by EMS deep inference) from the past three months of running data and performed LoRA fine-tuning on the unified base model. The rank of the adaptation matrix was set to r=16, the scaling factor alpha=32, and only the injected adaptation parameters (approximately 0.5% of the total parameters) were trained. The fine-tuning process took approximately 2 hours to complete, and the anomaly detection F1-score on the validation set improved from 0.91 to 0.96. After fine-tuning, knowledge distillation was re-performed on the teacher network using the updated unified base model to generate a new lightweight edge model, which was then pushed to all edge devices via OTA during low-load periods at night. The updated edge model improved the accuracy of identifying novel anomaly patterns by approximately 8 percentage points without increasing inference latency.

[0069] Example 3 Figure 4 This is a schematic diagram of a safety early warning system for an energy storage battery provided in Embodiment 3 of the present invention. This system can execute the safety early warning method for an energy storage battery provided in any embodiment of the present invention, and possesses the corresponding functional modules and beneficial effects of the method. For example... Figure 4 As shown, the system includes: The data acquisition module 310 is used to acquire multi-dimensional operating data of the energy storage battery within the target time window as the target operating data; The risk assessment module 320 is used to perform a risk assessment of the energy storage battery based on the target operating data using a first anomaly detection model deployed on the edge side to obtain the target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution. The battery early warning module 330 is used to determine the target early warning method based on the target evaluation result, and to perform early warning for the energy storage battery based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and the server side is deployed with a second anomaly detection model, which is obtained by performing structured pruning and volume compression on the reference model.

[0070] Optionally, the system further includes: a model training module, used for: Acquire historical normal operation data and historical fault operation data of energy storage batteries; Based on the historical normal operation data, a candidate model is obtained by reconstructing the initial model using a self-supervised learning method; wherein, the initial model adopts a multi-layer Transformer encoder architecture; Based on the historical fault operation data, a reference model is obtained by training the candidate model on a classification task using a supervised learning method.

[0071] Optionally, the risk assessment module 320 is used for: The target running data is subjected to time-domain and frequency-domain feature extraction to obtain target feature data; The target feature data and the target running data are concatenated to form the target input vector; The target input vector is input into the first anomaly detection model to perform anomaly detection on the energy storage battery and obtain the first anomaly probability distribution. The target assessment result is obtained by performing a risk assessment of the energy storage battery based on the first abnormal probability distribution.

[0072] Optionally, the risk assessment module 320 is further configured to: The target risk level is determined based on the maximum abnormal probability score in the first abnormal probability distribution; The entropy value of the first anomaly probability distribution is determined as the model uncertainty; wherein, the model uncertainty is used to reflect the prediction confidence of the first anomaly detection model for the first anomaly probability distribution; The target assessment result is determined based on the target risk level and the model uncertainty.

[0073] Optionally, the battery warning module 330 is used for: If the target risk level reaches the preset risk level, the target early warning method will be determined as server-side collaborative early warning; If the uncertainty of the model is greater than the preset uncertainty, then the target early warning method will be determined as server-side collaborative early warning.

[0074] Optionally, the battery warning module 330 is further used for: The model input distribution bias is determined based on the difference between the target input vector and the reference input vector; wherein, the reference input vector is the input vector of the first anomaly detection model during the training process; If the model input distribution deviation is greater than the preset distribution deviation, then the target early warning method is determined to be server-side collaborative early warning; If the current detection time reaches the preset collaborative inspection cycle, the target early warning method will be determined as server-side collaborative early warning.

[0075] Optionally, the battery warning module 330 is further used for: If the target early warning method is server-side collaborative early warning, then the target operating data and the first battery status representation vector are pushed to the server side; wherein, the first battery status representation vector is obtained by feature extraction of the target operating data through the first anomaly detection model; Using the second anomaly detection model deployed on the server side, the target battery fault diagnosis is performed based on the target operating data and the first battery state representation vector to obtain the target diagnosis result; wherein, the target diagnosis result includes fault type identification result and risk level determination result; Based on the target diagnostic results, a target control command is generated by querying a preset control strategy library. Based on the target diagnostic results and the target control command, an early warning for the energy storage battery is issued. The preset control strategy library is used to describe the mapping relationship between fault type, risk level and control command.

[0076] Optionally, the battery warning module 330 is further used for: The target operating data and the first battery state representation vector are input into the second anomaly detection model to perform energy storage battery anomaly detection, thereby obtaining the second battery state representation vector and the second anomaly probability distribution; wherein, the second battery state representation vector is obtained by the second anomaly detection model through feature extraction of the target operating data based on the first battery state representation vector; The fault type identification result is obtained by querying the preset fault mode library based on the second battery state representation vector; wherein, the preset fault mode library is used to describe the mapping relationship between the battery state representation vector and the fault type; The risk level determination result is obtained by conducting a risk assessment of the energy storage battery based on the second abnormal probability distribution. The target diagnostic result is determined based on the fault type identification result and the risk level determination result.

[0077] Optionally, the battery warning module 330 is further used for: If the target early warning method is edge-side local early warning, then the target inspection frequency is determined based on the target evaluation results; Early warning of energy storage batteries is generated based on the target assessment results and the target inspection frequency.

[0078] Optionally, the system further includes: a model update module, used for: If the cumulative amount of newly added operational data of the energy storage battery reaches the preset amount of data, the reference model will be updated based on the newly added operational data using a low-rank adaptation method. The first anomaly detection model and the second anomaly detection model are updated based on the updated reference model.

[0079] The energy storage battery safety early warning system provided in this embodiment of the invention can execute the energy storage battery safety early warning method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0080] Example 4 Figure 5 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0081] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0082] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0083] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as energy storage battery safety warning methods.

[0084] In some embodiments, the energy storage battery safety warning method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the energy storage battery safety warning method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the energy storage battery safety warning method by any other suitable means (e.g., by means of firmware).

[0085] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0086] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0087] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0088] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0089] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0090] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0091] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0092] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for early warning of energy storage battery safety, characterized in that, The method includes: Acquire multi-dimensional operational data of the energy storage battery within the target time window as the target operational data; The first anomaly detection model deployed at the edge is used to perform a risk assessment of the energy storage battery based on the target operating data to obtain the target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution; Based on the target evaluation results, a target early warning method is determined, and an early warning for energy storage batteries is performed based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and the server side is deployed with a second anomaly detection model, which is obtained by performing structured pruning and volume compression on the reference model.

2. The method according to claim 1, characterized in that, The training process of the reference model includes: Acquire historical normal operation data and historical fault operation data of energy storage batteries; Based on the historical normal operation data, a candidate model is obtained by reconstructing the initial model using a self-supervised learning method; wherein, the initial model adopts a multi-layer Transformer encoder architecture; Based on the historical fault operation data, a reference model is obtained by training the candidate model on a classification task using a supervised learning method.

3. The method according to claim 1, characterized in that, The first anomaly detection model deployed at the edge performs a risk assessment of the energy storage battery based on the target's operational data to obtain the target assessment result, including: The target running data is subjected to time-domain and frequency-domain feature extraction to obtain target feature data; The target feature data and the target running data are concatenated to form the target input vector; The target input vector is input into the first anomaly detection model to perform anomaly detection on the energy storage battery and obtain the first anomaly probability distribution. The target assessment result is obtained by performing a risk assessment of the energy storage battery based on the first abnormal probability distribution.

4. The method according to claim 3, characterized in that, The target assessment result is obtained by performing a risk assessment of the energy storage battery based on the first abnormal probability distribution, including: The target risk level is determined based on the maximum abnormal probability score in the first abnormal probability distribution; The entropy value of the first anomaly probability distribution is determined as the model uncertainty; wherein, the model uncertainty is used to reflect the prediction confidence of the first anomaly detection model for the first anomaly probability distribution; The target assessment result is determined based on the target risk level and the model uncertainty.

5. The method according to claim 4, characterized in that, Based on the target assessment results, the target early warning method is determined, including: If the target risk level reaches the preset risk level, the target early warning method will be determined as server-side collaborative early warning; If the uncertainty of the model is greater than the preset uncertainty, then the target early warning method will be determined as server-side collaborative early warning.

6. The method according to claim 5, characterized in that, Determining the target early warning method based on the target assessment results also includes: The model input distribution bias is determined based on the difference between the target input vector and the reference input vector; wherein, the reference input vector is the input vector of the first anomaly detection model during the training process; If the model input distribution deviation is greater than the preset distribution deviation, then the target early warning method is determined to be server-side collaborative early warning; If the current detection time reaches the preset collaborative inspection cycle, the target early warning method will be determined as server-side collaborative early warning.

7. The method according to any one of claims 1-6, characterized in that, Early warning of energy storage batteries based on the aforementioned target early warning method includes: If the target early warning method is server-side collaborative early warning, then the target operating data and the first battery status representation vector are pushed to the server side; wherein, the first battery status representation vector is obtained by feature extraction of the target operating data through the first anomaly detection model; Using the second anomaly detection model deployed on the server side, the target battery fault diagnosis is performed based on the target operating data and the first battery state representation vector to obtain the target diagnosis result; wherein, the target diagnosis result includes fault type identification result and risk level determination result; Based on the target diagnostic results, a target control command is generated by querying a preset control strategy library. Based on the target diagnostic results and the target control command, an early warning for the energy storage battery is issued. The preset control strategy library is used to describe the mapping relationship between fault type, risk level and control command.

8. The method according to claim 7, characterized in that, Using the second anomaly detection model deployed on the server side, the target diagnosis result is obtained by performing energy storage battery fault diagnosis based on the target operating data and the first battery state representation vector, including: The target operating data and the first battery state representation vector are input into the second anomaly detection model to perform energy storage battery anomaly detection, thereby obtaining the second battery state representation vector and the second anomaly probability distribution; wherein, the second battery state representation vector is obtained by the second anomaly detection model through feature extraction of the target operating data based on the first battery state representation vector; The fault type identification result is obtained by querying the preset fault mode library based on the second battery state representation vector; wherein, the preset fault mode library is used to describe the mapping relationship between the battery state representation vector and the fault type; The risk level determination result is obtained by conducting a risk assessment of the energy storage battery based on the second abnormal probability distribution. The target diagnostic result is determined based on the fault type identification result and the risk level determination result.

9. The method according to any one of claims 1-6, characterized in that, Early warning of energy storage batteries based on the aforementioned target early warning method includes: If the target early warning method is edge-side local early warning, then the target inspection frequency is determined based on the target evaluation results; Early warning of energy storage batteries is generated based on the target assessment results and the target inspection frequency.

10. The method according to any one of claims 1-6, characterized in that, The method further includes: If the cumulative amount of newly added operational data of the energy storage battery reaches the preset amount of data, the reference model will be updated based on the newly added operational data using a low-rank adaptation method. The first anomaly detection model and the second anomaly detection model are updated based on the updated reference model.

11. A safety early warning system for energy storage batteries, characterized in that, The system includes: The data acquisition module is used to acquire multi-dimensional operating data of the energy storage battery within the target time window as the target operating data; The risk assessment module is used to perform a risk assessment of the energy storage battery based on the target operating data using a first anomaly detection model deployed on the edge side to obtain the target assessment result; wherein, the first anomaly detection model is obtained by knowledge distillation and volume compression of a reference model, and the reference model is used to describe the mapping relationship between the operating state of the energy storage battery and the anomaly evolution. A battery early warning module is used to determine a target early warning method based on the target evaluation results, and to perform early warning for energy storage batteries based on the target early warning method; wherein, the early warning method includes edge-side local early warning and server-side collaborative early warning, and the server side is deployed with a second anomaly detection model, which is obtained by performing structured pruning and volume compression on the reference model.

12. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the energy storage battery safety warning method according to any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are used to cause a processor to execute the energy storage battery safety early warning method according to any one of claims 1-10.

Images