A safety early warning method for energy storage system based on multi-modal perception

Abstract

The application discloses a kind of energy storage system safety early warning methods based on multi-modal perception, multi-modal sensor array is deployed in key node of energy storage system, multi-modal data are collected by clock synchronization and complete noise cleaning, filtering smooth and space-time alignment preprocessing;Normal fluctuation interval is demarcated using statistical method, and continuous out-of-range data is marked as abnormal segment, based on abnormality judgment, each modal feature vector is extracted, and contribution weight is calculated through cross-modal dynamic attention network and weighted fusion;Weighted feature vector is input into bidirectional LSTM to extract time sequence characteristics and map into risk representation results, the similarity between the historical fault mode vector is calculated using the transfer learning early warning model, and the threshold value is triggered if it is over the threshold value;Early warning threshold is dynamically adjusted in combination with operating state and environmental parameters, generate multidimensional early warning information and push in multiple channels, and according to operation and maintenance feedback, parameters are optimized in closed loop, the application can effectively improve early warning accuracy and adaptive capacity.

Description

Technical Field

[0001] This application belongs to the field of energy storage system safety monitoring technology, specifically involving a safety early warning method for energy storage systems based on multimodal sensing. Background Technology

[0002] Energy storage system safety early warning refers to a comprehensive technical system that, during the operation of an energy storage system, identifies potential safety risks in advance and issues warning signals by real-time monitoring of key operating parameters, combined with data analysis and intelligent algorithms, in order to prevent accidents. Because energy storage systems involve high-energy-density electrochemical reactions, they are susceptible to safety hazards such as thermal runaway, overcharging and over-discharging, short circuits, gas leaks, fires, and even explosions; therefore, safety early warning systems are crucial.

[0003] Currently, most energy storage system safety early warning technologies employ single-modal monitoring or simple multimodal data overlay, lacking systematic processing and deep integration of multimodal data. Existing technologies are prone to issues such as time asynchrony and data distortion during multimodal data acquisition, leading to insufficient accuracy in anomaly identification. Furthermore, anomaly judgments often rely on fixed thresholds, failing to adapt to dynamic changes in energy storage system operating conditions and environmental parameters, resulting in false alarms and missed alarms.

[0004] In addition, existing early warning methods lack effective adaptive adjustment mechanisms, making it impossible to optimize monitoring and early warning parameters based on changes in system operating status. After long-term operation, the accuracy of early warning is prone to decline. Furthermore, a complete feedback optimization chain is not formed during the early warning process, making it difficult to continuously improve early warning performance through operation and maintenance data.

[0005] In summary, existing energy storage safety early warning technologies suffer from problems such as low anomaly identification accuracy, poor adaptability, high false alarm and false alarm rates, and insufficient long-term operational stability, which cannot meet the large-scale and high-safety operation requirements of energy storage systems. There is an urgent need for a more reliable energy storage safety early warning technology. Summary of the Invention

[0006] The purpose of this application is to provide a safety early warning method for energy storage systems based on multimodal perception, thereby improving the accuracy of safety early warning for energy storage systems.

[0007] The technical solution adopted in this application is as follows:

[0008] A safety early warning method for energy storage systems based on multimodal sensing includes the following steps:

[0009] S1. Deploy multimodal sensor arrays at key nodes of the energy storage system to collect multimodal data;

[0010] S2. Preprocess the multimodal data to obtain calibration data;

[0011] S3. Based on the calibration data, establish the normal fluctuation range for each mode using statistical methods; mark data with multiple consecutive sampling points (e.g., 3) that exceed the normal fluctuation range as abnormal segments;

[0012] S4. For abnormal segments, extract the feature vectors of each modality; construct a cross-modal dynamic attention network, calculate the contribution weight of each modality to the abnormal result through the attention mechanism based on the feature vectors of each modality, and perform weighted fusion of the feature vectors of each modality based on the contribution weight to obtain a weighted feature vector;

[0013] S5. Input the weighted feature vector into a bidirectional LSTM network, extract the temporal context features, and map them to the risk representation space to obtain risk representation results including multiple types of fault risk values.

[0014] S6. Input the risk characterization results into the early warning model constructed by the transfer learning framework; the early warning model performs online inference, calculates the similarity between the risk characterization results and the historical failure mode vectors, and if the similarity exceeds the preset threshold, the early warning model triggers an early warning.

[0015] In one possible implementation, in step S1, the multimodal sensor array includes temperature, high-frequency voltage, gas composition, and vibration acceleration sensors, which are used to collect temperature, high-frequency voltage, gas composition, and vibration acceleration data, respectively.

[0016] Each sensor uses a time synchronization protocol to unify its clock, ensuring that the time axis of the collected multimodal data is aligned, and providing a reference time anchor point for subsequent multimodal (multi-source) data fusion;

[0017] Furthermore, temperature sensors are installed on the surface of individual battery cells with an accuracy requirement of ±0.5℃, covering the positive and negative tabs of each battery module; a high-frequency voltage acquisition module is installed at the busbar with a fixed sampling frequency of 10kHz to capture millisecond-level voltage fluctuations; a gas composition detector is installed on the top of the chamber, with a detection limit of 1ppm for CO and H2, and the sampling point is located downstream of the battery compartment ventilation opening; a vibration accelerometer is arranged at the connection between the battery rack and the chamber, with a range of ±50g to cover the mechanical stress changes during charging and discharging; all sensors are synchronized with a clock via the NTP protocol, requiring a time deviation of less than 10ms to ensure the time axis alignment of the raw data and provide a unified time reference for subsequent multi-source data fusion; the collected raw data is packaged in 500ms increments and directly transmitted to the S2 module for preprocessing.

[0018] In one possible implementation, step S2, the preprocessing of the multimodal data, includes: noise cleaning, filtering, smoothing, and spatiotemporal alignment of the original data.

[0019] Furthermore, for the raw data collected by S1, a sliding window algorithm is first used for noise cleaning, with a window length set to 2 seconds. Impulse noise with amplitudes exceeding 5 times the standard deviation of the mean in the voltage data is removed. Temperature and vibration data are smoothed using Kalman filtering, with the process noise covariance matrix set to a diagonal matrix of 0.01 to balance real-time performance and smoothing effect. Due to the response delay of the gas sensor, its time axis needs to be shifted back by 0.8 seconds using the cross-correlation function method to achieve spatiotemporal alignment with the voltage and temperature data. The calibrated data must meet the quality requirement of a signal-to-noise ratio ≥20dB, with temperature data fluctuation ≤0.2℃ / s and voltage ripple ≤0.05V. These are passed to the S3 module as key indicators for confidence assessment of anomaly feature screening.

[0020] Further, step S3 specifically includes: based on the calibration data output from S2, establishing normal fluctuation ranges for each mode using a statistical process control method: temperature is based on 25℃, allowing fluctuations of ±5℃; unit voltage is centered at 3.2V, allowing deviations of ±0.1V; the upper limit of CO concentration is set at 5ppm; the effective value of vibration acceleration does not exceed 0.1g; when multiple consecutive sampling points of a certain mode data exceed the above ranges, it is marked as an "abnormal segment," and the abnormal start time, mode type, and deviation amplitude are recorded (e.g., the deviation is +17℃ at a temperature of 42℃). The selected abnormal segments are input as regions of interest into module S4 to reduce the computational load of subsequent deep feature extraction, while retaining complete data for a certain period before and after the anomaly for traceability.

[0021] In one possible implementation, in step S4, for temperature data, the gradient change rate at each time moment is calculated using a sliding window to construct a feature vector; for voltage data, the ripple coefficient (proportion of high-frequency components) at each time moment is extracted using Fourier transform to construct a feature vector; for gas data, the concentration growth rate at each time moment is calculated using differential operation to construct a feature vector; and for vibration data, the energy proportion (percentage of total energy) of a preset frequency band at each time moment is calculated using fast Fourier transform to construct a feature vector.

[0022] In one possible implementation, in step S4, the cross-modal dynamic attention network assigns a dynamic score to each modality, which is determined by the cosine similarity between the modality feature vector and the feature vector of the historical anomalous pattern: if the current feature vector of a certain modality has a higher similarity to the feature vector of that modality in historical thermal runaway cases, its score is improved.

[0023] The modal scores are converted into contribution weights through normalization. The sum of the contribution weights of each modality is 1. The larger the contribution weight of a modality, the greater its impact on subsequent risk calculations. For example, when the gas concentration growth rate reaches 1.2 ppm / s (exceeding the 95th percentile of historical normal data), the gas modality score is significantly higher than other modalities, and its contribution weight can be increased from the usual 0.25 to 0.6. At this time, the network will focus on the correlation between gas data and anomalies, while reducing the contribution weight of other modalities such as voltage.

[0024] The dynamic weights are calculated using the following innovative formula:

[0025]

[0026] In this formula, For the first Contribution weights of different modes and For modal type index, ,For example, =4, the four modes are temperature, voltage, gas, and vibration; and For the first Class modality and the first The dynamic score of a modality is determined by the cosine similarity between the current modality feature and the historical abnormal pattern feature (the higher the similarity, the greater the dynamic score); exp(·) is an exponential function used to amplify the differences in scores between different modalities; the denominator is the sum of the exponential scores of all modalities to ensure that the sum of the weights of each modality is 1.

[0027] In one possible implementation, in step S4, based on the contribution weight of each mode to the abnormal result, the sensor sampling frequency corresponding to S1 is adjusted accordingly. Specifically:

[0028] For modes with higher contribution weights, increase the sampling frequency of their corresponding sensors; for modes with lower contribution weights, decrease the sampling frequency of their corresponding sensors; thus achieving adaptive sensing where the more significant the anomaly, the denser the sampling.

[0029] In one possible implementation, in step S5, the weighted feature vector output from S4 is input into a bidirectional LSTM network to extract time series context features. The features processed by the LSTM are mapped to a multi-dimensional risk representation space through a fully connected layer to obtain risk representation results. Each dimension of the risk representation results corresponds to the risk value of a typical fault of a different category, with a value range of 0 to 1, where 0 represents no risk and 1 represents certain occurrence. The fused risk representation results are updated periodically at a preset time and serve as the input labels for the early warning model in S6, directly affecting the accuracy of subsequent early warning inference.

[0030] In one possible implementation, in step S6, the early warning model is constructed using a transfer learning framework, specifically including: in the pre-training stage, historical fault data is used to train the initial model through the cross-entropy loss function; in the online stage, incremental learning is performed with a preset learning rate (e.g., 0.001) to update the model parameters in real time.

[0031] Furthermore, during the training of the early warning model, the weight matrix of its fully connected layer is fed back to S4 every preset time (e.g., every hour) to optimize the weight calculation logic of the attention mechanism and ensure the network's adaptability to new abnormal patterns.

[0032] In one possible implementation, the method further includes:

[0033] S7. Combining the operating status of the energy storage system with seasonal environmental parameters, the fuzzy logic controller dynamically adjusts the early warning threshold according to the threshold rule table.

[0034] In step S7, the warning threshold is dynamically adjusted by a fuzzy logic controller based on the threshold rule table, taking into account the operating status of the energy storage system and seasonal environmental parameters: in the charging state, the thermal runaway risk threshold is reduced; in the quiescent state, the internal short circuit risk threshold is increased.

[0035] The threshold adjustment rule table contains multiple state combinations, which are updated by the historical warning accuracy output by S6. It requires the accuracy to always be maintained above the preset accuracy to ensure that the threshold adaptability matches the actual operating scenario.

[0036] S8. If the risk value in the risk characterization results exceeds the dynamically adjusted early warning threshold, generate composite early warning information including abnormal modes, risk levels, and development trends, and push it through multiple channels.

[0037] Furthermore, the abnormal modality labeling is the dominant abnormal modality; the risk levels include three levels: Level I alert (corresponding to risk value 0.5-0.7), Level II alarm (corresponding to risk value 0.7-0.9), and Level III emergency (corresponding to risk value > 0.9); the development trend refers to the change in risk value over a future period predicted by the ARIMA model. The input to the ARIMA model can be a historical risk value time series, which is formed by the risk characterization results in chronological order; the multi-channel push includes: synchronously pushing early warning information to the operation and maintenance APP, monitoring screen, and SMS platform through the edge computing terminal; the operation and maintenance APP displays key parameters in real-time pop-up windows; the monitoring screen uses red flashing markers to locate the abnormal location; and the SMS platform sends a risk level and trend summary, along with the original data curve of the abnormal segment in S4 for clicking to view.

[0038] In one possible implementation, the method further includes:

[0039] S9. Collect confirmation feedback from maintenance personnel regarding the early warning results. Based on false alarms or missed alarms, optimize the normal fluctuation range of the corresponding mode in S3 and the training parameters of the early warning model in S6 to form a closed-loop optimization.

[0040] For example, the system continuously collects confirmation actions from maintenance personnel regarding early warnings and compares the manually labeled results with the output of the S6 model. If there are multiple false alarms, the normal fluctuation range of the corresponding mode in S3 is adjusted. If there are missed alarms, the training sample weights for that fault type in the early warning model in S6 are increased. The optimized normal fluctuation range and sample weights are fed back to S3 and S6, forming a complete closed loop of "data collection - feature extraction - model training - early warning push - effect optimization", which continuously improves the system's ability to identify complex abnormal scenarios.

[0041] In summary, due to the adoption of the above technical solution, the beneficial effects of this application are:

[0042] 1. In this application, through a cross-modal dynamic attention network, the system can dynamically adjust the weights of each sensor modality based on real-time data. When an abnormal increase occurs in a certain modality, its proportion in risk calculation is automatically increased, focusing on capturing the abnormal signal of that modality and avoiding interference from other normally fluctuating modal data. This "on-demand focusing" mechanism allows the system to identify the true source of risk more quickly, reducing false alarms and missed alarms. Compared with traditional fixed-weight multimodal fusion methods, it is better at capturing key abnormal features, significantly improving the accuracy and timeliness of energy storage system safety early warning.

[0043] 2. In this application, a self-feedback optimization mechanism dynamically adjusts the anomaly screening range and early warning model training parameters of S3 based on actual feedback from maintenance personnel. If false alarms caused by ambient temperature fluctuations occur continuously, the system will automatically widen the normal fluctuation range of the temperature mode; if an internal short-circuit fault is missed, the training sample weight of that fault type will be increased. This closed-loop design of "using actual operating data to feed back optimization" allows the system to gradually adapt to the characteristics of energy storage systems under different seasons and operating conditions. Over long-term use, the early warning effect will become increasingly aligned with actual needs, avoiding the accuracy decline problem caused by fixed parameters in traditional methods, and enhancing the system's adaptability to complex scenarios and long-term reliability. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the process principle of one embodiment of this application. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] Reference Figure 1 This application provides a safety early warning method for energy storage systems based on multimodal sensing, comprising the following steps:

[0047] S1. Deploy multimodal sensor arrays at key nodes of the energy storage system to collect multimodal data.

[0048] In some embodiments, in step S1, the multimodal sensor array includes temperature, high-frequency voltage, gas composition, and vibration acceleration sensors, which are used to collect temperature, high-frequency voltage, gas composition, and vibration acceleration data, respectively.

[0049] Each sensor uses a time synchronization protocol to unify its clock, ensuring that the time axis of the collected multimodal data is aligned, and providing a reference time anchor point for subsequent multimodal (multi-source) data fusion;

[0050] Furthermore, temperature sensors are installed on the surface of individual battery cells with an accuracy requirement of ±0.5℃, covering the positive and negative tabs of each battery module; a high-frequency voltage acquisition module is installed at the busbar with a fixed sampling frequency of 10kHz to capture millisecond-level voltage fluctuations; a gas composition detector is installed on the top of the chamber, with a detection limit of 1ppm for CO and H2, and the sampling point is located downstream of the battery compartment ventilation opening; a vibration accelerometer is arranged at the connection between the battery rack and the chamber, with a range of ±50g to cover the mechanical stress changes during charging and discharging; all sensors are synchronized with a clock via the NTP protocol, requiring a time deviation of less than 10ms to ensure the time axis alignment of the raw data and provide a unified time reference for subsequent multi-source data fusion; the collected raw data is packaged in 500ms increments and directly transmitted to the S2 module for preprocessing.

[0051] S2. Preprocess the multimodal data to obtain calibration data.

[0052] In some embodiments, step S2, the preprocessing of multimodal data, includes: noise cleaning, filtering and smoothing, and spatiotemporal alignment of the original data.

[0053] Furthermore, for the raw data collected by S1, a sliding window algorithm is first used for noise cleaning, with a window length set to 2 seconds. Impulse noise with amplitudes exceeding 5 times the standard deviation of the mean in the voltage data is removed. Temperature and vibration data are smoothed using Kalman filtering, with the process noise covariance matrix set to a diagonal matrix of 0.01 to balance real-time performance and smoothing effect. Due to the response delay of the gas sensor (typically 0.8 seconds), its time axis needs to be shifted back by 0.8 seconds using the cross-correlation function method to achieve spatiotemporal alignment with the voltage and temperature data. The calibrated data must meet the quality requirement of a signal-to-noise ratio ≥20dB, with temperature data fluctuation ≤0.2℃ / s and voltage ripple ≤0.05V. These are passed to the S3 module as key indicators for confidence assessment of anomaly feature screening.

[0054] S3. Based on the calibration data, establish the normal fluctuation range for each mode using statistical methods; mark data with multiple consecutive sampling points (e.g., 3) that exceed the normal fluctuation range as abnormal segments.

[0055] The abnormal segment is a set of continuous time-series multimodal calibration data.

[0056] Further, step S3 specifically includes: based on the calibration data output from S2, establishing normal fluctuation ranges for each mode using a statistical process control method: temperature is based on 25℃, allowing fluctuations of ±5℃; unit voltage is centered at 3.2V, allowing deviations of ±0.1V; the upper limit of CO concentration is set at 5ppm; the effective value of vibration acceleration does not exceed 0.1g; when multiple consecutive sampling points of a certain mode data (e.g., 3 sampling points, i.e., within 1.5 seconds) exceed the above range, it is marked as an "abnormal segment," and the abnormal start time, mode type, and deviation amplitude are recorded (e.g., deviation of +17℃ at a temperature of 42℃). The selected abnormal segments are input as regions of interest into module S4 to reduce the computational load of subsequent deep feature extraction, while retaining complete data for a certain period of time (e.g., 2 seconds) before and after the anomaly for traceability.

[0057] S4. For abnormal segments, extract the feature vectors of each modality; construct a cross-modal dynamic attention network, calculate the contribution weight of each modality to the abnormal result through the attention mechanism based on the feature vectors of each modality, and perform weighted fusion of the feature vectors of each modality based on the contribution weight to obtain a weighted feature vector.

[0058] In some embodiments, in step S4, for temperature data, the gradient change rate (unit: °C / s) at each time moment is calculated using a sliding window to construct a feature vector; for voltage data, the ripple coefficient (proportion of high-frequency components) at each time moment is extracted using Fourier transform to construct a feature vector; for gas data, the concentration growth rate (unit: ppm / s) at each time moment is calculated using difference operations to construct a feature vector; for vibration data, the energy proportion (percentage of total energy) in the 100–500 Hz frequency band at each time moment is calculated using fast Fourier transform to construct a feature vector. The feature vectors for the four modes are denoted as follows: .

[0059] In some embodiments, in step S4, the cross-modal dynamic attention network assigns a dynamic score to each modality, which is determined by the cosine similarity between the modality feature vector and the historical anomalous pattern feature vector: if the current feature vector of a certain modality has a higher similarity to the feature vector of that modality in historical thermal runaway cases, its score is improved;

[0060] The modal scores are converted into contribution weights through normalization. The sum of the contribution weights of each modality is 1. The larger the contribution weight of a modality, the greater its impact on subsequent risk calculations. For example, when the gas concentration growth rate reaches 1.2 ppm / s (exceeding the 95th percentile of historical normal data), the gas modality score is significantly higher than other modalities, and its contribution weight can be increased from the usual 0.25 to 0.6. At this time, the network will focus on the correlation between gas data and anomalies, while reducing the contribution weight of other modalities such as voltage.

[0061] The dynamic weights are calculated using the following innovative formula:

[0062]

[0063] In this formula, For the first Contribution weights of different modes and For modal type index, ,For example, =4, the four modes are temperature, voltage, gas, and vibration; and For the first Class modality and the first The dynamic score of a modality is determined by the cosine similarity between the current modality feature and the historical abnormal pattern feature (the higher the similarity, the greater the dynamic score); exp(·) is an exponential function used to amplify the differences in scores between different modalities; the denominator is the sum of the exponential scores of all modalities to ensure that the sum of the weights of each modality is 1.

[0064] The innovation of this formula lies in achieving nonlinear amplification of the "abnormally dominant mode" through exponential normalization (i.e., the softmax function)—when the score of a certain mode is only 1 higher than that of other modes, its weight will be higher than that of linear normalization (such as directly taking the value of 1). The abnormal dominant mode is about 2.7 times higher, thus highlighting the key role of the abnormal dominant mode in risk assessment.

[0065] In some embodiments, in step S4, based on the contribution weight of each modality to the abnormal result, the sensor sampling frequency corresponding to S1 is adjusted accordingly. Specifically:

[0066] For modes with higher contribution weights, increase the sampling frequency of their corresponding sensors; for modes with lower contribution weights, decrease the sampling frequency of their corresponding sensors; thus achieving adaptive sensing where the more significant the anomaly, the denser the sampling.

[0067] S5. Input the weighted feature vector into a bidirectional LSTM network, extract the temporal context features, and map them to the risk representation space to obtain risk representation results including multiple types of fault risk values.

[0068] In this context, each dimension of the risk characterization result corresponds to a risk value for different fault types.

[0069] Further, in step S5, the weighted feature vector output from S4 is input into a bidirectional LSTM network with a hidden layer dimension of 64 to extract time-series context features (such as the implicit temporal evolution patterns of a continuously deteriorating fusion state and a step-like increase in risk level). The features processed by the bidirectional LSTM network are mapped to an N-dimensional (e.g., 8-dimensional) risk representation space through a fully connected layer. Each dimension corresponds to the risk value of 8 typical faults such as thermal runaway, internal short circuit, and insulation failure, with a value range of 0 to 1 (0 indicates no risk, and 1 indicates inevitable occurrence). The fused risk representation result is updated at a preset time (e.g., 1 second) and serves as the input label for the real-time early warning model of module S6, directly affecting the accuracy of subsequent early warning inference.

[0070] Among them, the fully connected layer refers to one or more fully connected layers after the bidirectional LSTM, which are used to map the temporal features to the risk representation space. It is a key bridge from feature extraction (LSTM) to risk quantification (risk representation vector), and its weight matrix directly encodes the information of "which input features are most important for judging which fault risk".

[0071] Furthermore, during model training, the weight matrix of the fully connected layer in S5, which connects the bidirectional LSTM network to the risk representation space, is analyzed at preset intervals (e.g., every hour) to generate an input feature importance vector, which is then fed back to the S4 module. Based on this feedback vector, the S4 module adjusts the bias parameters for calculating the dynamic scores of each modality in its cross-modal dynamic attention network, thereby optimizing the attention weight calculation logic and ensuring that the network can continuously learn and adapt to new abnormal patterns.

[0072] Specifically, the dimension of the weight matrix W of the fully connected layer is [the feature dimension after processing by the bidirectional LSTM network, N], which can be analyzed in its reverse direction (from risk type back to input features). By calculating the gradient of each type of failure risk (e.g., "thermal runaway") with respect to each input feature, a feature importance vector is summarized. Although this feature importance vector does not directly correspond to a certain original mode, it reflects which abstract features the model in S5 considers most critical for the final risk judgment. S4 can maintain a learnable mapping matrix, converting the feature importance vector fed back from S5 into bias terms for different modes. The calculation logic for dynamic scores of each mode is adjusted. Based on the updated bias term, S4's cross-modal attention mechanism will proactively increase attention to key modes. For example, in new anomalous events, even if the gas mode has obvious anomalous features, attention will prioritize the vibration mode that is more decisive for fault mode identification, thereby capturing new fault modes faster and more accurately.

[0073] S6. Input the risk characterization results into the early warning model constructed by the transfer learning framework; the early warning model performs online inference, calculates the similarity (such as cosine similarity) between the risk characterization results and the historical failure mode vectors, and if the similarity exceeds the preset threshold (such as 0.85), the early warning model triggers an early warning.

[0074] In some embodiments, in step S6, the early warning model is constructed using a transfer learning framework, specifically including: in the pre-training stage, historical fault data (such as 1000 thermal runaway cases and 500 internal short circuit cases) is used to train the initial model through the cross-entropy loss function; in the online stage, incremental learning is performed with a preset learning rate (such as 0.001) to update the model parameters in real time.

[0075] In some embodiments, the method further includes:

[0076] S7. Combining the operating status of the energy storage system with seasonal environmental parameters, the fuzzy logic controller dynamically adjusts the early warning threshold according to the threshold rule table.

[0077] In step S7, the warning threshold is dynamically adjusted by a fuzzy logic controller based on the energy storage system's operating status (charging, discharging, and resting) and seasonal environmental parameters (e.g., a summer ambient temperature of 30°C). Under charging conditions, the thermal runaway risk threshold is lowered (e.g., from the usual 0.6 to 0.5 to improve sensitivity); under resting conditions, the internal short circuit risk threshold is raised (e.g., from 0.7 to 0.75 to reduce false alarms).

[0078] The threshold adjustment rule table contains a variety of state combinations (such as 20 state combinations, including winter discharge, summer charging, etc.). It is updated by the historical warning accuracy output by S6, and the accuracy is required to always be maintained above the preset accuracy (such as 92%) to ensure that the threshold adaptability matches the actual operation scenario.

[0079] S8. If the risk value in the risk characterization results exceeds the dynamically adjusted early warning threshold, generate composite early warning information including abnormal modes, risk levels, and development trends, and push it through multiple channels.

[0080] Furthermore, the abnormal mode labeling is dominated by the abnormal mode (e.g., "gas (CO concentration growth rate 1.2 ppm / s) + temperature (42℃ for 5 minutes)"); the risk levels include three levels: Level I alert (corresponding to risk value 0.5-0.7), Level II alarm (corresponding to risk value 0.7-0.9), and Level III emergency (corresponding to risk value > 0.9); the development trend refers to the risk value change predicted by the ARIMA model over a future period of time (e.g., 10 minutes) (e.g., "the risk value is expected to rise to 0.85 in 10 minutes"). The input to the ARIMA model can be a historical risk value time series, which is formed by the risk characterization results in chronological order; the multi-channel push includes: synchronously pushing early warning information to the operation and maintenance APP, monitoring screen, and SMS platform through the edge computing terminal; the operation and maintenance APP displays key parameters in real-time pop-up windows; the monitoring screen uses red flashing markers to locate the abnormal location; and the SMS platform sends a risk level and trend summary, along with the original data curve of the abnormal segment in S4 for clicking to view.

[0081] In some embodiments, the method further includes:

[0082] S9. Collect confirmation feedback from maintenance personnel regarding the early warning results. Based on false alarms or missed alarms, optimize the normal fluctuation range of the corresponding mode in S3 and the training parameters of the early warning model in S6 to form a closed-loop optimization.

[0083] For example, the system continuously collects confirmation actions from maintenance personnel regarding early warnings (such as "confirming thermal runaway risk" or "marking as false alarm"), and compares the manually labeled results with the output of the S6 model: if there are multiple (3) consecutive false alarms (such as model output > 0.7 but no actual risk), the normal fluctuation range of the corresponding mode in S3 is adjusted (such as relaxing the upper limit of CO concentration from 5ppm to 7ppm); if there is a missed alarm (such as model output < 0.5 but actual risk exists), the training sample weight of the fault type in the early warning model in S6 is increased (such as increasing the sample weight of internal short circuit from 1.0 to 1.5). The optimized normal fluctuation range and sample weights are fed back to S3 and S6, forming a complete closed loop of "data collection - feature extraction - model training - early warning push - effect optimization", continuously improving the system's ability to identify complex abnormal scenarios.

[0084] In this application, a cross-modal dynamic attention network is used to dynamically adjust the weights of each sensor modality based on real-time data. When an abnormal increase occurs in a certain modality, its proportion in risk calculation is automatically increased to focus on capturing the abnormal signal of that modality and avoid interference from other normally fluctuating modal data. This "on-demand focusing" mechanism allows the system to identify the true source of risk more quickly, reducing false alarms and missed alarms. Compared with traditional fixed-weight multimodal fusion methods, it is better at capturing key abnormal features and significantly improves the accuracy and timeliness of safety early warning for energy storage systems.

[0085] In this application, the self-feedback optimization mechanism dynamically adjusts the anomaly screening range and model training parameters of S3 based on actual feedback from maintenance personnel. If false alarms caused by ambient temperature fluctuations occur continuously, the system will automatically widen the normal fluctuation range of the temperature mode; if an internal short-circuit fault is missed, the weight of the training sample for that fault type will be increased. This closed-loop design of "using actual operating data to feed back optimization" allows the system to gradually adapt to the characteristics of energy storage systems under different seasons and operating conditions. Over long-term use, the early warning effect will become increasingly aligned with actual needs, avoiding the accuracy decline problem caused by fixed parameters in traditional methods, and enhancing the system's adaptability to complex scenarios and long-term reliability.

[0086] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0087] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A safety early warning method for energy storage systems based on multimodal sensing, characterized in that, Includes the following steps: S1. Deploy multimodal sensor arrays at key nodes of the energy storage system to collect multimodal data; S2. Preprocess the multimodal data to obtain calibration data; S3. Based on calibration data, establish normal fluctuation ranges for each mode using statistical methods; Data with multiple consecutive sampling points exceeding the normal fluctuation range are marked as abnormal segments; S4. For abnormal segments, extract the feature vectors of each modality respectively; A cross-modal dynamic attention network is constructed. Based on the feature vectors of each modality, the contribution weight of each modality to the abnormal results is calculated through the attention mechanism. Based on the contribution weights, the feature vectors of each modality are weighted and fused to obtain a weighted feature vector. S5. Input the weighted feature vector into a bidirectional LSTM network, extract the temporal context features, and map them to the risk representation space to obtain risk representation results including multiple types of fault risk values. S6. Input the risk characterization results into the early warning model constructed by the transfer learning framework; The early warning model performs online inference, calculates the similarity between the risk characterization results and the historical failure mode vectors, and triggers an early warning if the similarity exceeds a preset threshold.

2. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, In step S1, the multimodal sensor array includes temperature, high-frequency voltage, gas composition, and vibration acceleration sensors, which are used to collect temperature, high-frequency voltage, gas composition, and vibration acceleration data, respectively. Each sensor uses a time synchronization protocol to unify its clock, ensuring that the time axis of the acquired multimodal data is aligned.

3. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 2, characterized in that, In step S2, the preprocessing of multimodal data includes: noise cleaning, filtering and smoothing, and spatiotemporal alignment of the original data.

4. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 3, characterized in that, In step S4, for temperature data, the gradient change rate at each time moment is calculated using a sliding window to construct a feature vector; for voltage data, the ripple coefficient at each time moment is extracted using Fourier transform to construct a feature vector; for gas data, the concentration growth rate at each time moment is calculated using difference operation to construct a feature vector; and for vibration data, the energy proportion of a preset frequency band at each time moment is calculated using fast Fourier transform to construct a feature vector.

5. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, In step S4, the cross-modal dynamic attention network assigns a dynamic score to each modality, which is determined by the cosine similarity between the modality feature vector and the historical abnormal pattern feature vector. The modality scores are converted into contribution weights through a normalization operation, using the following formula: ； In this formula, For the first Contribution weights of different modes and For modal type index, , and For the first Class modality and the first The dynamic score of a modality is determined by the cosine similarity between the current modality feature and the historical abnormal pattern feature; exp(·) is an exponential function used to amplify the differences in scores between different modalities; the denominator is the sum of the exponential scores of all modalities to ensure that the sum of the weights of each modality is 1.

6. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, In step S4, based on the contribution weight of each mode to the abnormal result, the sensor sampling frequency corresponding to S1 is adjusted accordingly. Specifically: For modes with higher contribution weights, increase the sampling frequency of their corresponding sensors; for modes with lower contribution weights, decrease the sampling frequency of their corresponding sensors.

7. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, In step S5, the weighted feature vector output from S4 is input into a bidirectional LSTM network to extract time series context features. The features processed by LSTM are mapped to a multi-dimensional risk representation space through a fully connected layer to obtain the risk representation result. Each dimension of the risk representation result corresponds to the risk value of a different category of typical faults, with a value range of 0 to 1, where 0 indicates no risk and 1 indicates that it will definitely happen. The fused risk characterization results are updated at preset time intervals and used as input labels for the early warning model in S6.

8. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, In step S6, the early warning model is constructed using a transfer learning framework, which specifically includes: in the pre-training stage, historical fault data is used to train the initial model through the cross-entropy loss function; in the online stage, incremental learning is performed with a preset learning rate to update the model parameters in real time. During the training of the early warning model, the weight matrix of its fully connected layer is fed back to step S4 every preset time to optimize the weight calculation logic of the attention mechanism.

9. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, The method further includes: S7. Combining the operating status of the energy storage system with seasonal environmental parameters, the fuzzy logic controller dynamically adjusts the early warning threshold according to the threshold rule table. S8. If the risk value in the risk characterization results exceeds the dynamically adjusted early warning threshold, generate composite early warning information including abnormal modes, risk levels, and development trends, and push it through multiple channels.

10. The method for safety early warning of energy storage systems based on multimodal sensing as described in claim 1, characterized in that, The method further includes: S9. Collect confirmation feedback from maintenance personnel regarding the early warning results. Based on false alarms or missed alarms, optimize the normal fluctuation range of the corresponding mode in S3 and the training parameters of the early warning model in S6 to form a closed-loop optimization.

Images