A method and system for monitoring and early warning of thermal runaway of an immersed energy storage battery pack in multiple modes

Abstract

The application discloses a kind of methods and systems for monitoring and early warning of thermal runaway of submerged energy storage battery pack, relating to the technical field of battery thermal runaway early warning, comprising S1, in the internal of submerged energy storage battery pack, thin film pressure-temperature composite sensor, infrared spectrum probe, top multi-in-one gas sensor are arranged synchronously, and combined with battery management system, real-time collection of electric parameters, temperature, characteristic gas, stress, spectral multidimensional physical field time series data;S2, based on the multidimensional physical field time series data, the mapping database of the degradation physical mechanism, ultrasonic acoustic characteristic parameter and multidimensional online sensor data is established by using offline ultrasonic detection, and the multidimensional physical field time series data of each time window is labeled with degradation level label and attached with ultrasonic velocity, attenuation value and characteristic waveform as auxiliary reference features.The application solves the problem of insufficient multidimensional perception of early degradation signal in submerged closed environment and the difficulty of deep integration of offline ultrasonic knowledge and online early warning.

Description

Technical Field

[0001] This invention relates to the field of battery thermal runaway early warning technology, and in particular to a multimodal monitoring and early warning method and system for thermal runaway of an immersed energy storage battery pack. Background Technology

[0002] With the increasing demands for energy density and safety in electrochemical energy storage power stations, immersion liquid cooling thermal management technology has been widely adopted in large-scale energy storage battery packs due to its excellent temperature uniformity and heat dissipation efficiency. For this special structure, the industry has developed various thermal runaway monitoring solutions, among which online early warning technology based on multi-sensor fusion is a common approach. This typically involves deploying temperature, voltage, current, and characteristic gas sensors within the battery pack, combined with the battery management system to collect multi-dimensional operational data in real time. Subsequently, machine learning algorithms are used to comprehensively assess the risk of thermal runaway. This method, to a certain extent, overcomes the limitations of single-parameter early warning, improving the accuracy of judgment through the complementarity of multi-source data, and has become one of the important technical routes in the current field of energy storage safety monitoring.

[0003] Existing multi-sensor fusion solutions for use in submerged energy storage battery packs suffer from limitations in sensing dimensions. On one hand, temperature and characteristic gas fluctuations in the immersion environment are small; on the other hand, there is a lack of methods for synchronously monitoring microscopic mechanical deformation within the cells and changes in the physicochemical properties of the immersion liquid. If these weak signals are not effectively captured, they will directly affect the lead time for thermal runaway warnings. Although ultrasonic detection technology can reflect changes in the modulus and interface of materials within the battery and has explanatory power for degradation mechanisms such as lithium plating and gas generation, its deployment in online monitoring systems is currently difficult due to limitations in the cost of ultrasonic sensors, the complexity of acoustic coupling in the liquid environment, and the difficulty of online signal processing. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-modal monitoring and early warning method for thermal runaway of immersion energy storage battery packs, which solves the problems of insufficient multi-dimensional perception of early degradation signals in immersion sealed environments and the inability to deeply integrate offline ultrasonic knowledge for online early warning.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for monitoring and early warning of thermal runaway in an immersion energy storage battery pack in multiple modes, comprising, S1, simultaneously arranging a thin-film pressure-temperature composite sensor, an infrared spectral probe, and a top multi-in-one gas sensor inside the immersion energy storage battery pack and combining them with a battery management system to collect real-time time-series data of electrical parameters, temperature, characteristic gases, stress, and spectral multi-dimensional physical field. S2, based on the multidimensional physical field time series data, an offline ultrasonic detection is used to establish a mapping database of degradation physical mechanism, ultrasonic acoustic characteristic parameters and multidimensional online sensor data, and a degradation level label is added to the multidimensional physical field time series data of each time window, and ultrasonic velocity, attenuation value and characteristic waveform are attached as auxiliary reference features. S3. Based on the mapping database, train a multimodal temporal Transformer model with degradation level as the main task label and multidimensional online sensor and battery management system data as auxiliary supervision signals, so that the model can learn implicit representations related to the changes in multidimensional physical field temporal signals from ultrasonic acoustic feature data. S4 deploys the trained multimodal temporal Transformer model on an edge computing gateway after knowledge distillation and quantization compression. It performs online inference by collecting complete multidimensional physical field temporal data once per second and maintaining a sliding time window, outputting a degradation level probability vector and obtaining a risk score. Based on the set multi-level warning thresholds, it triggers corresponding graded warning signals.

[0007] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, wherein: in step S1, the thin-film pressure-temperature composite sensor is attached between the large surfaces of adjacent cells, with one thin-film pressure-temperature composite sensor arranged every three cells to sense the expansion / contraction strain and bulk temperature in the thickness direction of the cells; the infrared spectroscopy probe is arranged in the outlet area of ​​the battery pack immersion liquid circulation system to characterize the physicochemical deterioration of the immersion liquid through the change of absorption peaks at specific wavenumbers; and the top multi-functional gas sensor is used to monitor the concentration of characteristic gases and the gas phase temperature.

[0008] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, the infrared spectral probe operates in a band covering 2.5~5μm. As an economically efficient in-situ sensing method, it reflects trace amounts of moisture or electrolyte decomposition products by monitoring the absorption peak near 2.8μm, and reflects abnormal gas production and dissolution by monitoring the absorption peak near 4.3μm.

[0009] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, wherein: in step S2, the step of establishing a mapping database using offline ultrasonic detection includes: using a high-precision ultrasonic scanning system as a non-in-situ detection method, building an offline equivalent test platform of the same batch as the online battery pack, and simultaneously deploying a thin-film pressure sensor, an infrared spectroscopy probe, and a top-mounted multi-functional gas sensor in a sealed laboratory container to simulate the degradation process of the submerged energy storage battery pack under typical operating conditions and typical abnormal conditions such as overcharging, thermal abuse, and electrical crosstalk, and introducing high-precision ultrasonic scanning... The system was checked; a typical abnormal operating condition degradation experiment was performed, and all online sensor data were recorded synchronously at the first sampling frequency. At the same time, ultrasonic velocity, attenuation coefficient and backscatter signal were recorded through ultrasonic equipment at the second sampling frequency. Combining ultrasonic scan images, acoustic parameter mutation points and post-disassembly analysis, the battery degradation process was divided into multiple degradation levels, and the corresponding degradation level label was marked on the multidimensional physical field time series data of each time window. Ultrasonic velocity and attenuation value were attached as auxiliary reference features to construct a dataset of immersion battery thermal runaway time series feature signals that integrates in-situ multidimensional signals and non-in-situ ultrasonic features.

[0010] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, the degradation levels include at least Level 0 (normal), Level 1 (SEI decomposition / trace lithium plating), Level 2 (significant lithium plating / small amount of gas production), Level 3 (large amount of gas production / electrolyte decomposition), and Level 4 (thermal runaway outbreak). The accurate identification of Level 1 and Level 2 degradation states is used to achieve very early warning of thermal runaway.

[0011] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, wherein: in step S3, the input of the multimodal time-series Transformer model is multidimensional physical field time-series data within a preset time window, including at least one of the following: film pressure (mean / variance), infrared absorbance ratio, pack pressure, characteristic gas concentration, tab temperature, gas phase temperature, battery bulk temperature, battery terminal voltage, current, SOC, and internal resistance estimation; the model output includes a main task and an auxiliary task, wherein the main task is used to predict the probability distribution of degradation level within a future time window, and the auxiliary task is used to regress and predict the corresponding ultrasonic acoustic features. Through joint iterative training of massive in-situ sensor data and corresponding non-in-situ ultrasonic labeled data, the model learns the implicit mapping relationship from multidimensional online time-series signals to the internal degradation state of the battery, thereby improving the robustness and generalization ability of the model; the loss function during training is the sum of the weighted cross-entropy loss of the main task and the mean square error loss of the auxiliary task.

[0012] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, in step S4, the risk score is defined as the weighted sum of the probabilities of the degradation level reaching a preset level, and a multi-level early warning threshold is set according to the risk score. When the risk score exceeds the corresponding threshold, an early warning signal of the corresponding level is triggered, and the early warning signal is uploaded to the monitoring system and cloud platform to trigger a safety linkage strategy.

[0013] The risk scoring expression is: ; in, To score the risk, This is the highest deterioration level number. For discrete variables representing the degree of degradation, The lower threshold of the preset level, To correspond to the degradation level The preset weighting coefficients, Currently at a degraded level The probability of.

[0014] Deterioration level Preset weighting coefficients The expression is:

[0015] in, To preset the baseline weights, This represents the absolute value of the temperature change. This represents the absolute value of the voltage change. This represents the absolute value of the pressure change between battery cells. This represents the absolute value of the pressure change within the package. The deviation from the state of charge. The characteristic value is the concentration of the characteristic gas. , , , , τ and τ are both pre-calibrated adaptive adjustment coefficients.

[0016] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, the multi-level early warning thresholds include: a risk score <0.2 for safe, 0.2 to 0.4 for concern, 0.4 to 0.7 for warning, and >0.7 for danger; before the early warning is triggered, a false alarm prevention strategy is also included, which confirms the warning for multiple consecutive inference cycles to improve the reliability of the system's early warning under complex operating conditions.

[0017] As a preferred embodiment of the multimodal monitoring and early warning method for thermal runaway of the submerged energy storage battery pack described in this invention, in step S4, the trained multimodal time-series Transformer model is subjected to knowledge distillation and INT8 quantization compression, and deployed at an edge computing gateway next to the battery pack. This eliminates the need for ultrasonic detection equipment during online operation. During online operation, complete multidimensional physical field time-series data is collected once per second, a 300-second sliding time window is maintained, and forward inference is performed every 60 data points (1 minute) to output a degradation level probability vector. This effectively reduces the platform's computational load and inference pressure while ensuring monitoring timeliness, and achieves very early warning of thermal runaway based solely on in-situ sensors and BMS data.

[0018] Secondly, this invention provides a multimodal monitoring and early warning system for thermal runaway in an immersed energy storage battery pack, comprising a sensor module, a mapping database module, a model training module, and an edge computing gateway module. The sensor module is used to synchronously arrange a thin-film pressure-temperature composite sensor, an infrared spectroscopy probe, and a top-mounted multi-functional gas sensor between the large surfaces of the cells inside the immersed energy storage battery pack, combined with a battery management system, serving as a low-cost, long-term operational in-situ sensing front-end to collect multidimensional physical field time-series data in real time. The mapping database module is used to store the mapping relationship between the degradation physical mechanism, ultrasonic acoustic characteristic parameters, and multidimensional online sensor data established by offline ultrasonic detection. The model training module is used to train a multimodal time-series Transformer model based on the mapping database, with degradation level as the main task label and ultrasonic acoustic characteristics as the auxiliary supervision signal. The edge computing gateway module is used to deploy the multimodal time-series Transformer model after knowledge distillation and quantization compression, and perform online inference and hierarchical early warning.

[0019] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the immersion energy storage battery pack thermal runaway multimodal monitoring and early warning method as described in the first aspect of the present invention.

[0020] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the immersion energy storage battery pack thermal runaway multimodal monitoring and early warning method as described in the first aspect of the present invention.

[0021] The beneficial effects of this invention are as follows: By constructing a mapping database that integrates in-situ multidimensional signals and non-in-situ ultrasonic features through step S2, and using a high-precision ultrasonic scanning system as a "physical ruler," low-cost in-situ sensing data such as thin-film pressure, infrared spectroscopy, and gas concentration are calibrated and their degradation levels are labeled at the microscopic mechanism level in an offline environment. This solves the problem that in-situ sensors in immersion environments are difficult to directly characterize early microscopic degradation such as SEI decomposition and trace lithium plating due to liquid disturbance and thermal coupling interference. It realizes the transformation of difficult-to-integrate ultrasonic detection capabilities into transferable implicit knowledge, providing a high-quality dataset with physical interpretability for model training. By training a multimodal temporal Transformer model through step S3 and introducing ultrasonic acoustic features as auxiliary supervision signals, the model is forced to learn the deep nonlinear mapping relationship between multidimensional online temporal signals and internal ultrasonic features during backpropagation. This breaks through the dimensional limitations of single physical field data in feature extraction, and significantly improves the model's identification accuracy of early weak fault features and its cross-condition generalization ability. By further combining the knowledge distillation and quantization compression processing of the S4 step, the complex model enhanced with ultrasonic knowledge is deployed in a lightweight manner on the edge computing gateway. This completely eliminates the dependence on ultrasonic hardware during the online inference stage. The very early warning with millisecond-level response can be achieved by using only in-situ sensors and BMS data through a sliding time window. This effectively balances monitoring accuracy and system economy, and ultimately achieves high-precision and high-reliability thermal runaway early warning with a low-cost sensing architecture in the immersion energy storage scenario. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0023] Figure 1 A flowchart for a multimodal monitoring and early warning method for thermal runaway in submerged energy storage battery packs.

[0024] Figure 2 This is a schematic diagram of a multimodal monitoring and early warning system for thermal runaway of an immersed energy storage battery pack. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Reference Figures 1-2 These are three embodiments of the present invention. This embodiment provides a multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack, comprising the following steps: Example 1 Example 1: Early thermal runaway monitoring and warning of a 52-cell (1P52S) type submerged lithium iron phosphate battery pack This embodiment implements multi-modal monitoring and early warning of thermal runaway for a 52 kWh submerged 314Ah lithium iron phosphate (LFP) cell-configured energy storage battery pack.

[0029] Step S1: In-situ multidimensional physical field data acquisition configuration Inside the battery pack, a thin-film pressure-temperature composite sensor is attached between the large surfaces of every three cells, totaling 16 measuring points. This is used to monitor the thickness-direction expansion / contraction strain and bulk temperature of the cells during charging and discharging in real time. At the outlet pipe of the immersion liquid circulation system, a miniature infrared spectroscopy probe with a working wavelength of 2.5~5μm is integrated to monitor absorption peak changes at 2.8μm (trace moisture / electrolyte decomposition products) and 4.3μm (dissolution of abnormal gases such as CO2). A multi-functional gas sensor (detecting H2, CO, and VOCs) and a gas phase temperature sensor are installed in the gas chamber at the top of the battery pack housing. Simultaneously, the system is connected to the BMS to read the cell voltage, total current, SOC, and estimated internal resistance values ​​in real time. All data is aggregated to the edge computing gateway via a CAN bus, with a sampling frequency of 1Hz.

[0030] Step S2: Construct the mapping database and data annotation In a laboratory environment, an equivalent testing platform for battery packs from the same batch as those used in the field was constructed. In addition to deploying the same in-situ sensors as described above, a high-precision ultrasonic scanning system (center frequency 5MHz) was introduced as a non-in-situ detection method. Overcharge, thermal abuse, and high / low temperature cycling shock experiments were performed to simulate the entire process of a battery from normal operation to thermal runaway. During the experiments, changes in ultrasonic velocity and attenuation coefficient were recorded at a frequency of 1kHz, and the lithium plating and gas production within the battery were identified using ultrasonic C-scan images. Combined with post-experiment disassembly verification, the battery degradation process was strictly divided into 5 levels: Level 0 (normal), Level 1 (SEI film decomposition / trace lithium plating), Level 2 (significant lithium plating / small amount of gas production), Level 3 (large amount of gas production / electrolyte decomposition), and Level 4 (thermal runaway triggered). The time-series data from the in-situ sensors collected in the experiments were aligned with the corresponding ultrasonic characteristics (velocity, attenuation value) and degradation level labels to construct a mapping database containing 5000 hours of data.

[0031] Step S3: Training the multimodal temporal Transformer model Based on the aforementioned mapping database, a multimodal time-series Transformer model is trained. The model input consists of a sliding time window of data from the past 300 seconds, including 36 channels of thin-film pressure mean / variance, infrared absorbance ratio, pack pressure, characteristic gas concentration, tab temperature, battery bulk temperature, and BMS electrical parameters. The model output is configured with a main task and an auxiliary task: the main task uses a weighted cross-entropy loss function to predict the probability distribution of degradation levels within the future window; the auxiliary task uses a mean squared error (MSE) loss function to regress and predict the corresponding ultrasonic velocity and attenuation values. Through this multi-task joint training mechanism, the model is forced to learn how to invert the internal microscopic physical state of the battery from low-cost online sensor signals.

[0032] Step S4: Edge Deployment and Online Early Warning The trained model undergoes knowledge distillation and INT8 quantization compression to reduce its size by more than 70%, and is deployed on an ARM architecture edge computing gateway. During online operation, the system maintains a 300-second sliding window, performing forward inference every minute (every 60 accumulated data points) and outputting the current degradation level probability vector P(grade). The risk score is calculated according to the risk scoring formula in claim 7: ; The weighting coefficient is set as follows: =0.1, =0.3, =0.6, =1.0.

[0033] Deterioration level Preset weighting coefficients The expression is:

[0034] in, To preset the baseline weights, This represents the absolute value of the temperature change. This represents the absolute value of the voltage change. This represents the absolute value of the pressure change between battery cells. This represents the absolute value of the pressure change within the package. The deviation from the state of charge. The characteristic value is the concentration of the characteristic gas. , , , , τ and τ are both pre-calibrated adaptive adjustment coefficients.

[0035] The system has set multiple warning thresholds: R < 0.2 indicates safety; 0.2 ≤ R < 0.4 indicates concern; 0.4 ≤ R < 0.7 indicates a warning; and R ≥ 0.7 indicates danger. To prevent false alarms, the system has a confirmation strategy for three consecutive inference cycles. When the risk score exceeds 0.4 three times consecutively, the system triggers a "warning" signal. In addition to uploading to the cloud platform, it immediately executes a first-level linkage: initiating emergency circulation of the immersion liquid to accelerate heat removal and attempting active balancing via the BMS to eliminate crosstalk effects; if the score exceeds 0.7, a "danger" signal is immediately triggered, cutting off the relay and activating the fire suppression system.

[0036] Example 2: Early thermal runaway monitoring and warning of a 26-cell (1P26S) type submerged lithium iron phosphate battery pack This embodiment implements multi-modal monitoring and early warning of thermal runaway for a 23.3 kWh submerged 280Ah lithium iron phosphate (LFP) cell-configured energy storage battery pack.

[0037] Step S1: In-situ multidimensional physical field data acquisition configuration Inside the battery pack, a thin-film pressure-temperature composite sensor is attached between the large surfaces of every three cells, totaling eight measuring points. This is used to monitor the thickness-direction expansion / contraction strain and bulk temperature of the cells during charging and discharging in real time. At the outlet pipe of the immersion liquid circulation system, a miniature infrared spectroscopy probe with a working wavelength of 2.5~5μm is integrated to monitor absorption peak changes at 2.8μm (trace moisture / electrolyte decomposition products) and 4.3μm (dissolution of abnormal gases such as CO2). A multi-functional gas sensor (detecting H2, CO, and VOCs) and a gas phase temperature sensor are installed in the gas chamber at the top of the battery pack housing. Simultaneously, the system is connected to the BMS to read the cell voltage, total current, SOC, and estimated internal resistance values ​​in real time. All data is aggregated to the edge computing gateway via a CAN bus, with a sampling frequency of 1Hz.

[0038] Step S2: Construct the mapping database and data annotation In a laboratory environment, an equivalent testing platform for battery packs from the same batch as those used in the field was constructed. In addition to deploying the same in-situ sensors as described above, a high-precision ultrasonic scanning system (center frequency 5MHz) was introduced as a non-in-situ detection method. Overcharge, thermal abuse, and high / low temperature cycling shock experiments were performed to simulate the entire process of a battery from normal operation to thermal runaway. During the experiments, changes in ultrasonic velocity and attenuation coefficient were recorded at a frequency of 1kHz, and the lithium plating and gas production within the battery were identified using ultrasonic C-scan images. Combined with post-experiment disassembly verification, the battery degradation process was strictly divided into 5 levels: Level 0 (normal), Level 1 (SEI film decomposition / trace lithium plating), Level 2 (significant lithium plating / small amount of gas production), Level 3 (large amount of gas production / electrolyte decomposition), and Level 4 (thermal runaway triggered). The time-series data from the in-situ sensors collected in the experiments were aligned with the corresponding ultrasonic characteristics (velocity, attenuation value) and degradation level labels to construct a mapping database containing 5000 hours of data.

[0039] Step S3: Training the multimodal temporal Transformer model Based on the aforementioned mapping database, a multimodal time-series Transformer model is trained. The model input consists of a sliding time window of data from the past 300 seconds, including 36 channels of thin-film pressure mean / variance, infrared absorbance ratio, pack pressure, characteristic gas concentration, tab temperature, battery bulk temperature, and BMS electrical parameters. The model output is configured with a main task and an auxiliary task: the main task uses a weighted cross-entropy loss function to predict the probability distribution of degradation levels within the future window; the auxiliary task uses a mean squared error (MSE) loss function to regress and predict the corresponding ultrasonic velocity and attenuation values. Through this multi-task joint training mechanism, the model is forced to learn how to invert the internal microscopic physical state of the battery from low-cost online sensor signals.

[0040] Step S4: Edge Deployment and Online Early Warning The trained model undergoes knowledge distillation and INT8 quantization compression to reduce its size by more than 70%, and is deployed on an ARM-based edge computing gateway. During online operation, the system maintains a 300-second sliding window, performing forward inference once per minute (every 60 accumulated data points) and outputting the current degradation level probability vector P(grade).

[0041] The system has set multiple warning thresholds: R < 0.2 indicates safety; 0.2 ≤ R < 0.4 indicates concern; 0.4 ≤ R < 0.7 indicates a warning; and R ≥ 0.7 indicates danger. To prevent false alarms, the system has a confirmation strategy for three consecutive inference cycles. When the risk score exceeds 0.4 three times consecutively, the system triggers a "warning" signal. In addition to uploading to the cloud platform, it immediately executes a first-level linkage: initiating emergency circulation of the immersion liquid to accelerate heat removal and attempting active balancing via the BMS to eliminate crosstalk effects; if the score exceeds 0.7, a "danger" signal is immediately triggered, cutting off the relay and activating the fire suppression system.

[0042] Example 3: Application of 52-cell (1P52S) type submerged lithium iron phosphate battery pack under electrical crosstalk conditions This embodiment describes the implementation of multi-modal monitoring and early warning for thermal runaway in an energy storage battery pack configured with 52 kWh submerged 314Ah lithium iron phosphate (LFP) cells.

[0043] Step S1: Multidimensional sensor network layout Inside the battery pack, a thin-film pressure-temperature composite sensor is attached between the large surfaces of every three cells, totaling 16 measuring points. This is used to monitor the thickness-direction expansion / contraction strain and bulk temperature of the cells during charging and discharging in real time. At the outlet pipe of the immersion liquid circulation system, a miniature infrared spectroscopy probe with a working wavelength of 2.5~5μm is integrated to monitor absorption peak changes at 2.8μm (trace moisture / electrolyte decomposition products) and 4.3μm (dissolution of abnormal gases such as CO2). A multi-functional gas sensor (detecting H2, CO, and VOCs) and a gas phase temperature sensor are installed in the gas chamber at the top of the battery pack housing. Simultaneously, the system is connected to the BMS to read the cell voltage, total current, SOC, and estimated internal resistance values ​​in real time. All data is aggregated to the edge computing gateway via a CAN bus, with a sampling frequency of 1Hz.

[0044] Step S2: Construction of the mapping database under complex working conditions In the offline equivalent testing platform, the focus is on simulating the "electrical crosstalk" condition (i.e., abnormal local heating caused by excessive resistance of connecting plates or uneven air-cooling in adjacent cells, which is then conducted to surrounding cells). A high-precision data acquisition system records the abnormal voltage drop (OCV), dynamic resistance changes, and acoustic characteristic changes within the cells affected by crosstalk under this condition. Through high-precision ultrasonic attenuation coefficient measurement and pressure feedback, early signs of SEI film decomposition and gas production (corresponding to Level 1 degradation) are identified within the LFP cells even before a significant temperature rise occurs. Multi-parameter data (temperature, pressure, hydrogen concentration, OCV) under this condition are expanded into the mapping database, with enhancement specifically for Level 1 and Level 2 degradation data, ensuring the model can identify this atypical but highly hazardous early hazard in air-cooled LFP battery packs.

[0045] Step S3: Model Optimization and Generalization Ability Improvement When training the multimodal time-series Transformer model, an adversarial training strategy is introduced to enable the model to learn to maintain stability in recognizing Level 1 and Level 2 degradation states under different environmental temperatures (such as -20°C low-temperature charging and 45°C high-temperature float charging). This allows the model to not only learn patterns from the data but also understand the physical mechanisms through ultrasonic features, thus accurately extracting the precursor features of thermal runaway in the battery cell even under complex electrical crosstalk interference.

[0046] Step S4: Real-time inference and hierarchical linkage control The compressed model is deployed on a container-level edge computing gateway. An overlapping sliding window strategy (inference every 30 seconds) is employed to further shorten response time. In the risk score calculation, the weighting coefficients for Level 1 and Level 2 degradation states are increased (w2 is increased to 0.4) for faster response. When the risk score exceeds 0.4 three times consecutively, the system triggers a "warning" signal and activates maximum power cooling in conjunction with the BMS; if the score exceeds 0.7, a "danger" signal is immediately triggered, and the system cuts off cluster-level contactors within milliseconds, opens container valves for directional pressure relief, and activates the fire suppression system, effectively preventing heat spread throughout the container due to a single battery cell failure.

[0047] This embodiment also provides a multi-modal monitoring and early warning system for thermal runaway of an immersed energy storage battery pack, including: The system comprises a sensor module, a mapping database module, a model training module, and an edge computing gateway module. The sensor module is used to synchronously deploy a thin-film pressure sensor, an infrared spectroscopy probe, and a top-mounted multi-functional gas sensor within the submerged energy storage battery pack, combined with a battery management system, serving as a low-cost, long-term operational in-situ sensing front-end to collect multi-dimensional physical field time-series data in real time. The mapping database module stores the mapping relationships between degradation physical mechanisms, ultrasonic acoustic characteristic parameters, and multi-dimensional online sensor data established through offline ultrasonic detection. The model training module trains a multimodal time-series Transformer model based on the mapping database, using degradation level as the primary task label and ultrasonic acoustic characteristics as auxiliary supervision signals. The edge computing gateway module deploys the multimodal time-series Transformer model after knowledge distillation and quantization compression, and performs online inference and hierarchical early warning.

[0048] This embodiment also provides a computer device applicable to the multimodal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the multimodal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as proposed in the above embodiment.

[0049] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0050] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the multi-modal monitoring and early warning method for thermal runaway of an immersion energy storage battery pack as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0051] In summary, this invention constructs a mapping database that integrates in-situ multidimensional signals and non-in-situ ultrasonic features through step S2. Utilizing a high-precision immersion ultrasonic scanning system as a "physical benchmark," it calibrates and labels the degradation levels at the microscopic mechanism level for low-cost in-situ sensing data such as thin-film pressure, infrared spectroscopy, and gas concentration in an offline environment. This solves the problem that in-situ sensors in immersion environments are difficult to directly characterize early microscopic degradation such as SEI decomposition and trace lithium plating due to liquid disturbance and thermal coupling interference. It transforms expensive ultrasonic detection capabilities into transferable implicit knowledge, providing a high-quality dataset with physical interpretability for model training. Step S3 trains a multimodal temporal Transformer model and introduces ultrasonic acoustic features as auxiliary supervision signals, forcing the model to learn the deep nonlinear mapping relationship between multidimensional online temporal signals and internal ultrasonic features during backpropagation. This overcomes the dimensional limitations of single physical field data in feature extraction, significantly improving the model's accuracy in identifying early weak fault features and its cross-condition generalization ability. By further combining the knowledge distillation and quantization compression processing of the S4 step, the complex model enhanced with ultrasonic knowledge is deployed in a lightweight manner on the edge computing gateway. This completely eliminates the dependence on ultrasonic hardware during the online inference stage. The very early warning with millisecond-level response can be achieved by using only in-situ sensors and BMS data through a sliding time window. This effectively balances monitoring accuracy and system economy, and ultimately achieves high-precision and high-reliability thermal runaway early warning with a low-cost sensing architecture in the immersion energy storage scenario.

[0052] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-modal monitoring and early warning method for thermal runaway of an immersion energy storage battery pack, characterized in that: Including S1, a thin-film pressure-temperature composite sensor, an infrared spectroscopy probe, and a top multi-in-one gas sensor are synchronously arranged between the large surfaces of the cells inside the submerged energy storage battery pack, and combined with the battery management system, to collect real-time time-series data of electrical parameters, temperature, characteristic gases, stress, and spectral multi-dimensional physical field. S2, based on the multidimensional physical field time series data, an offline ultrasonic detection is used to establish a mapping database of degradation physical mechanism, ultrasonic acoustic characteristic parameters and multidimensional online sensor data, and a degradation level label is added to the multidimensional physical field time series data of each time window, and ultrasonic velocity, attenuation value and characteristic waveform are attached as auxiliary reference features. S3. Based on the mapping database, train a multimodal temporal Transformer model with degradation level as the main task label and multidimensional online sensor and battery management system data as auxiliary supervision signals, so that the model can learn implicit representations related to the changes in multidimensional physical field temporal signals from ultrasonic acoustic feature data. S4 deploys the trained multimodal temporal Transformer model on an edge computing gateway after knowledge distillation and quantization compression. It performs online inference by collecting complete multidimensional physical field temporal data once per second and maintaining a sliding time window, outputting a degradation level probability vector and obtaining a risk score. Based on the set multi-level warning thresholds, it triggers corresponding graded warning signals.

2. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 1, characterized in that: In step S1, the thin-film pressure-temperature composite sensor is attached between the large surfaces of adjacent battery cells, with one thin-film pressure sensor arranged every three battery cells to sense the expansion / contraction strain and bulk temperature in the thickness direction of the battery cells; the infrared spectroscopy probe is arranged in the outlet area of ​​the battery pack immersion liquid circulation system to characterize the physicochemical deterioration of the immersion liquid through the change of absorption peaks at specific wavenumbers; and the top multi-functional gas sensor is used to monitor the concentration of characteristic gases and the gas phase temperature.

3. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 2, characterized in that: The infrared spectral probe operates in the 2.5~5μm band. As an economically viable in-situ sensing method, it reflects trace amounts of moisture or electrolyte decomposition products by monitoring the absorption peak near 2.8μm, and reflects abnormal gas production and dissolution by monitoring the absorption peak near 4.3μm.

4. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 3, characterized in that: In step S2, the step of establishing a mapping database using offline ultrasonic testing includes: using a high-precision ultrasonic scanning system as a non-in-situ detection method, building an offline equivalent test platform of the same batch as the online battery pack, and simultaneously deploying a thin-film pressure-temperature composite sensor, an infrared spectral probe, and a top multi-in-one gas sensor in a sealed laboratory container to simulate the degradation process of the submerged energy storage battery pack under typical operating conditions and typical abnormal conditions such as overcharging, thermal abuse, and electrical crosstalk, and introducing a high-precision ultrasonic scanning system; performing typical abnormal condition degradation experiments, synchronously recording all online sensor data at a first sampling frequency, and simultaneously recording ultrasonic velocity, attenuation coefficient, and backscatter signal through ultrasonic equipment at a second sampling frequency; combining ultrasonic scanning images, acoustic parameter mutation points, and post-disassembly analysis, dividing the battery degradation process into multiple degradation levels, and labeling the multidimensional physical field time series data of each time window with the corresponding degradation level, while attaching ultrasonic velocity and attenuation value as auxiliary reference features, and constructing a dataset of submerged battery thermal runaway time series characteristic signals that integrates in-situ multidimensional signals and non-in-situ ultrasonic features.

5. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 4, characterized in that: The degradation levels include at least Level 0 (normal), Level 1 (solid electrolyte membrane (SEI) decomposition / trace lithium plating), Level 2 (significant lithium plating / small amount of gas production), Level 3 (large amount of gas production / electrolyte decomposition), and Level 4 (thermal runaway triggering). Accurate identification of Level 1 and Level 2 degradation states is used to achieve very early warning of thermal runaway.

6. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 1, characterized in that: In step S3, the input to the multimodal time-series Transformer model is multidimensional physical field time-series data within a preset time window, including at least one of the following: film pressure (mean / variance), infrared absorbance ratio, pack pressure, characteristic gas concentration, tab temperature, gas phase temperature (collected by a multi-sensor), battery bulk temperature (collected by a composite film pressure-temperature sensor between the cell surfaces), battery terminal voltage, current, battery state of charge (SOC), and internal resistance estimation. The model output includes a main task and an auxiliary task. The main task is used to predict the probability distribution of degradation level within a future time window, and the auxiliary task is used to regress and predict the corresponding ultrasonic acoustic features. Through joint iterative training with massive in-situ sensor data and corresponding non-in-situ ultrasonic annotation data, the model learns the implicit mapping relationship from multidimensional online time-series signals to the internal degradation state of the battery, improving the model's robustness and generalization ability. During training, the loss function is the sum of the weighted cross-entropy loss of the main task and the mean square error loss of the auxiliary task.

7. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 1, characterized in that: In step S4, the risk score is defined as the weighted sum of the probabilities of the degradation level reaching a preset level. Multi-level warning thresholds are set based on the risk score. When the risk score exceeds the corresponding threshold, a warning signal of the corresponding level is triggered and uploaded to the monitoring system and cloud platform, triggering a security linkage strategy. The risk score expression is: ； in, To score the risk, This is the highest deterioration level number. For discrete variables representing the degree of degradation, The lower threshold of the preset level, To correspond to the degradation level The preset weighting coefficients, Currently at a degraded level The probability, the level of degradation Preset weighting coefficients The expression is: ;in, To preset the baseline weights, This represents the absolute value of the temperature change. This represents the absolute value of the voltage change. This represents the absolute value of the pressure change between battery cells. This represents the absolute value of the pressure change within the package. The deviation from the state of charge. The characteristic value is the concentration of the characteristic gas. , , , , τ and τ are both pre-calibrated adaptive adjustment coefficients.

8. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 5, characterized in that: The multi-level early warning thresholds include: risk score <0.2 for safe, 0.2-0.4 for attention, 0.4-0.7 for warning, and >0.7 for danger; before the early warning is triggered, a false alarm prevention strategy is also included, which confirms the warning for multiple consecutive inference cycles to improve the reliability of the system's early warning under complex operating conditions.

9. The multi-modal monitoring and early warning method for thermal runaway of an immersed energy storage battery pack as described in claim 6, characterized in that: In step S4, the trained multimodal temporal Transformer model undergoes knowledge distillation and INT8 quantization compression, and is deployed on an edge computing gateway next to the battery pack. This eliminates the need for ultrasonic detection equipment during online operation. During online operation, complete multidimensional physical field temporal data is collected once per second, a 300-second sliding time window is maintained, and forward inference is performed every 60 data points (1 minute) to output a degradation level probability vector. This effectively reduces the platform's computational load and inference pressure while ensuring monitoring timeliness, enabling early warning of thermal runaway based solely on in-situ sensors and battery management system (BMS) data.

10. A multi-modal monitoring and early warning system for thermal runaway of an immersion energy storage battery pack, based on the multi-modal monitoring and early warning method for thermal runaway of an immersion energy storage battery pack according to any one of claims 1 to 9, characterized in that: It includes a sensor module, a mapping database module, a model training module, and an edge computing gateway module; the sensor module is used to synchronously arrange thin-film pressure sensors, infrared spectroscopy probes, and top multi-in-one gas sensors between the large surfaces of the cells inside the submerged energy storage battery pack and combine them with the battery management system, serving as a low-cost, long-term operating in-situ sensing front end to collect multi-dimensional physical field time-series data in real time. The mapping database module is used to store the mapping relationship between the degradation physical mechanism, ultrasonic acoustic feature parameters and multidimensional online sensor data established by offline ultrasonic detection; the model training module is used to train a multimodal temporal Transformer model with degradation level as the main task label and ultrasonic acoustic features as the auxiliary supervision signal according to the mapping database; the edge computing gateway module is used to deploy the multimodal temporal Transformer model after knowledge distillation and quantization compression, and perform online inference and hierarchical early warning.

Images