Thermal runaway early warning method and device for sodium battery energy storage system

Abstract

The invention relates to the technical field of sodium battery energy storage safety, and discloses a thermal runaway early warning method and device for a sodium battery energy storage system. Multi-modal data such as acoustic emission, gas dielectric characteristics and electrical parameters are collected in real time through multiple sensors, pre-processing and feature extraction are performed through an edge calculation unit, and the data are transmitted to a central decision-making unit for multi-source information fusion and risk assessment; the thermal runaway risk probability is dynamically calculated by adopting a deep learning algorithm based on an attention mechanism, a grading early warning instruction is generated according to the risk grade, and accurate response is realized through multi-system linkage execution. According to the invention, early-stage monitoring of internal microscopic damage and gas environment abnormity of the sodium battery is realized, timeliness and accuracy of thermal runaway early warning are improved, and safe and stable operation of a sodium battery energy storage system is guaranteed.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of sodium battery energy storage safety, in particular to a sodium battery energy storage system thermal runaway early warning method and device. BACKGROUND

[0002] The sodium battery energy storage system is a complete energy solution that uses sodium ion batteries as the core electric energy storage carrier, and is equipped with energy storage converters, battery management systems, monitoring systems and other devices to realize the functions of electric energy storage, scheduling and release. Compared with mainstream lithium battery energy storage, it has the advantages of lower cost, higher safety, stronger environmental adaptability and lower resource dependence. The battery cabin is an important structure in the sodium battery energy storage system, which is a closed or semi-closed box for containing and protecting the sodium ion battery pack and supporting equipment in the system. As the core physical carrier of the system, it can isolate outdoor high temperature, low temperature, rain, dust and other harsh environments, protect the battery safety through the built-in fire extinguishing system, temperature and smoke sensor, and can also integrate the battery pack, BMS local module and other devices to reduce field wiring, facilitate overall transportation and rapid deployment. Its typical structure includes cabin shell, battery cluster area, control and monitoring area, heat dissipation and fire extinguishing area and auxiliary facility area, which plays a key role in the stable operation of sodium batteries in industrial scenarios.

[0003] The most important problem faced by the sodium battery energy storage system during operation is the risk of thermal runaway. Sodium batteries may have irreversible exothermic reactions due to internal abnormalities, causing the battery temperature to rise sharply, and then causing serious consequences such as electrolyte decomposition, gas generation, fire and explosion. Once a thermal runaway accident occurs, it will not only cause equipment loss and economic loss, but also endanger personnel safety and surrounding environmental safety.

[0004] Therefore, it is particularly important to perform thermal runaway early warning for the sodium battery energy storage system. The existing early warning methods mainly rely on traditional electrical parameter monitoring, including monitoring of macroscopic parameters such as voltage, current, temperature, etc., which has the problem of response lag and can only detect when the battery has obvious abnormal signs, often missing the best early warning and intervention opportunity. When these parameters change abnormally, the microscopic damage process inside the battery may have lasted for a considerable period of time, and the thermal runaway process may have entered an irreversible stage. In addition, the traditional early warning method is mainly based on simple threshold judgment, using fixed early warning thresholds and binary normal-abnormal judgment logic. This simplified judgment mechanism is prone to false positives and false negatives. False positives will lead to unnecessary system downtime and economic losses, while false negatives may lead to serious safety accidents. At the same time, the traditional method is difficult to accurately quantify the risk level and cannot provide fine-grained risk assessment and graded response.

[0005] In view of the problems and deficiencies in the prior art, there is an urgent need to develop a new sodium battery energy storage system thermal runaway early warning method and system to comprehensively improve the safe operation level of the sodium battery energy storage system and provide reliable safety protection for large-scale industrial application of sodium battery energy storage technology. SUMMARY

[0006] To solve the problems in the background art, the present application provides a sodium battery energy storage system thermal runaway early warning method, comprising the following steps: S1, real-time data acquisition by multiple sensors; synchronous acquisition of multi-modal monitoring data by high-frequency acoustic emission sensors, microwave resonant cavity sensors and electrical sensors deployed in the sodium battery energy storage system battery cabin to monitor the internal state changes of the sodium battery; S2, multi-source data preprocessing and feature extraction; the original sensor data collected in step S1 are input into an edge computing unit for preprocessing, including extracting feature indicators reflecting internal microscopic damage of the battery from acoustic emission signals through a signal processing algorithm and extracting feature indicators reflecting abnormality of the gas environment in the battery cabin from dielectric property changes; S3, multi-source information fusion and risk assessment; the feature indicators extracted in step S2 are transmitted to a central decision unit together with voltage, current and temperature electrical parameters, and the central decision unit calculates the thermal runaway risk of the current sodium battery energy storage system; S4, early warning decision and execution; when the thermal runaway risk probability calculated in step S3 exceeds a preset safety threshold, the central decision unit generates a graded early warning instruction and transmits it to the battery management system and / or the fire linkage system through a communication interface.

[0007] Further, the step S1 specifically comprises: S11, acoustic emission signal acquisition; the high-frequency acoustic emission sensor is in close coupled contact with the sodium battery cell shell to real-time collect high-frequency acoustic emission signals generated by internal electrochemical reactions, mechanical stress changes and microscopic crack propagation during battery operation; S12, gas dielectric property acquisition; the microwave resonant cavity sensor emits microwave signals of a specific frequency to the space above the battery module and receives reflected signals to measure the resonance frequency change of the environment, and the relative change amount of the dielectric constant is calculated according to the formula: ; The relative change amount of the dielectric constant is dimensionless. The calibration coefficient related to the resonant cavity geometry and material properties is dimensionless. The calibration coefficient related to the resonant cavity geometry and material properties is dimensionless. The difference between the current measured resonance frequency and the reference resonance frequency is dimensionless. This is the reference resonant frequency established under standard healthy conditions; S13. Electrical parameter acquisition: Voltage and current sensors are connected to the positive and negative terminals of the sodium battery module, respectively, and temperature sensors are connected to several temperature monitoring points of the sodium battery module; real-time acquisition of terminal voltage, charging and discharging current and monitoring point temperature data reflecting the electrical and thermal states of the battery.

[0008] Furthermore, step S2 specifically includes: S21. Acoustic emission signal feature extraction: The edge computing unit receives the raw acoustic emission signal collected by the high-frequency acoustic emission sensor, decomposes the time-domain signal into different frequency bands through the wavelet packet transform algorithm, and identifies the characteristic frequency bands related to the abnormal state inside the battery. S22, Acoustic emission energy entropy calculation; edge computing unit targets the characteristic frequency band identified in step S21. The wavelet packet energy entropy is calculated as an acoustic emission characteristic index using the formula: ; Perform calculations; where is a dimensionless characteristic index of acoustic emission energy entropy. For the first The normalized energy ratio of each sub-band, through Calculated; For frequency band Inner Wavelet packet energy of individual frequency bands; To analyze the total energy within the frequency band; S23, Transmission of gas characteristic quantities in the battery compartment; The edge computing unit transmits the relative change in the dielectric constant of the gas in the battery compartment calculated in step S12. As a characteristic indicator reflecting abnormalities in the cabin gas environment, it is related to acoustic emission characteristic indicators. They are transmitted together to the central decision-making unit.

[0009] Furthermore, step S3 specifically includes: S31. Asynchronous time series input processing; The central decision-making unit receives asynchronous time series data from different sensor channels and constructs a comprehensive input vector: ; in for The combined input vector at time step; for The characteristic index of acoustic emission energy entropy at any given moment; for The relative change in the dielectric constant of the gas inside the battery compartment at any given time; for The measured value of the battery terminal voltage at a given time, in volts (V); for The measured values ​​of charging and discharging current at each moment; for Temperature measurements at key points in time; S32. Attention Weight Calculation: The early warning model analyzes the importance of each input feature to the current state judgment through an attention mechanism and calculates a dynamic weight vector. ; in for Attention weight vectors for each input feature at each time step; It is a normalized exponential function; This is the attention scoring function, used to calculate the relevance score between the hidden state and the input features; For the model in The hidden state vector at time step; S33. Calculation of thermal runaway risk probability; the early warning model is based on weighted input features and historical state information, using the formula: ; Calculate the probability of thermal runaway at the current moment; where for The probability value of thermal runaway risk at any given moment, ranging from 0 to 1; It is the sigmoid activation function; The weight matrix obtained during training; This represents the element-wise multiplication operator; This is the bias vector.

[0010] Furthermore, step S4 specifically includes: S41. Multi-level risk threshold discrimination; The central decision-making unit will use the thermal runaway risk probability calculated in step S33. The system compares the data with the preset four-level security thresholds level by level. When it is determined to be a low-risk level, When it was determined to be a medium-risk level, When it is determined to be a high-risk level, It was determined to be at an extremely high risk level. S42. Generation of graded early warning instructions: The central decision-making unit generates corresponding early warning instructions based on the risk level determined in step S41. When the risk level is low, it generates status monitoring enhancement instructions and maintenance personnel reminder notices. When the risk level is medium, it generates charging and discharging power limitation instructions and on-site inspection notices for technical personnel. When the risk level is high, it generates emergency circuit breaker instructions for the battery management system and early warning preparation instructions for the fire protection system. When the risk level is extremely high, it generates forced circuit breaker instructions for the battery management system, immediate activation instructions for the fire protection system, and emergency evacuation instructions for personnel. S43. Multi-system coordinated execution: The central decision-making unit transmits battery control commands to the battery management system through the CAN bus communication interface, fire control commands to the fire linkage system through the Ethernet communication interface, and personnel notification commands to the operation and maintenance management system through the wireless communication interface, so as to realize multi-system coordinated response and graded handling of thermal runaway risk.

[0011] The present invention also designs a thermal runaway early warning device for a sodium battery energy storage system, including a sensing module, an edge processing module, a central decision-making module and an execution module; The sensing module includes a high-frequency acoustic emission sensor, a microwave resonant cavity sensor, a voltage sensor, a current sensor, and a temperature sensor, used for real-time data acquisition. The high-frequency acoustic emission sensor is fixedly installed on the outer surface of the sodium battery cell casing and forms an acoustic coupling connection with the cell casing through a high-temperature coupling agent. The microwave resonant cavity sensor is suspended in the space above the battery module inside the battery compartment, with its resonant cavity opening facing the battery module to form a gas detection chamber. The voltage sensor, current sensor, and temperature sensor are respectively connected to the electrical terminals and temperature monitoring point of the sodium battery module. The edge processing module includes multiple edge computing units for multi-source data preprocessing and feature extraction. Each edge computing unit establishes a communication connection with the high-frequency acoustic emission sensor and microwave resonant cavity sensor of the corresponding battery module through a data acquisition interface. The edge computing unit has a built-in digital signal processor and memory for executing wavelet packet transform algorithm and energy entropy calculation algorithm. The central decision-making module includes a central decision-making unit, which is used to realize multi-source information fusion and risk assessment. The central decision-making unit establishes a communication connection with all edge computing units through a data communication bus, and at the same time establishes a direct connection with voltage sensors, current sensors and temperature sensors through sensor interfaces. The central decision-making unit has a built-in processor and a pre-trained thermal runaway early warning model. The execution module includes a battery management system and a fire alarm linkage system, which are used to realize early warning decision-making and execution; the central decision-making unit establishes communication connections with the battery management system and the fire alarm linkage system respectively through a control communication interface.

[0012] In a preferred embodiment, the high-frequency acoustic emission sensor employs a piezoelectric ceramic transducer structure; the high-frequency acoustic emission sensor is connected to the sodium battery cell housing via a high-temperature resistant acoustic coupling agent.

[0013] In a preferred embodiment, the microwave resonant cavity sensor adopts a cylindrical resonant cavity structure; the volume of the microwave resonant cavity sensor matches the volume of the gas space above the monitored battery module, and the inner wall of the resonant cavity is coated with a conductive material to form an electromagnetic shielding layer.

[0014] In a preferred embodiment, the edge computing unit has a built-in ARM architecture microprocessor and a dedicated digital signal processor; the memory of the edge computing unit is pre-installed with a wavelet packet transform function library and a fast Fourier transform algorithm library; and the edge computing unit establishes a data communication connection with the central decision-making unit through a CAN bus interface.

[0015] In a preferred embodiment, the central decision-making unit adopts an industrial-grade embedded computing platform with a built-in multi-core processor and a large-capacity memory; the battery management system is connected to the central decision-making unit via an RS485 communication interface to receive circuit breaker control commands; the fire alarm linkage system is connected to the central decision-making unit via an Ethernet interface to receive pre-cooling start commands; the central decision-making unit is also equipped with a human-machine interface to display system operating status and early warning information.

[0016] The beneficial effects achieved by this invention are as follows:

[0017] First, the thermal runaway early warning method for sodium battery energy storage systems designed in this invention adopts multi-sensor fusion technology. It constructs a comprehensive battery status monitoring network by using acoustic emission sensors to monitor the microscopic damage process inside the battery, microwave resonant cavity sensors to detect changes in the gas environment, and traditional electrical sensors to collect voltage, current, and temperature parameters. By using multi-modal data fusion, it overcomes the shortcomings of traditional single sensors, such as limited information and delayed response. It enables the detection of internal microscopic changes before macroscopic anomalies occur in the battery, achieving advanced early warning capability and significantly improving the timeliness and accuracy of the early warning.

[0018] Second, this invention employs a deep learning algorithm based on an attention mechanism for risk assessment. By adaptively analyzing the importance of each input feature at different times, it achieves dynamic weight allocation and decision-making. It can process asynchronous time-series data from different sensors, automatically identify key features, and calculate continuous thermal runaway risk probability values, providing a more accurate quantitative risk assessment compared to traditional qualitative judgment. The introduction of the attention mechanism enables the adjustment of the weights of each sensor data according to the current state, improving the intelligence level and accuracy of the early warning.

[0019] Third, this invention designs a four-level risk classification and graded early warning response mechanism, formulating corresponding response strategies based on different risk probability ranges. This avoids the false alarms or over-handling problems that may arise from the simple binary judgment of normal and abnormal situations in traditional early warning systems. Monitoring is strengthened during low-risk periods, power is limited during medium-risk periods, emergency circuit breaker is activated during high-risk periods, and fire protection systems and personnel evacuation are initiated during extremely high-risk periods. Through this refined risk management strategy, economic benefits are maximized while ensuring safety, achieving a precise match between early warning measures and risk levels.

[0020] Fourth, the sodium battery energy storage system thermal runaway early warning device designed in this invention adopts a hierarchical distributed architecture design. Each edge computing unit has independent signal processing capabilities and can execute complex algorithms such as wavelet packet transform and energy entropy calculation. It also has a certain degree of autonomous decision-making capability and can still operate independently when communication is interrupted. The distributed processing architecture not only improves the reliability and fault tolerance of the system, but also lays the foundation for the modular expansion of the system.

[0021] Fifth, the thermal runaway early warning device for the sodium battery energy storage system designed in this invention adopts a modular integrated design. Each functional module is relatively independent and connected through a standardized industrial communication protocol, facilitating system upgrades, maintenance, and functional expansion. The multi-system linkage execution mechanism enables rapid transmission and coordinated response of early warning information among the battery management system, fire alarm system, and operation and maintenance management system. It establishes a multi-level safety assurance system at the sensor level, edge computing level, central decision-making level, and execution level, ensuring the system's fault-safe characteristics. Even if a fault occurs at one level, other levels can still provide basic safety protection, significantly improving the overall system reliability and safety. Attached Figure Description

[0022] Fig. 1 This is a flowchart of the thermal runaway early warning method for sodium battery energy storage system of the present invention; Fig. 2 This is a schematic diagram of the structure of the thermal runaway early warning device for the battery energy storage system of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figs. 1-2The present invention provides a method for early warning of thermal runaway in sodium battery energy storage systems. This method employs multi-sensor fusion technology, monitors the internal state changes of the sodium battery in real time, and utilizes advanced artificial intelligence algorithms for risk assessment. This enables early warning and tiered response to thermal runaway accidents, addressing the technical problems of delayed response and insufficient accuracy in thermal runaway early warning in existing sodium battery energy storage systems. The method significantly improves the accuracy and timeliness of early warning through multi-modal data fusion. The specific steps are as follows: Step S1: Real-time data acquisition from multiple sensors. Step S1 constructs a comprehensive battery status monitoring network through the collaborative work of various types of sensors. Traditional battery monitoring systems mainly rely on electrical parameters such as voltage, current, and temperature. These parameters often have a lag and cannot reflect microscopic changes inside the battery in a timely manner. This invention introduces acoustic emission sensors and microwave resonant cavity sensors, which can detect internal microscopic damage and changes in the gas environment before macroscopic anomalies occur in the battery, achieving early warning. The specific implementation steps of S1 are as follows: S11, High-frequency acoustic emission signal acquisition; the high-frequency acoustic emission sensor adopts a piezoelectric ceramic transducer structure, which has good high-frequency response characteristics and temperature stability. Acoustic emission (AE) refers to the elastic wave phenomenon generated by the adjustment of atomic or molecular structure when a material is subjected to external force or internal stress changes. During the operation of sodium batteries, characteristic acoustic emission signals are generated by electrochemical reactions, ion migration, volume changes of electrode materials, formation and destruction of solid electrolyte interfacial films, and the initiation and propagation of internal microcracks.

[0025] The sensor forms a tight acoustic coupling connection with the sodium battery cell casing via a high-temperature resistant acoustic coupling agent. The role of the acoustic coupling agent is to eliminate air gaps between the sensor and the cell casing, ensuring efficient sound wave transmission and preventing signal attenuation and distortion. Its high-temperature resistance ensures stable coupling within the battery's operating temperature range.

[0026] The acoustic emission signals acquired by the sensor in real time typically range from 20 kHz to 1 MHz, with a sampling frequency of no less than 2 MHz to satisfy the Nyquist sampling theorem. These high-frequency signals can reflect transient processes inside the battery, including lattice changes during ion insertion / extraction, the rupture of bubbles generated by electrolyte decomposition, and mechanical friction between active material particles.

[0027] S12, Gas dielectric property acquisition; The microwave resonant cavity sensor adopts a cylindrical resonant cavity structure, and its working principle is based on the resonance phenomenon of electromagnetic waves in a closed cavity. When the battery malfunctions, gases such as hydrogen, carbon monoxide, carbon dioxide, and hydrocarbon compounds may be generated. The dielectric constant of these gases differs significantly from that of air, which will cause changes in the resonant frequency of the resonant cavity.

[0028] The sensor transmits a microwave signal of a specific frequency into the space above the battery module, typically choosing the ISM band such as 2.4 GHz or 5.8 GHz to avoid interference with other communication devices. The transmission power is controlled at the milliwatt level to ensure electromagnetic compatibility. After the reflected signal is received by the receiving antenna, the system analyzes its spectral characteristics and extracts the resonant frequency information.

[0029] The formula for calculating the relative change in dielectric constant is: ; in, It is the relative change in dielectric constant, a dimensionless parameter that reflects the degree of change in the current gas environment relative to the reference state. The calibration coefficients are determined by the geometry of the resonant cavity, including its length, diameter, shape, and the conductivity and permeability characteristics of the inner wall material. The calibration coefficients are established through calibration experiments in a known gas environment and are calibrated once during the system installation and commissioning phase. This is the difference between the currently measured resonant frequency and the reference resonant frequency, expressed in Hertz (Hz). The reference resonant frequency is a reference value established under standard healthy conditions, typically in a clean air environment with a temperature of 25°C and a relative humidity of 50%.

[0030] S13: Electrical Parameter Acquisition; The electrical parameter acquisition system includes three subsystems: a voltage sensor, a current sensor, and a temperature sensor. The voltage sensor uses a differential measurement method and is directly connected to the positive and negative terminals of the sodium battery module. The measurement range is typically 0-4V, with an accuracy of no less than 0.1%. The current sensor uses the Hall effect principle or the shunt principle to measure the charging and discharging current, and its range is determined according to the battery capacity.

[0031] Temperature sensors employ RTD or NTC technologies and are deployed at key temperature monitoring points in the battery module, including cell surfaces, busbar connections, and cooling pipe inlets and outlets, to promptly detect localized overheating. The number of monitoring points is determined based on the size of the battery module, with each battery pack configured with no fewer than four temperature monitoring points.

[0032] The electrical parameter acquisition system collects data at a frequency of 1Hz to 10Hz. The selection of the acquisition frequency takes into account the balance between the rate of change of battery electrical characteristics and data processing capabilities. Too high an acquisition frequency will generate a large amount of redundant data and increase the system load; too low an acquisition frequency may miss rapidly changing abnormal signals.

[0033] Step S2 involves multi-source data preprocessing and feature extraction. Step S2 transforms the raw sensor data into feature indicators with clear physical meaning, providing high-quality input for subsequent fusion analysis. Since data from different types of sensors have different physical dimensions, numerical ranges, and temporal characteristics, direct fusion analysis can lead to feature weight imbalance and model performance degradation. Distributed preprocessing using edge computing units not only reduces data transmission burden but also enables signal processing tasks with high real-time requirements. The specific implementation steps of S2 are as follows: S21, Acoustic emission signal feature extraction; After receiving the raw acoustic emission signal collected by the high-frequency acoustic emission sensor, the edge computing unit first performs signal preprocessing, including noise filtering, baseline correction, and amplitude normalization. Noise filtering uses a bandpass filter to remove environmental noise such as power frequency interference and mechanical vibration.

[0034] Wavelet packet transform is an important tool for signal time-frequency analysis. Compared to the traditional Fourier transform, wavelet packet transform has good time-frequency locality and can simultaneously provide both time and frequency information of the signal. The algorithm decomposes the time-domain acoustic emission signal into different frequency bands, each corresponding to a specific physical process or anomalous pattern.

[0035] The mathematical expression for wavelet packet transform is: ; ; in, Indicates the first Layer Wavelet packet coefficients of each node, and These are the coefficients of the low-pass and high-pass filters, respectively. Through multi-level decomposition, the signal is broken down into different frequency bands, each reflecting the energy distribution within a specific frequency range.

[0036] The identification of characteristic frequency bands is based on statistical analysis of historical data and expert experience. Practical experience has shown that the acoustic emission signals generated by sodium batteries during normal operation are mainly concentrated in the 50-150kHz frequency band, while abnormal states such as internal short circuits produce significant signal enhancement in the 200-500kHz frequency band. By establishing a correspondence between frequency bands and abnormal states through training data, the system automatically identifies characteristic frequency bands related to abnormal states within the battery.

[0037] S22, Calculation of Acoustic Emission Energy Entropy; Energy entropy is an important indicator in information theory describing signal complexity and uncertainty, and is often used in signal processing to characterize the frequency domain distribution characteristics of a signal. For acoustic emission signals, energy entropy can quantify the degree of energy dispersion of the signal in different frequency bands, thus reflecting the complexity and anomaly degree of the internal processes of the battery. The calculation process of wavelet packet energy entropy is as follows: First, calculate the wavelet packet energy of each sub-band: ; in, For the first The wavelet packet energy of each frequency band is expressed in joules (J). These are the corresponding wavelet packet coefficients.

[0038] Then, calculate the total energy: ; Next, the normalized energy ratio of each sub-band is calculated: ; For the first The normalized energy ratio of each sub-band is a dimensionless parameter that satisfies The constraints.

[0039] Finally, calculate the acoustic emission energy entropy: ; This is a dimensionless indicator of acoustic emission energy entropy. Its physical meaning is: when the battery is in normal condition, the energy of the acoustic emission signal is mainly concentrated in a few frequency bands, and the energy entropy value is relatively small; when the battery malfunctions, the energy in multiple frequency bands increases, the energy distribution becomes more dispersed, and the energy entropy value increases.

[0040] S23, Transmission of gas characteristic quantities in the battery compartment; the edge computing unit calculates the relative change in the dielectric constant of the gas in the battery compartment obtained in step S12. As a characteristic indicator reflecting abnormalities in the cabin's gas environment, the physical meaning of this indicator is clear: This indicates that the dielectric constant of the current gas environment is higher than that of the reference state, and polar gases such as water vapor and organic solvent vapor may be present. This indicates that the dielectric constant is lower than the reference state, and non-polar gases such as hydrogen and alkanes may be present.

[0041] Edge computing units transmit acoustic emission characteristics via industrial communication protocols such as the CAN bus. and gas characteristic indicators The data is transmitted together to the central decision-making unit. The data transmission uses a timestamp synchronization mechanism to ensure that data from different sensors can be correctly mapped to the same time point.

[0042] Step S3, Multi-source Information Fusion and Risk Assessment; Step S3 achieves the fusion of multi-source heterogeneous data, overcoming the shortcomings of limited information from a single sensor and the susceptibility to false alarms. An asynchronous recurrent neural network architecture based on an attention mechanism adaptively analyzes the importance of each input feature and considers historical state information to achieve an accurate quantitative assessment of thermal runaway risk. The specific implementation steps of S3 are as follows: S31, Asynchronous Time Series Input Processing: The central decision unit receives asynchronous time series data from different sensor channels. Due to differences in sampling frequency, transmission delay, and data processing time among the sensors, the data arriving at the central decision unit is often asynchronous. The system converts the asynchronous data into a synchronous time series using timestamp calibration and data interpolation techniques.

[0043] The formula for constructing the comprehensive input vector is: ; in, for The comprehensive input vector at time step is a 5-dimensional vector containing complete state information of the system at the current time step. The physical meaning and numerical characteristics of each component are as follows: : The acoustic emission energy entropy characteristic index at any given time is dimensionless, with a normal range of 0.1-0.8, and can reach above 1.0 in abnormal cases; : The relative change in the dielectric constant of the gas inside the battery compartment at any given time is dimensionless, with a normal range of ±0.01 and an abnormal range of over ±0.1. : The measured value of the battery terminal voltage at any given time, in volts (V), which is typically in the range of 2.0-3.8V for sodium batteries; : The charging and discharging current measurement value at any given time, in amperes (A), with positive values ​​indicating charging and negative values ​​indicating discharging; : The key temperature measurement values ​​at any given time are in degrees Celsius (°C). The normal operating temperature is usually in the range of 15-45°C. S32, Attention Weight Calculation; Attention mechanism is an important technique in deep learning. Its core idea is to enable the model to automatically identify important parts of the input data and assign them different weights. In thermal runaway early warning scenarios, the importance of each sensor data point changes dynamically at different times, and the attention mechanism can adaptively adjust the weights of each feature.

[0044] The mathematical expression for calculating attention weights is: ; in, for The attention weight vector for each input feature at each time step is a 5-dimensional vector, with the sum of its components being 1. The larger the weight value, the more important the corresponding feature is in the current state judgment.

[0045] The function is the normalized exponential function, defined as: ; This function ensures that all weight values ​​are positive and their sum is 1, which meets the requirements of the probability distribution.

[0046] For attention scoring functions, dot product attention or additive attention mechanisms are typically used: ; in, The attention weight matrix is ​​trainable and optimized using the backpropagation algorithm.

[0047] For the model in The hidden state vector at each moment contains state information from historical moments, reflecting the system's temporal memory capability.

[0048] S33, Calculation of thermal runaway risk probability; The calculation of thermal runaway risk probability is the final output of the entire early warning algorithm. The sigmoid activation function is used to ensure that the output value is between 0 and 1, which conforms to the definition of probability.

[0049] The calculation formula is: ; in, for The probability value of thermal runaway risk at any given moment ranges from 0 to 1. A value close to 0 indicates an extremely low risk, while a value close to 1 indicates that thermal runaway is imminent.

[0050] The sigmoid activation function is defined as follows: ; This function possesses favorable mathematical properties: its output range is (0,1), and it is continuously differentiable. The area is most sensitive to changes.

[0051] The weight matrix was obtained through machine learning training on a large amount of historical data. The training dataset includes normal operation data, data on various abnormal states, and data from actual thermal runaway accidents.

[0052] represents the element-wise multiplication operator (Hadamard product), i.e.: ; This operation realizes the dynamic modulation of the input features by the attention weights, where the values of important features are amplified and the influence of secondary features is suppressed. is the bias vector, which is used to adjust the judgment reference point of the model and is automatically optimized through the training process.

[0053] Step S4, early warning decision and execution; Step S4 realizes the refined classification of risk levels and corresponding differential responses. Through multi-level risk threshold discrimination, it avoids the false alarm or missed alarm problems that may be brought by simple binary normal / abnormal judgments. The classified early warning instructions ensure that the response measures match the risk level, which not only ensures safety but also avoids unnecessary economic losses. The multi-system linkage execution realizes the rapid transmission and collaborative response of early warning information. The specific implementation steps of S4 are as follows: S41, multi-level risk threshold discrimination; The central decision-making unit adopts a four-level risk level classification, and each level corresponds to a specific probability interval and response strategy. The setting of risk thresholds is based on a large amount of experimental data and on-site operation experience, and is statistically verified to ensure high accuracy and low false alarm rate. The specific risk level classification is as follows: Low risk level: When it indicates that the system is in a sub-healthy state, with slight abnormalities but no immediate danger. At this time, mainly strengthen monitoring and provide reminder information for operation and maintenance personnel.

[0054] Medium risk level: When it indicates that the degree of system abnormality has further increased, and preventive measures need to be taken. At this stage, the battery load is usually reduced by restricting the charge and discharge power to slow down the development of the abnormal state.

[0055] High risk level: When it indicates that the system is approaching a dangerous state, and emergency preventive measures need to be taken. At this time, the battery circuit will be disconnected and the fire protection system warning preparation will be started.

[0056] Extremely high risk level: When it indicates that a thermal runaway accident is about to occur or has already started, and the highest-level emergency response measures need to be taken immediately.<00​​​​S42, tiered early warning instruction generation: The central decision-making unit generates corresponding early warning instructions based on the risk level. Each instruction has a clearly defined execution target, execution content, and time requirements. Instruction generation follows a standardized communication protocol to ensure that each subsystem can correctly parse and execute it. Low-risk level response measures include enhanced status monitoring instructions, increasing the data acquisition frequency from 1Hz to 5Hz, extending the data retention period from 7 days to 30 days, and activating detailed trend analysis functions; email and SMS notifications are sent to relevant personnel through the operation and maintenance management system to remind them to pay attention to the system status and suggest arranging routine checks.

[0059] Medium-risk response measures include charging and discharging power limitation instructions, limiting the maximum charging power to 80% of the rated power and the maximum discharging power to 70% of the rated power, thereby mitigating the development of abnormal conditions by reducing battery load; and notifying professional technicians to arrive on-site to conduct detailed equipment inspections, including electrical connection checks, cooling system checks, and sensor calibration.

[0060] High-risk response measures include an emergency circuit breaker command from the battery management system, which disconnects the battery from the external circuit via the main contactor, stops charging and discharging operations, and places the battery in a safe state; activating the fire protection system's early warning mode; and checking the status of fire extinguishing devices to ensure they can be activated immediately if needed.

[0061] Extremely high risk level response measures include a forced circuit breaker command for the battery management system, which completely isolates the battery system through multi-level redundant circuit breakers, disconnecting all possible current paths; immediately activating the gas extinguishing system or water sprinkler system, selecting the appropriate extinguishing method according to the type of fire; and issuing an evacuation alarm through the emergency broadcast system to guide on-site personnel to evacuate to a safe area according to the predetermined route.

[0062] S43 enables multi-system coordinated execution. This is achieved through standardized industrial communication protocols, ensuring that early warning commands are quickly and accurately transmitted to all execution systems. The system employs a redundant communication design, automatically switching to a backup link in the event of a primary communication link failure, guaranteeing communication reliability.

[0063] The central decision-making unit transmits battery control commands to the battery management system via the CAN bus communication interface. The CAN bus is a widely used real-time communication protocol in industrial settings, characterized by high reliability, strong anti-interference capabilities, and deterministic real-time response. The command frame format conforms to the SAE J1939 standard and includes information such as command type, priority, and execution time.

[0064] Fire control commands are transmitted to the fire alarm system via an Ethernet communication interface. The Ethernet protocol supports larger data transmission volumes and more complex command formats, making it suitable for transmitting fire control commands containing detailed parameters. Communication utilizes the TCP / IP protocol stack to ensure the integrity and reliability of data transmission.

[0065] The system transmits personnel notification instructions to the operations and maintenance management system via a wireless communication interface. The advantage of wireless communication is its wide coverage, enabling timely notification of operations and maintenance personnel not currently on-site. The system supports multiple notification methods, including app push notifications, SMS, and email.

[0066] The time from the completion of risk probability calculation to the issuance of the instruction shall not exceed 1 second, the time from the issuance of the instruction to the execution system response shall not exceed 3 seconds, and the total time of the entire early warning response process shall not exceed 5 seconds. This time requirement is based on the development timescale of thermal runaway accidents to ensure sufficient time window for taking effective countermeasures.

[0067] The thermal runaway early warning method for sodium battery energy storage systems of this invention monitors microscopic damage inside the battery through acoustic emission signals. It can detect abnormalities hours or even days before macroscopic parameters become abnormal, providing an early warning time 6-24 hours earlier than traditional methods. Multi-source information fusion technology reduces the probability of false alarms from single sensors, and intelligent judgment is achieved through artificial intelligence algorithms, achieving an early warning accuracy rate of over 95% and a false alarm rate controlled below 2%. Outputting risk probabilities with continuous values ​​of 0-1 provides more accurate risk quantification compared to traditional qualitative judgments, supporting refined risk management and decision-making. Four risk levels correspond to different response strategies, avoiding a one-size-fits-all approach and maximizing economic benefits while ensuring safety. Simultaneously, a single system realizes the entire process of monitoring, analysis, early warning, and response, reducing system interface and maintenance costs and improving overall reliability.

[0068] This invention also designs a thermal runaway early warning device for a sodium battery energy storage system. It adopts a hierarchical distributed architecture design. The system consists of four main functional modules: a sensing module, an edge processing module, a central decision-making module, and an execution module. It realizes local processing of data acquisition, centralized control of decision analysis, and multi-system execution response. The modules of the system establish a two-way communication link for data transmission and control commands through a standardized industrial communication protocol, forming a complete closed-loop control system.

[0069] The high-frequency acoustic emission sensor employs a piezoelectric ceramic transducer structure. The sensor body consists of a piezoelectric ceramic element, a metal housing, a signal conditioning circuit, and a connector. The piezoelectric ceramic element is made of lead zirconate titanate (PZT) material, which has a high voltage constant and good temperature stability. The operating frequency range is 20kHz to 1MHz, and the resonant frequency is set around 150kHz to obtain optimal sensitivity.

[0070] High-frequency acoustic emission sensors are fixedly mounted on the outer surface of the sodium battery cell casing. Each cell cluster is equipped with no fewer than two sensors, installed near the positive and negative terminals of the cell, respectively. The distance between the sensor and the cell casing is controlled within 5mm, and a dedicated mechanical clamp ensures a tight fit between the sensor and the casing surface. The installation location avoids areas such as welded seams and labels on the cell that may affect sound wave propagation, while also ensuring that the sensor does not interfere with normal battery assembly and maintenance operations.

[0071] The sensor forms an acoustic coupling connection with the battery cell housing via a high-temperature resistant acoustic coupling agent. The acoustic coupling agent is made of silicon-based material, with an operating temperature range of -40℃ to +85℃ and an acoustic impedance matching coefficient greater than 0.9, ensuring efficient transmission of sound waves between the sensor and the battery cell without significant attenuation. The coupling agent layer thickness is controlled between 0.1-0.5mm; too thick a layer will cause high-frequency signal attenuation, while too thin a layer cannot effectively fill the gaps caused by surface roughness.

[0072] The output signal of each acoustic emission sensor is transmitted to the corresponding edge computing unit via a coaxial cable. Differential signal transmission is used to improve anti-interference capability. The coaxial cable is a low-loss cable with a characteristic impedance of 50Ω, and its length is controlled within 10m to avoid signal attenuation and phase distortion. The sensor output signal amplitude range is -5V to +5V, and the frequency bandwidth is DC to 1MHz.

[0073] The microwave resonant cavity sensor adopts a cylindrical resonant cavity structure, mainly composed of a resonant cavity body, a microwave transmitter, a microwave receiver, a frequency synthesizer, and signal processing circuitry. The resonant cavity body is made of high-conductivity copper, and its inner wall is precision-machined and silver-plated to reduce microwave loss. The cavity dimensions are precisely designed according to the operating frequency to obtain stable resonant characteristics.

[0074] The microwave resonant cavity sensor is suspended in the space above the battery module inside the battery compartment, 20-50cm from the top of the battery module, ensuring that the resonant cavity can effectively cover the gas space above the battery module. The sensor is fixed to the top of the battery compartment by an adjustable-height bracket made of non-conductive engineering plastic to avoid interference with the microwave signal. Each battery compartment is equipped with 1-2 microwave resonant cavity sensors, with a spacing of no less than 1m between the sensors to avoid mutual interference.

[0075] The resonant cavity's opening faces the battery module, forming a closed or semi-closed gas detection chamber. The chamber volume matches the gas space volume above the monitored battery module, with the matching ratio controlled between 1:5 and 1:10 to ensure representativeness and sensitivity of the detection. The inner wall of the resonant cavity is coated with a conductive material to form an electromagnetic shielding layer, preventing external electromagnetic interference from affecting measurement accuracy.

[0076] The microwave transmitter emits 2.4 GHz or 5.8 GHz microwave signals into the cavity, with the transmission power controlled within the range of 10-100 mW. The receiver receives the reflected signal from inside the cavity and calculates the resonant frequency change by comparing the frequency difference between the transmitted and received signals. The sensor's output signal is a digital signal, transmitted to the edge computing unit via an RS485 bus, with a data update frequency of 1 Hz.

[0077] The electrical parameter sensor subsystem includes three subsystems: a voltage sensor, a current sensor, and a temperature sensor, which are responsible for monitoring the electrical and thermal states of the battery, respectively.

[0078] The voltage sensor employs a differential amplifier and a precision resistor divider network, with a measurement range of 0-5V, an accuracy class of 0.1%, and a response time of less than 1ms. The sensor is directly connected to the positive and negative terminals of the sodium battery module via a dedicated voltage acquisition cable. The connecting cable uses twisted-pair shielded wire with a diameter of at least 1.5mm² and a length limited to 5m to minimize voltage drop and noise interference.

[0079] The current sensor utilizes the Hall effect principle and consists of a Hall element, signal conditioning circuitry, and an iron core. The measurement range is determined by the battery system capacity; for a 1MWh system, it is typically 0-2000A, with an accuracy class of 0.5% and a response time of less than 10ms. The sensor employs a through-hole mounting method, where the battery's main circuit wire passes through the center hole of the sensor's iron core, enabling current measurement without disconnecting the main circuit. The sensor is installed on the busbar of the battery module, with one current sensor configured for each battery cluster.

[0080] The temperature sensor employs a resistance temperature detector (RTD) using a Pt100 platinum resistance element, with a measurement range of -50℃ to +150℃, accuracy class A, and a response time of less than 5 seconds. The sensor housing is made of stainless steel, offering excellent corrosion resistance and thermal conductivity. Each battery module is equipped with 4-6 temperature sensors, installed at key temperature monitoring points such as the cell surface, busbar connection points, and cooling system inlets and outlets. The sensors are connected to the data acquisition system via a three-wire connection, using shielded twisted-pair cable to improve measurement accuracy.

[0081] All electrical parameter sensors output analog signals with a range of 4-20mA or 0-10V. The signals are transmitted via shielded cables to the analog input module of the central decision unit, with a transmission distance of up to 100m. The central decision unit incorporates a high-precision ADC (Analog-to-Digital Converter) with a 16-bit sampling resolution and a 10Hz sampling frequency, ensuring accurate measurement and real-time monitoring of electrical parameters.

[0082] The edge computing unit adopts an embedded computing architecture, mainly composed of a microprocessor, a digital signal processor, memory, a communication interface, and a power management unit. The microprocessor is an ARM Cortex-A53 quad-core processor with a clock speed of 1.2GHz, featuring good real-time performance and low power consumption. The digital signal processor uses a dedicated DSP chip, such as TI's TMS320C6678, which has powerful floating-point arithmetic and parallel processing capabilities, specifically designed for computationally intensive tasks such as wavelet packet transform of acoustic emission signals.

[0083] The edge computing unit is equipped with 2GB of DDR4 memory for program execution and data caching, and 32GB of eMMC flash memory for storing the operating system, applications, and data files. The memory is pre-installed with wavelet packet transform function libraries, fast Fourier transform algorithm libraries, digital filter libraries, and signal processing tools, providing fundamental support for real-time signal analysis.

[0084] Each edge computing unit establishes a communication connection with the high-frequency acoustic emission sensor and microwave resonant cavity sensor of the corresponding battery module through a data acquisition interface. The data acquisition interface includes an analog input channel, a digital input / output channel, and a serial communication port. The analog input channel is used to receive analog signals from the acoustic emission sensor, with a sampling frequency of 2MHz and a sampling precision of 16 bits. The digital input / output channel is used for digital communication with the microwave resonant cavity sensor, supporting multiple communication protocols such as RS485 and CAN.

[0085] Edge computing units undertake distributed preprocessing tasks within the system. They receive configuration instructions and task assignments from the central decision-making unit, independently execute signal processing algorithms, and report the processing results to the central decision-making unit. Edge computing units possess a degree of autonomous decision-making capability; when abnormal signals are detected, they can automatically increase the sampling frequency and processing precision to ensure that critical information is not overlooked.

[0086] The wavelet packet transform processing module implements time-frequency analysis of acoustic emission signals. It employs the Daubechies wavelet basis function for 8-level wavelet packet decomposition, dividing the signal with a frequency range of 0-1MHz into 256 sub-bands. The module adopts a pipelined architecture design, enabling near real-time signal processing with a processing latency of less than 100ms. The module incorporates an adaptive threshold algorithm that automatically adjusts the signal detection threshold based on the ambient noise level.

[0087] The energy entropy calculation module is responsible for calculating the energy distribution and entropy values ​​of each frequency band, employing a parallel computing architecture to improve computational efficiency. The module supports multiple entropy calculation methods, including Shannon entropy and Renyi entropy, which can be selected according to application requirements. Calculation results are updated in real time, with an update cycle of 1 second.

[0088] The data fusion preprocessing module is responsible for synchronizing, converting, and initially fusing data from different sensors. The module uses timestamp technology to ensure data temporal consistency and employs standardization algorithms to eliminate dimensional differences between data from different sensors, providing high-quality input data for subsequent central decision analysis.

[0089] The edge computing unit establishes a data communication connection with the central decision-making unit via a CAN bus interface. The CAN bus uses two-wire differential signal transmission, with a communication rate of 500kbps, and supports multi-master communication and error detection functions. Each edge computing unit is assigned a unique node address, and the communication protocol follows the CANopen standard to ensure compatibility with other industrial equipment.

[0090] As a complement to the CAN bus, the edge computing unit is equipped with a 100Mbps Ethernet interface for large data transmission and system maintenance. The Ethernet interface supports the TCP / IP protocol stack, enabling remote monitoring and firmware upgrades.

[0091] The central decision-making unit adopts an industrial-grade embedded computing platform, mainly composed of a multi-core processor, large-capacity memory, AI accelerator, communication interface module, and human-machine interface. The processor is selected from Intel Core i7 industrial-grade processors or ARM Cortex-A78 multi-core processors, which have powerful computing capabilities and rich peripheral interfaces.

[0092] The central decision-making unit features an 8-core CPU clocked at 2.5GHz, 16GB of DDR4 memory, and a 1TB SSD, ensuring it can handle large-scale data fusion and analysis tasks. The processor supports virtualization technology, allowing multiple independent applications to run simultaneously without interference.

[0093] To improve the inference speed of deep learning models, the central decision-making unit is equipped with an AI accelerator containing NVIDIA Jetson series GPUs. The AI ​​accelerator is specifically optimized for neural network computation, providing a 10-100x performance improvement compared to CPU computation, ensuring that the thermal runaway early warning model can achieve real-time inference.

[0094] The central decision-making unit's memory pre-loads a thermal runaway early warning model developed using the TensorFlow framework. This model employs an asynchronous recurrent neural network architecture with an attention mechanism. The model file is approximately 100MB in size and includes pre-trained weight parameters and network structure definitions. The memory also contains a historical database to store nearly 30 days of runtime data, supporting trend analysis and online model optimization.

[0095] The central decision-making unit establishes communication connections with all edge computing units via a data communication bus. The communication bus adopts a star topology, with the central decision-making unit acting as the master station and each edge computing unit as a slave station. The main communication bus uses the CAN-FD protocol, with a communication rate of 5Mbps, supporting connections of 64 nodes. The backup communication bus uses industrial Ethernet, with a communication rate of 1Gbps, and automatically switches over in the event of a main bus failure.

[0096] The central decision-making unit establishes direct connections with voltage, current, and temperature sensors via a multi-channel analog input module. The analog input module is configured with 32 differential input channels, each supporting a ±10V input range, a 16-bit sampling resolution, and a 10kHz sampling frequency. The module incorporates signal conditioning circuitry, including amplifiers, filters, and isolators, to ensure accurate signal acquisition and system electrical safety.

[0097] The early warning model employs a multi-layer Long Short-Term Memory (LSTM) network architecture combined with an attention mechanism. The network consists of an input layer, three LSTM hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. Each LSTM layer contains 128 neurons, with a total of approximately 500,000 parameters. The attention mechanism uses a self-attention structure, enabling it to automatically learn the importance of each input feature at different times.

[0098] The model was trained using a large amount of historical and experimental data. The training dataset contains 1 million records of normal operation, 100,000 records of abnormal states, and 1,000 records of thermal runaway incidents. Training employed the Adam optimization algorithm with a learning rate of 0.001, a batch size of 64, and 1,000 training epochs. The model achieved an accuracy of 95.2%, a recall of 93.8%, and an F1 score of 94.5% on the validation set.

[0099] To meet real-time requirements, the model was optimized using quantization and pruning techniques, converting the 32-bit floating-point model to an 8-bit integer model, reducing the model size by 75% and increasing inference speed by 4 times. The optimized model has an inference time of less than 50ms on the central decision-making unit, meeting the system's 1-second response time requirement.

[0100] The Battery Management System (BMS), as a crucial component of the execution module, is responsible for executing battery control commands issued by the central decision-making unit. The BMS mainly consists of a main control unit, a contactor control module, a relay drive module, and a safety protection module.

[0101] The central decision-making unit connects to the battery management system via an RS485 communication interface at a communication rate of 115.2 kbps, supporting the Modbus RTU protocol. The communication cable is a shielded twisted-pair cable with a diameter of 1.5 mm², allowing for a transmission distance of up to 1000 meters. The communication interface is equipped with an opto-isolator with an isolation voltage of 2500V to ensure electrical safety between systems.

[0102] The control commands sent from the central decision-making unit to the BMS include charging enable / disable, discharging enable / disable, power limiting, and emergency circuit breaker. Upon receiving the commands, the BMS performs safety verification, including command validity checks, operation permission verification, and system status confirmation. After successful verification, the BMS executes the corresponding control operation and reports the execution result back to the central decision-making unit.

[0103] The BMS incorporates a multi-level safety protection mechanism, including software and hardware protection. Software protection uses algorithms to determine the rationality of commands and prevent misoperation; hardware protection uses an independent safety circuit to forcibly disconnect the battery circuit when a dangerous state is detected. The BMS has fail-safe characteristics, and can independently perform basic safety protection functions even if communication is interrupted.

[0104] The fire alarm linkage system is responsible for activating corresponding fire-fighting measures when the risk of thermal runaway escalates. It mainly consists of a fire controller, fire extinguishing device, smoke exhaust system and emergency lighting system.

[0105] The central decision-making unit connects to the fire alarm system via an Ethernet interface, communicating using the TCP / IP protocol at a speed of 100Mbps. The Ethernet interface supports redundant configuration, with automatic switching between the primary and backup links to ensure reliable communication. The fire alarm system is configured with an independent IP address and communication port, supporting standard automatic fire alarm system communication protocols.

[0106] The control commands sent by the central decision-making unit to the fire alarm linkage system include early warning preparation, system self-test, fire extinguishing activation, and smoke extraction activation. Upon receiving the commands, the fire alarm linkage system first performs a system status check to confirm that all subsystems are ready, and then executes fire-fighting measures according to the predetermined procedures. The system supports both manual and automatic control modes. In automatic mode, it executes entirely according to the commands of the central decision-making unit, while in manual mode, on-site operators can directly control the system.

[0107] Based on the characteristics of sodium batteries, the fire suppression system combines gaseous extinguishing and water spraying. The gaseous extinguishing system uses clean gas as the extinguishing medium, which will not cause secondary damage to the battery system. The water spraying system is used for cooling down the battery in the event of thermal runaway.

[0108] The operation and maintenance management system is responsible for receiving and processing personnel notifications and instructions, ensuring that relevant personnel can understand the system status in a timely manner and take appropriate actions.

[0109] The central decision-making unit connects to the operation and maintenance management system via a wireless communication interface, supporting multiple communication methods such as WiFi and 4G / 5G. The wireless communication module is equipped with multiple carrier SIM cards, supporting automatic network switching to ensure communication reliability. The communication interface employs VPN encryption technology to guarantee data transmission security.

[0110] The operation and maintenance management system receives information including system status reports, early warning notifications, fault alarms, and maintenance reminders. The information is encapsulated in JSON format and includes fields such as timestamp, event type, severity, and detailed description. The system supports multiple notification methods, including mobile app push notifications, SMS, email, and voice calls, ensuring that information is delivered to relevant personnel in a timely manner.

[0111] The operation and maintenance management system offers two user interfaces: a web interface and a mobile app. The web interface is suitable for large-screen monitoring in the control room, displaying information such as the overall system status, real-time data curves, and historical trend analysis. The mobile app is suitable for on-the-go monitoring by operation and maintenance personnel, supporting functions such as real-time alarm push notifications, remote system control, and maintenance record management.

[0112] The central decision-making unit features a human-machine interface for displaying system operating status and early warning information. The display system consists of an industrial-grade LCD screen, a touchscreen controller, and a graphics processing module. The screen is a 15-inch industrial-grade LCD with a resolution of 1024×768, a brightness of 500 cd / m², a contrast ratio of 1000:1, and an operating temperature range of -20℃ to +70℃. The touchscreen uses resistive touch technology, supports multi-touch, has a touch accuracy of ±1mm, and a lifespan exceeding 10 million touches.

[0113] The human-machine interface (HMI) adopts a graphical design, displaying system topology diagrams, key parameter values, and warning status indicators on the main screen. The interface supports multi-level menus, allowing users to view detailed information step-by-step. Warning information uses color coding: green for normal, yellow for caution, orange for warning, red for danger, and flashing red for emergency. Through the HMI, operators can view system status, set parameters, perform manual operations, and view historical data. The interface includes access control, with different user levels having different permissions. All operations are logged in detail, supporting operation traceability and auditing. The system uses dual AC 220V power supplies, with automatic switching between main and backup power in less than 10ms. Each power supply is regulated by a UPS (Uninterruptible Power Supply) to ensure power quality and reliability. The UPS capacity is designed for 4 hours of full-load operation, and the batteries are maintenance-free lead-acid batteries. Internally, the system uses multiple DC power supplies, including ±24V, ±12V, +5V, and +3.3V voltage levels. Each power supply is equipped with overvoltage, undervoltage, and overcurrent protection circuits. The power supply module employs switching power supply technology, achieving an efficiency greater than 90% and a ripple of less than 50mV. The system power supply utilizes a multi-stage filtering design, including common-mode filters, differential-mode filters, and power supply filters, ensuring electromagnetic compatibility. All power cables are shielded, with the shielding layer uniformly grounded within the distribution cabinet. The system adopts a modular integration approach, with each functional module developed and tested independently before integration via standard interfaces.

[0114] This system is the first to combine acoustic emission sensing, microwave resonant cavity sensing, and traditional electrical sensing technologies, achieving comprehensive monitoring of the internal state of sodium batteries. Compared to traditional single-sensor methods, multimodal fusion technology can detect abnormal signs earlier and more accurately, with an average warning time of more than 6 hours. The system employs a deep learning algorithm specifically designed for sodium battery thermal runaway early warning, adaptively analyzing the importance of data from each sensor through an attention mechanism to achieve dynamic weight allocation and intelligent decision-making. The algorithm achieves a warning accuracy rate of over 95% and a false alarm rate controlled below 2%. The system adopts an edge computing architecture, distributing signal preprocessing tasks to various edge computing units, which reduces the burden on the central processing unit and improves the system's real-time performance and reliability. The edge computing units have a certain degree of autonomous decision-making capability and can operate independently when communication is interrupted.

[0115] The system adopts a modular design, with each functional module relatively independent, facilitating system upgrades and functional expansion. Adding a new battery module only requires adding the corresponding sensors and edge computing units, without modifying the central decision-making system. The modular design also reduces system maintenance costs and the scope of failure impact. The system establishes a multi-level security system encompassing sensor level, edge computing level, central decision-making level, and execution level, with each level possessing independent security protection functions. Even if a level fails, other levels can still provide basic security protection, ensuring the system's fail-safe characteristics.

[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for thermal runaway early warning of a sodium battery energy storage system, characterized in that, The method comprises the following steps: S1, real-time data acquisition by multiple sensors; synchronous acquisition of multi-modal monitoring data by high-frequency acoustic emission sensors, microwave resonant cavity sensors and electrical sensors deployed in the battery cabin of the sodium battery energy storage system to monitor internal state changes of the sodium battery; S2, multi-source data preprocessing and feature extraction; the original sensor data collected in step S1 are input into an edge computing unit for preprocessing, including extracting feature indexes reflecting micro-damage inside the battery from acoustic emission signals through a signal processing algorithm, and extracting feature indexes reflecting gas environment abnormalities of the battery cabin from dielectric property changes; S3, multi-source information fusion and risk assessment; the feature indexes extracted in step S2 are transmitted to a central decision unit together with voltage, current and temperature electrical parameters, and the central decision unit calculates the thermal runaway risk of the current sodium battery energy storage system; S4, early warning decision and execution; when the thermal runaway risk probability calculated in step S3 exceeds a preset safety threshold, the central decision unit generates a graded early warning instruction and transmits it to the battery management system and / or the fire linkage system through a communication interface.

2. The thermal runaway early warning method for a sodium battery energy storage system according to claim 1, wherein, The step S1 specifically comprises: S11, acoustic emission signal acquisition; the high-frequency acoustic emission sensor acquires high-frequency acoustic emission signals generated by internal electrochemical reactions, mechanical stress changes and micro-crack propagation during the operation of the battery through close coupling contact with the sodium battery cell shell; S12, gas dielectric property acquisition; the microwave resonant cavity sensor measures the resonance frequency change of the environment by emitting microwave signals of a specific frequency to the space above the battery module and receiving reflected signals according to the formula ； a relative change in the dielectric constant is calculated; wherein is the relative change in the dielectric constant, dimensionless; is a calibration factor related to the resonant cavity geometry and material properties, the microwave resonant cavity sensor being arranged within the battery compartment for detecting changes in the dielectric constant of the gas within the battery compartment using its resonant cavity structure; is the difference between the current measured resonant frequency and the reference resonant frequency; is the reference resonant frequency established under standard healthy conditions; S13, electrical parameter acquisition; the voltage sensor and the current sensor are connected to the positive and negative terminals of the sodium battery module, and the temperature sensor is connected to several temperature monitoring points of the sodium battery module; end voltage, charge and discharge current and monitoring point temperature data reflecting the electrical and thermal states of the battery are acquired in real time.

3. The thermal runaway early warning method for sodium battery energy storage systems of claim 1, wherein, The step S2 specifically comprises: S21, acoustic emission signal feature extraction; the edge computing unit receives the original acoustic emission signals collected by the high-frequency acoustic emission sensor, decomposes the time domain signals into different frequency bands through wavelet packet transform algorithm, and identifies the characteristic frequency bands related to the abnormal state inside the battery; S22, acoustic emission energy entropy calculation; the edge computing unit calculates the energy entropy of the feature frequency band identified in step S21 The wavelet packet energy entropy is calculated as the acoustic emission feature index, and the formula is ； performing the calculation; wherein is an acoustic emission energy entropy characteristic index, dimensionless; is a normalized energy ratio of the th sub-band, calculated by ; is a wavelet packet energy of the th sub-band in the frequency band ; is a total energy in the analysis band; S23, battery cabin gas characteristic quantity transmission; the edge computing unit transmits the relative change of the dielectric constant of the gas in the battery cabin calculated in step S12 as a characteristic index reflecting the abnormality of the gas environment in the cabin, together with the acoustic emission characteristic index to the central decision unit.

4. The sodium battery energy storage system thermal runaway early warning method of claim 1, wherein, The step S3 specifically comprises: S31, asynchronous time series input processing; the central decision unit receives asynchronous time series data from different sensor channels, and constructs a comprehensive input vector ； wherein is a comprehensive input vector at the moment; is an acoustic emission energy entropy feature index at the moment; is a relative change of the dielectric constant of the gas in the battery cabin at the moment; is a battery end voltage measurement value at the moment, in volts (V); is a charge and discharge current measurement value at the moment; is a key point temperature measurement value at the moment; S32, attention weight calculation; the early warning model analyzes the importance of each input feature to the current state judgment through the attention mechanism, and calculates a dynamic weight vector ； wherein is the attention weight vector for each input feature at time is a normalized exponential function; is an attention score function for calculating a relevance score of the hidden state and the input feature; is the hidden state vector of the model at time is the hidden state vector of the model at time S33, thermal runaway risk probability calculation; the early warning model calculates the thermal runaway risk probability based on the weighted input features and historical state information through the formula ； calculating a thermal runaway risk probability at the current time; wherein is a thermal runaway risk probability value at the current time, ranging from 0 to 1; sigmoid is a sigmoid activation function; W is a weight matrix obtained by training; represents an element-wise multiplication operator; bias is a bias vector.

5. The thermal runaway early warning method for sodium battery energy storage systems of claim 1, wherein, The step S4 specifically comprises: S41, multi-level risk threshold discrimination; the central decision unit compares the thermal runaway risk probability calculated in step S33 with the preset four-level safety threshold in sequence determines as low risk level when determines as medium risk level when determines as high risk level when determines as extremely high risk level when determines as extremely high risk level. S42, hierarchical early warning instruction generation; the central decision unit generates corresponding early warning instructions according to the risk level determined in step S41, generates state monitoring enhancement instructions and operation and maintenance personnel reminder notifications when it is a low risk level, generates charge and discharge power limitation instructions and technical personnel on-site inspection notifications when it is a medium risk level, generates battery management system emergency shutdown instructions and fire extinguishing system early warning preparation instructions when it is a high risk level, and generates battery management system forced shutdown instructions, fire extinguishing system immediate start instructions and personnel emergency evacuation instructions when it is an extremely high risk level; S43, multi-system linkage execution; the central decision unit transmits the battery control instructions to the battery management system through the CAN bus communication interface, transmits the fire control instructions to the fire linkage system through the Ethernet communication interface, and transmits the personnel notification instructions to the operation and maintenance management system through the wireless communication interface, so as to realize multi-system collaborative response and hierarchical treatment of thermal runaway risk.

6. A thermal runaway early warning device for a sodium battery energy storage system, for implementing the method of any one of claims 1 to 5, characterized in that, It comprises a sensing module, an edge processing module, a central decision module and an execution module; The sensing module comprises a high-frequency acoustic emission sensor, a microwave resonant cavity sensor, a voltage sensor, a current sensor and a temperature sensor, which are used for realizing real-time data acquisition; the high-frequency acoustic emission sensor is fixedly installed on the outer surface of the shell of the sodium battery cell, and is acoustically coupled to the cell shell through a high-temperature coupling agent; the microwave resonant cavity sensor is suspendedly installed above the battery module in the battery cabin, and the resonant cavity opening thereof faces the battery module to form a gas detection chamber; the voltage sensor, the current sensor and the temperature sensor are respectively connected to the electrical terminals and temperature monitoring points of the sodium battery module; The edge processing module comprises a plurality of edge computing units, which are used for realizing multi-source data preprocessing and feature extraction; each edge computing unit is communicatively connected to the high-frequency acoustic emission sensor and the microwave resonant cavity sensor of the corresponding battery module through a data acquisition interface, and a digital signal processor and a memory are built-in the edge computing unit, which are used for executing wavelet packet transform algorithm and energy entropy calculation algorithm; The central decision module comprises a central decision unit, which is used for realizing multi-source information fusion and risk assessment; the central decision unit is communicatively connected to all edge computing units through a data communication bus, and is directly connected to the voltage sensor, the current sensor and the temperature sensor through a sensor interface; a processor and a pre-trained thermal runaway early warning model are built-in the central decision unit; The execution module comprises a battery management system and a fire linkage system, which are used for realizing early warning decision and execution; the central decision unit is communicatively connected to the battery management system and the fire linkage system through a control communication interface.

7. The thermal runaway early warning device for sodium battery energy storage systems of claim 6, wherein, The high-frequency acoustic emission sensor adopts a piezoelectric ceramic transducer structure; the high-frequency acoustic emission sensor is connected to the sodium battery cell shell through a high-temperature acoustic coupling agent.

8. The thermal runaway early warning device for sodium battery energy storage systems of claim 6, wherein, The microwave resonant cavity sensor adopts a cylindrical resonant cavity structure; the volume of the resonant cavity of the microwave resonant cavity sensor matches the volume of the gas space above the monitored battery module, and the inner wall of the resonant cavity is coated with a conductive material to form an electromagnetic shielding layer.

9. The thermal runaway early warning device for sodium battery energy storage systems of claim 6, wherein, The edge computing unit is built-in with a microprocessor of ARM architecture and a special digital signal processor; the edge computing unit is pre-installed with a wavelet packet transform function library and a fast Fourier transform algorithm library in a memory; and the edge computing unit is connected with the central decision unit through a CAN bus interface to establish a data communication connection.

10. The thermal runaway early warning device for sodium battery energy storage systems of claim 6, wherein, The central decision unit adopts an industrial embedded computing platform and is built-in with a multi-core processor and a large-capacity memory; the battery management system is connected with the central decision unit through an RS485 communication interface to receive a circuit breaking control instruction; the fire-fighting linkage system is connected with the central decision unit through an Ethernet interface to receive a pre-cooling starting instruction; and the central decision unit is further provided with a man-machine interaction interface to display system running states and early warning information.

Images