Real-time monitoring and early warning method and device for electrochemical energy storage power station

Abstract

The invention discloses a real-time monitoring and early warning method and device for an electrochemical energy storage power station. The method comprises the following steps: acquiring multi-parameter time sequence data; carrying out fusion analysis and state prediction on the data; establishing and online updating a self-adaptive safety threshold value of each parameter; matching the real-time observation sequence with a pre-constructed multi-mode fault evolution knowledge graph to obtain a mode matching confidence coefficient; predicting the probability that the battery is in a dangerous state in the future, and taking the probability as state prediction confidence; calculating the standard exceeding amplitude and duration of the current parameter value relative to the safety threshold, and obtaining threshold standard exceeding confidence; carrying out weighted fusion on the mode matching confidence coefficient, the state prediction confidence coefficient and the threshold exceeding confidence coefficient, and calculating to obtain a comprehensive risk confidence coefficient; and according to the comprehensive risk confidence, triggering an early warning signal and an emergency response measure of a corresponding level. According to the invention, the safety of the energy storage power station is effectively improved, the stable operation of the energy storage power station is ensured, and the accuracy, robustness and adaptability of early warning are improved.

Description

A method and device for real-time monitoring and early warning of electrochemical energy storage power stations Technical Field

[0001] This invention relates to a real-time monitoring and early warning method in the field of battery monitoring technology, and more particularly to a real-time monitoring and early warning method for electrochemical energy storage power stations, and also to a real-time monitoring and early warning device for electrochemical energy storage power stations. Background Technology

[0002] With the continuous development of power systems, batteries, as a crucial component of energy storage, have become increasingly important for monitoring their safe operating status. During charging and discharging, batteries may generate hazardous factors such as temperature increases and gas releases. Failure to monitor and address these hazards in a timely manner could lead to serious safety accidents. Therefore, real-time monitoring and dynamic data acquisition of battery operating status are of great significance for ensuring the safe operation of power systems.

[0003] Currently, there are various monitoring systems based on Internet of Things (IoT) technology on the market. For example, CN112185090A discloses a remote monitoring system and method for agricultural greenhouses based on NB-IoT. This system consists of a main processor, a data acquisition module, a data storage module, a power supply module, a positioning module, a wireless communication module, a cloud server, and a user management terminal, achieving remote monitoring through NB-IoT wireless communication. CN111552219A proposes a coalbed methane storage and transportation end-to-end monitoring system and method, which, by setting up a main control device, a data acquisition device, and a remote monitoring device, monitors key environmental information, safety status, and logistics information throughout the coalbed methane storage and transportation process.

[0004] In the field of gas monitoring, CN107888414A discloses a biogas leak monitoring system based on NB-IoT. This system includes a methane concentration detection module, a main controller, a display module, an audible and visual alarm module, a power supply module, an NB wireless transmission module, an IoT cloud platform, and a user terminal, enabling the detection of biogas leaks and centralized monitoring on the platform. CN110057971A introduces a multi-component toxic gas monitoring device based on IoT, including multiple gas sensor detection modules, a main control unit module, a display and alarm module, and a wireless transmission module. It can monitor the content of multiple toxic gases in real time and transmit relevant information through IoT technology.

[0005] However, existing data acquisition terminals still have the following problems in the field of battery monitoring: 1. Incomplete data acquisition: Existing monitoring systems often only focus on a single or a few parameters, such as temperature and voltage, while ignoring the comprehensive monitoring of various dangerous gases (such as hydrogen sulfide, methane, carbon monoxide, etc.) generated during battery charging and discharging, resulting in the inability to comprehensively assess the battery's operating status.

[0006] 2. Poor monitoring timeliness: Most systems use a fixed frequency data acquisition method, which cannot dynamically adjust the acquisition frequency according to the battery status. In case of abnormal situations, they cannot respond in time, thus delaying the early warning opportunity.

[0007] 3. Incomplete data processing: The existing system lacks effective fusion analysis of the collected data, and cannot achieve multi-parameter collaborative judgment, which reduces the accuracy and reliability of monitoring.

[0008] 4. Inadequate energy management: Existing terminal devices generally suffer from high power consumption and poor battery life, requiring frequent battery replacements or charging, which increases maintenance costs and workload.

[0009] These issues lead to inaccurate and untimely analysis of the operating status of power plant batteries, increasing operation and maintenance costs and safety risks. Therefore, there is an urgent need for a technology that can achieve multi-parameter fusion, real-time monitoring, and dynamic data acquisition to improve the comprehensiveness, accuracy, and timeliness of battery operating status monitoring. Summary of the Invention

[0010] To address the technical problems of incomplete data collection, poor timeliness, and incomplete data processing in existing monitoring and early warning methods for electrochemical energy storage power stations, this invention provides a real-time monitoring and early warning method and device for electrochemical energy storage power stations.

[0011] This invention employs the following technical solution: a real-time monitoring and early warning method for electrochemical energy storage power stations, comprising the following steps: real-time acquisition of battery temperature, concentrations of various characteristic gases, voltage, current, and internal resistance parameters to obtain multi-parameter time series data; fusion analysis and state prediction of the time series data: based on historical operating data under battery health conditions, establishing and updating adaptive safety thresholds for each parameter online; matching the real-time observation sequence with a pre-constructed multimodal fault evolution knowledge graph to obtain pattern matching confidence; inputting the real-time observation sequence into a pre-trained Hidden Markov Model to predict the probability of the battery being in a dangerous state in the future, and using this as the state prediction confidence; calculating the magnitude and duration of the current parameter value exceeding the safety threshold to obtain threshold exceeding confidence; weighted fusion of the pattern matching confidence, the state prediction confidence, and the threshold exceeding confidence to calculate a comprehensive risk confidence; triggering corresponding level early warning signals and emergency response measures according to the preset numerical range of the comprehensive risk confidence.

[0012] This invention provides real-time monitoring of power station batteries. It collects and stores multi-timescale sensor data related to the safety of batteries and critical circuits in chemical energy storage power stations, enabling timely detection of abnormal data trends. This effectively improves the safety and ensures the stable operation of energy storage power stations. Through four innovative modules—dynamic thresholds, multimodal knowledge graphs, hidden Markov model state prediction, and multi-source fusion decision-making—it significantly improves the accuracy, robustness, and adaptability of early warning systems for batteries. It is particularly suitable for the health management of batteries in electrochemical energy storage power stations under various models and operating conditions, solving the technical problems of incomplete data collection, poor timeliness, and incomplete data processing in existing monitoring and early warning methods for electrochemical energy storage power stations.

[0013] As a further improvement to the above scheme, the adaptive safety threshold update method includes the following steps: based on the statistical 3σ principle, calculate the initial mean and standard deviation of each parameter data sampled over a long period of time under battery health conditions, and set an initial safety threshold; use an online statistical algorithm to dynamically update the mean and variance of the parameters using the latest data within the sliding window, and update the adaptive safety threshold in real time; introduce the battery health status value as feedback, and dynamically adjust the size of the sliding window according to the battery health status value to adapt to the degradation of battery performance throughout its entire life cycle.

[0014] As a further improvement to the above scheme, the method for constructing the multimodal fault evolution knowledge graph includes the following steps: under controllable laboratory conditions, inducing the battery to experience multiple known faults, and simultaneously collecting the time series data from the initial stage of the fault to the entire process of thermal runaway; dividing the time series data into stages and labeling states based on characteristic inflection points; and based on the labeled data, mining frequently occurring state association rules through data mining methods to construct a graph representing the evolution law of parameters and the relationship between state transitions.

[0015] As a further improvement to the above scheme, the hidden state set of the Hidden Markov Model is defined as S = {S1: normal, S2: initial stage of thermal abuse, S3: later stage of thermal abuse, S4: thermal diffusion stage, S5: thermal runaway stage}, and the initialization parameters of the observed state set O are defined as λ = (A, B, π), π = [1, 0, 0, 0, 0]^T, A is the state transition matrix, B is the observation probability matrix, and λ is trained unsupervised by inputting labeled battery fault sequence data; the calculation method of the state prediction confidence includes the following steps: the real-time observation sequence O1, O2, ..., O t Input the Hidden Markov Model and use the forward algorithm to calculate the probability P(S) that the battery will be in a thermal runaway state at the k-th time step in the future. t+k = S5 | O1, O2, ..., O t ), where t represents time.

[0016] As a further improvement to the above scheme, the calculation formula for the threshold overshoot confidence level is: C3 = amplitude factor * duration factor; where C3 is the threshold overshoot confidence level, the amplitude factor is obtained by mapping the relative amplitude of the current parameter value exceeding the adaptive safety threshold through an S-shaped function, and the duration factor is obtained by mapping the duration of the parameter continuously exceeding the adaptive safety threshold through an exponential decay function.

[0017] As a further improvement to the above scheme, the formula for calculating the comprehensive risk confidence level is: C final = ω1 * C1 + ω2 * C2 + ω3 * C3; where C final Let C1, C2, and C3 be the pattern matching confidence, the state prediction confidence, and the threshold exceeding confidence, respectively, and let ω1, ω2, and ω3 be the corresponding weight coefficients, satisfying ω1+ω2+ω3=1.

[0018] Furthermore, the weighting coefficients are dynamically adjusted according to different stages of battery fault development: during the fault latency period, ω2>ω1>ω3 is set to enhance sensitivity to weak trends; during the fault development period, ω1=ω2>ω3 is set to perform cross-validation by combining trend prediction and pattern matching; and during the late stage of the fault, ω1>ω3>ω2 is set to rely on clear fault patterns and direct evidence of exceeding limits.

[0019] Furthermore, the preset numerical range includes at least four levels, each corresponding to a different response measure: when 0.1 <C final When the value is ≤ 0.3, it is defined as Level 1 monitoring level, and only exception logs are recorded; when it is 0.3... <C final When the value is ≤ 0.5, it is defined as Level 2 alert level, and a warning notification is sent to the user terminal; when it is 0.5... <C final When C is ≤ 0.8, it is defined as Level 3 warning level, triggering the local audible and visual alarm on the data acquisition terminal and sending an emergency alarm signal to the cloud; when C final When the value is greater than 0.8, it is defined as Level 4 hazard level, and the highest level emergency response is executed. The highest level emergency response includes linking the fire protection system and emergency circuit cutoff.

[0020] This invention also provides a real-time monitoring and early warning device for electrochemical energy storage power stations, which applies any of the aforementioned real-time monitoring and early warning methods for electrochemical energy storage power stations. The real-time monitoring and early warning device includes: a data acquisition terminal, comprising a main control center, a wireless communication module, a display module, an audible and visual alarm module, and a sensor array for collecting temperature, characteristic gas concentration, voltage / current, and internal resistance, and for collecting and uploading multi-parameter time-series data and executing local early warning; the display module for displaying the collected parameter information; the audible and visual alarm module for detecting whether the parameter information exceeds a set threshold, triggering an audible and visual alarm when the parameter is at a dangerous value, and transmitting the alarm information to the main control center; the main control center for summarizing the parameter information collected by the sensor array, uploading it through the wireless communication module, and adjusting the working mode of the data acquisition terminal upon receiving alarm information; a cloud server for receiving and storing the data uploaded by the data acquisition terminal, performing fusion analysis and state prediction on the time-series data, and executing fusion analysis, state prediction, and graded early warning decisions; and a user terminal for receiving and displaying battery status information and early warning information from the cloud server.

[0021] As a further improvement to the above scheme, the sensor array includes a temperature and humidity sensor, a hydrogen sulfide concentration sensor, a methane concentration sensor, a carbon monoxide concentration sensor, and a voltage / current sensor arranged around the circumference of the battery under test; wherein, the data acquisition terminal is generally cylindrical in shape.

[0022] Compared with existing monitoring and early warning methods and devices for electrochemical energy storage power stations, the real-time monitoring and early warning method and device for electrochemical energy storage power stations of the present invention has the following beneficial effects: 1. The real-time monitoring and early warning method for electrochemical energy storage power stations can monitor the power station batteries in real time, collect and store multi-timescale sensor data related to the safety of batteries and key circuits in the electrochemical energy storage power station in real time, and can capture the changing trend of abnormal data in a timely manner, effectively improving the safety of the energy storage power station and ensuring the stable operation of the energy storage power station. Through four innovative modules, namely dynamic threshold, multimodal knowledge graph, hidden Markov model state prediction, and multi-source fusion decision, the accuracy, robustness and adaptability of early warning of batteries are significantly improved. It is especially suitable for battery health management of electrochemical energy storage power stations under multiple models and multiple operating conditions, and solves the technical problems of incomplete data collection, poor timeliness and incomplete data processing in existing monitoring and early warning methods for electrochemical energy storage power stations.

[0023] 2. This real-time monitoring and early warning method for electrochemical energy storage power stations achieves dynamic self-adaptation. Thresholds are smoothly adjusted as the battery ages without manual intervention, making it suitable for operation in resource-constrained embedded data acquisition terminals. It consumes minimal memory and computing resources, ensuring system real-time performance and low power consumption. This method draws on the ideas of information fusion and ensemble learning, greatly improving the system's robustness by integrating multiple independent and complementary information sources, enabling more reliable and accurate final decisions than any single information source.

[0024] 3. This real-time monitoring and early warning device for electrochemical energy storage power stations operates simultaneously through multiple data sensors, collaborating with a cloud server to process data, reducing the time the processor core waits for data and improving system response speed. By intelligently controlling two operating modes from the main control center—a low-power consumption mode when the battery is operating normally and a high-frequency acquisition mode when abnormalities occur—power consumption and labor costs are reduced. This device can accurately determine the operating status of the batteries in the electrochemical energy storage power station, greatly preventing safety accidents and providing strong technological support for the safe operation of the energy storage power station.

[0025] 4. This real-time monitoring and early warning device for electrochemical energy storage power stations utilizes sensor data fusion technology to collect and store multi-timescale sensor data related to the safety of batteries and critical circuits in the electrochemical energy storage power station in real time. Regarding the data acquisition frequency, it is dynamically adjusted according to the operating status of the energy storage power station. During normal operation, a lower acquisition frequency is used to reduce power consumption and data transmission pressure; when an anomaly is detected, the acquisition frequency is automatically increased to ensure timely capture of abnormal data trends. This dynamic adjustment strategy ensures both data real-time performance and improves the overall efficiency of the system. Attached Figure Description

[0026] Figure 1 is a flowchart of the real-time monitoring and early warning method for electrochemical energy storage power stations according to Embodiment 1 of the present invention.

[0027] Figure 2 is a simplified structural diagram of the real-time monitoring and early warning device for an electrochemical energy storage power station according to Embodiment 2 of the present invention.

[0028] Figure 3 is a flowchart of the real-time monitoring and early warning device for electrochemical energy storage power stations shown in Figure 2.

[0029] Figure 4 shows the concentration changes of H2S, a characteristic gas, in the real-time monitoring and early warning device for electrochemical energy storage power stations shown in Figure 2.

[0030] Figure 5 shows the concentration changes of CO, a characteristic gas, in the real-time monitoring and early warning device for electrochemical energy storage power stations shown in Figure 2.

[0031] Figure 6 shows the concentration changes of CH4, a characteristic gas, in the real-time monitoring and early warning device for electrochemical energy storage power stations shown in Figure 2. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example 1 (Referring to Figure 1) provides a real-time monitoring and early warning method for electrochemical energy storage power stations. This method applies multi-parameter fusion real-time monitoring and dynamic analysis technology in the monitoring and early warning system of electrochemical energy storage power stations. It can combine with a cloud server to analyze battery status, achieve real-time monitoring of the power station batteries, and collect and store multi-timescale sensor data related to the safety of batteries and key circuits in the electrochemical energy storage power station in real time. It can promptly capture the changing trends of abnormal data, effectively improving the safety of the energy storage power station. The main steps of this real-time monitoring and early warning method are as follows.

[0034] Step 1: Data Acquisition: Real-time acquisition of battery temperature, concentrations of various characteristic gases, voltage, current, and internal resistance parameters to obtain multi-parameter time series data. In this embodiment, data acquisition can be performed using a sensor data acquisition module. This module can sense charging and discharging hazard factors and risk characteristic parameters. It can include multiple sensors, which are arranged around the battery under test to form a comprehensive parameter monitoring network.

[0035] Step 2: Perform fusion analysis and state prediction on time series data: (1) Based on historical operating data under the battery health state, establish and update the adaptive safety thresholds of each parameter online; (2) Match the real-time observation sequence with the pre-constructed multimodal fault evolution knowledge graph to obtain the pattern matching confidence; (3) Input the real-time observation sequence into the pre-trained hidden Markov model to predict the probability that the battery will be in a dangerous state in the future, and use it as the state prediction confidence; (4) Calculate the magnitude and duration of the current parameter value exceeding the safety threshold to obtain the threshold exceeding confidence.

[0036] The adaptive safety threshold update method includes the following steps: First, based on the statistical 3σ principle, the initial mean and standard deviation of each parameter data sampled over a long period of time under battery health conditions are calculated to set the initial safety threshold; then, an online statistical algorithm is used to dynamically update the mean and variance of the parameters using the latest data within the sliding window, and the adaptive safety threshold is updated in real time; finally, the battery health status value is introduced as feedback, and the size of the sliding window is dynamically adjusted according to the battery health status value to adapt to the degradation of battery performance throughout its entire life cycle.

[0037] In this embodiment, under healthy operating conditions, the temperature, mixed gas concentration, and internal resistance of the lithium battery are continuously sampled over a sufficiently long period of time, and long-term operating data of these three parameters are obtained. Based on the statistical 3σ principle, it is assumed that the parameter data of a healthy battery follows a normal distribution, and 99.73% of the data points should fall within the (μ ± 3σ) interval. The probability of exceeding this range is extremely low, and therefore it can be used as an anomaly criterion.

[0038] During the initial learning phase, the system collects various battery parameter data at a high frequency. For each parameter sequence, it calculates the mean μ and standard deviation σ, and sets an initial safety threshold: Threshold. initial = μ + 3σ / μ - 3σ. X represents the temperature sensor value; Y represents an array of three values ​​[YCO, YH2S, YCH4] representing the concentrations of CO, H2S, and CH4; Z represents the battery internal resistance value and records the statistical characteristics of the data. The initial value is set to: [X base Y base Z base After the system enters normal operation, it records all data of the battery charge and discharge cycles within the window size, and calculates all mean (μ), standard deviation (σ), and distribution shape. The mean and standard deviation of the data within the window are updated in real time using the following algorithm: (a) Whenever a new data point x new When the time arrives, update the counter n: n = n + 1.

[0039] (b) Calculate the deviation Delta between the current data point and the old mean: Delta = x new –Mean x-1 .

[0040] (c) Update the mean: Mean x = Current_mean x-1 + Delta / n(d) Update the squared deviation and M2: M2 = M2 + delta * (x new - Mean x (e) The standard deviation required for the final threshold calculation: Variance = M2 / (n-1) Considering the degradation of battery performance over time, a battery state of health (SOH) feedback mechanism is introduced based on the original dynamic threshold learning, so that the sliding window size can be automatically adjusted according to the degree of battery aging. First, a real-time SOH evaluation module is established to calculate the state of health through battery physical parameters: (a) Calculate the rate of change of internal resistance, the ratio of the current internal resistance value to the initial internal resistance value: r R = R c / R i .

[0041] (b) Calculate the cycle decay factor, where β is the experimentally calibrated aging coefficient: C d = math.exp(-β * c) .

[0042] (c) Calculate SOH, limiting this value to a reasonable range (default 50%~90%) to avoid calculation distortion in extreme cases: SOH = r R * C d .

[0043] (d) Change the original fixed window to a dynamic adjustment mechanism. Calculate the battery's health degradation: D level = 1 - (SOH c / SOH i ) .

[0044] (e) Calculate the new window. The α-sensitivity parameter is a configurable parameter used to prevent the window size from shrinking drastically with aging; it also sets the window's boundary protection (upper and lower limits): W new = W0* (1 + α * D level ) .

[0045] (f) Finally, the adaptive mechanism is integrated into the original statistical learning: When the system starts, the fixed window of the original scheme is used for initial learning, and the initial internal resistance value of the battery is recorded at the same time. During normal system operation, a health status assessment and window size check are automatically performed after processing every 100 data points. When a sudden change in key parameters such as internal resistance is detected, the assessment process is triggered immediately.

[0046] Window Adjustment and Statistical Maintenance: When a new window size is calculated, the system adjusts the buffer used to store historical data. If the new window is smaller than the current window, the oldest excess data in the buffer is discarded. Subsequently, the system recalculates all statistical characteristics, such as the mean and variance, based on all remaining data in the buffer, ensuring that the statistics are strictly consistent with the new window range and avoiding bias caused by data truncation. After completing the window adjustment and statistical update, the original algorithm's "3σ principle" continues to be used to calculate the dynamic safety threshold using the new mean and variance.

[0047] This achieves dynamic adaptation, with the threshold smoothly adjusting as the battery ages without manual intervention, making it suitable for operation in resource-constrained embedded data acquisition terminals. It consumes minimal memory and computing resources, ensuring system real-time performance and low power consumption.

[0048] In this embodiment, the method for constructing a multimodal fault evolution knowledge graph includes the following steps: 1. Under controlled laboratory conditions, induce multiple known faults in the battery and synchronously collect time series data from the initial stage of the fault to the entire process of thermal runaway; 2. Divide the time series data into stages and label the states according to the characteristic inflection points; 3. Based on the labeled data, mine frequently occurring state association rules through data mining methods to construct a graph representing the evolution law of parameters and the relationship between state transitions.

[0049] Under controlled laboratory conditions, various known faults were induced in the battery, and high-frequency time-series data on temperature, voltage, various gas concentrations, and internal resistance were collected throughout the entire process, from the initial stage of fault occurrence to thermal runaway, using a data acquisition terminal. This process was mainly divided into two stages: data acquisition and preprocessing, and the construction and storage of a knowledge graph.

[0050] In the data acquisition and preprocessing stage (completed in a laboratory environment), a batch of new battery samples of the same model were selected, and various controllable fault abuse experiments were designed to create a known, single fault origin, ensuring that the acquired data has a clear causal relationship with specific fault modes. High-speed data acquisition cards were used to synchronously acquire data from all sensors (voltage, gas concentration, temperature, etc.), and millisecond-level time synchronization was performed on the multi-sensor data acquired in the laboratory. Based on characteristic inflection points, the entire fault process was divided into continuous stages. Generally, battery faults can be divided into: t0-t1: Normal period; t1-t2: Initial stage of thermal abuse (characteristic: slight increase in characteristic gases such as CO); t2-t3: Late stage of thermal abuse (characteristic: accelerated temperature rise); t3-t4: Thermal diffusion period (characteristic: sudden voltage drop); t4-t∞: Thermal runaway period (characteristic: temperature > 800°C, smoke emission). The knowledge graph is constructed during the storage stage, utilizing the labeled data produced in the previous stage to construct a computable knowledge graph through data mining techniques. For each labeled fault test segment, the statistical characteristics of each parameter in each stage are calculated, and a complete fault test data is transformed into a state sequence: {S1: Normal, S2: Initial stage of thermal abuse, S3: Late stage of thermal abuse, S4: Thermal diffusion stage, S5: Thermal runaway stage}.

[0051] Each failure stage is defined as a node. Node attributes include the typical parameter range and main characterizing parameters of that stage, constructing the skeleton of the graph. Frequently occurring association rules are mined from all state sequences. Based on support and confidence metrics, strong causal and temporal relationships are automatically discovered from the data. A failure sequence database D = {I1,I2,…I...} is constructed through data acquisition and preprocessing stages. x}, and set a minimum support level Sup min and minimum confidence Conf min.

[0052] Scan the database, generate candidate 1-itemsets C1, and calculate the weighted support Sup: Sup x = sequence {I x Pruning the sum of weights of all states in the set / the total weight of all sequences determines the frequent 1-itemset L1: if Sup x >= Sup min , will I x Add to L1 and repeat the above steps to find frequent 2-itemsets, 3-itemsets... until L1 k For an empty set: Sup x1,x2 = sequence {I x1 ,I x2 The sum of the weights of all states in} / the total weight of all sequences, L2 = {(I x1 I x2 Sup x1,x2,x3 =sequence{I x1 ,I x2 ,I x3 The sum of the weights of all states in} / the total weight of all sequences, L3 = {(I x1 I x2 I x3 ),...} by L k-1 Generate candidate k-itemsets C k When performing a self-join, the items of the two itemsets are arranged in a certain order, and the two itemsets can be joined if their first k-2 items are the same: L （1） L （2） All are (k-1)-itemsets; C k = L （1） + L （2） ={L （1） [1], L （1） [2],…, L （1） [k-1],L （2） [k-1]};Any subset of a frequent k-itemset must be a frequent itemset, defined by the set C generated during the join. k It is necessary to verify and exclude infrequent k-itemsets that do not meet the support requirement. After filtering, frequent k-itemsets L are determined. k .

[0053] For each frequent itemset L, generate all its non-empty proper subsets S. For each subset S, construct the rule S -> (l - S) to generate all potential candidate rules, and calculate the confidence Conf of each rule: Conf = Sup (L) / Sup (s).

[0054] If Conf x >= Confmin If so, then rule X is retained and determined as a strongly associated rule.

[0055] Finally, the nodes, edges, and their attributes are stored in a graph database to form a queryable and reasonable multimodal fault evolution knowledge graph.

[0056] Based on the various processes in the above stages, real-time, multi-dimensional time-series observation data of the battery are used to achieve more refined state tracking and risk prediction through a model that can model continuous hidden states and their correlations.

[0057] The hidden state set of the Hidden Markov Model (HMM) is defined as S = {S1: normal, S2: initial stage of thermal abuse, S3: later stage of thermal abuse, S4: thermal diffusion stage, S5: thermal runaway stage}. The initialization parameters of the observed state set O are defined as λ = (A, B, π), π = [1, 0, 0, 0, 0]^T, A is the state transition matrix, B is the observation probability matrix, and λ is trained unsupervised using labeled battery fault sequence data. The observed states are directly read by sensors (including parameters such as carbon monoxide, methane, hydrogen sulfide gas concentration, temperature, and voltage). Since the data is continuous, a continuous observation HMM is usually used, and it is assumed that the observation probability follows a Gaussian mixture model. The method for calculating the state prediction confidence includes the following steps: real-time observation sequences O1, O2, ..., O t Input a Hidden Markov Model and use the forward algorithm to calculate the probability P(Sk) that the battery will be in a thermal runaway state at the k-th time step in the future. t+k = S5 | O1, O2, ..., O t ), where t represents time. The unsupervised training process is as follows: Define the forward probability a t (i): At time t, O1→O is observed. t And at this time the battery is in state S i The probability of the backward direction; the probability β t (i): At time t, the battery is in state S i Under the premise that O can be observed in the future t+1 →O t The probability of a state; the state probability γ t (i): Given the entire observation sequence, at time t, the battery is in state S. i The probability. This is the core output of the E-step; the transition probability ξ. t (i,j): Given the entire observation sequence, the state of the battery at time t is S. i And at time t+1, it transitions to state S. j The probability of.

[0058] Based on the previously measured data, the model parameter λ 0 = (A0 B 0 , π 0 Assign an initial estimate and begin iteration.

[0059] (1) Calculate the expected statistic for each observation sequence O in the training set, using the current model parameters λ. n Calculate a t (i) β t (i): Based on the data collected in Phase 2, based on the state S at time t i The support and the support at time t when state Si ends are calculated as a. t (i), β t (i).

[0060] Calculate the state probability γ using the current model parameters. t (i): γ t (i) = P(S t = S i | O, λ n ) = (α t (i) * β t (i)) / P(O | λ n Calculate the transition probability ξ using the current model parameters. t (i,j): ξ t (i,j) = P(S t =S i , S t+1 =S j | O, λ n ) = (α t (i) * A ij * B j (O t+1 ) * β t+1 (j)) / P(O | λ n (2) Update the model parameters using all γ calculated in the expectation step. t (i) and ξ t (i,j) (sum over all training sequences), re-estimate the model parameters to obtain the next generation parameters λ^{n+1}: update the initial state distribution π, state transition matrix A, and observation probability distribution B: π i ^{n+1} = γ1(i)A ij ^{n+1} = (Sum of the expected number of transitions from i to j at all times) / (Sum of the expected number of transitions out of i at all times) with mean μ jmUpdate: The average of all observations belonging to state j and the m-th Gaussian component, weighted by their respective probabilities; covariance Σ jm Update: The covariance of the above observations is weighted by their respective probabilities; weight c jm Update: The proportion of data points belonging to state j and the m-th Gaussian component out of the total; update the parameter λ. n+1 As input for the next iteration, repeat the expectation step and the maximization step until the model's total likelihood P(O | λ) no longer increases significantly, or the parameter change is less than a preset threshold. At this point, the model is considered to have converged and training is complete.

[0061] The real-time multidimensional data streams (temperature, gas, voltage, etc.) acquired from the sensors are arranged into an observation sequence O1, O2, ..., O according to the training time step. t Input the observed sequence into the trained HMM model, and calculate the hidden state sequence Q1, Q2, ..., Q that is most likely to have generated the observed sequence. t .

[0062] Using the forward algorithm, calculate the probability P(S) of being in a dangerous state (e.g., S5: thermal runaway) after k time steps. t+k = S5 | O1, O2, ..., O t If P(S) t+k = S5)>Preset threshold, then trigger early warning: (1) Using the HMM model after training, the initial state probability distribution π:[1.0, 0.0, 0.0, …], calculate the forward probability vector α at the current time. t (i)=P( O1, O2,…, O t ,S t = S i |λ), define the preset threshold P0.

[0063] Initialize t=1: α1(i) = π i * B i (O1); Recursively calculate t=2, 3, …,t: α t (j) = [Σ i (α t-1 (i) * A ij )] * B j (O t ); where α t-1 (i): The probability that the battery was in state i at the previous moment, A ij B represents the probability that the battery transitions from state i in the previous time step to state j in the current time step. j (O t): Observation probability.

[0064] (2) Predict the probability distribution of the state in the next k steps: P(S t+k | O1, O2, ..., Ot) = α t * A^k where α t Let A be a 1xN row vector (where N is the number of states), representing the probability distribution of the current state, and let A be an NxN state transition matrix. A^k denotes matrix A multiplied by itself k times, representing the state transition probability after k steps.

[0065] Ultimately, if P(S) t+k If Si)>P0, an early warning is triggered, and the warning signal is sent to the computer terminal.

[0066] The formula for calculating the confidence level of threshold exceedance is: C3 = Amplitude Factor * Duration Factor; where C3 is the confidence level of threshold exceedance, the amplitude factor is obtained by mapping the relative amplitude of the current parameter value exceeding the adaptive safety threshold through an S-shaped function, and the duration factor is obtained by mapping the duration of the parameter continuously exceeding the adaptive safety threshold through an exponential decay function. The calculation process of the confidence level in this embodiment is described below.

[0067] (1) A time series curve X = (x1, x2,..., x) of a historical fault parameter (such as temperature) extracted from the knowledge base based on the confidence level C1 of historical fault mode matching. n The currently collected, evolving time series curve Y = (y1, y2, ..., y) is the same parameter. m Using these two sets of data, initialize an (m x n) matrix D and begin constructing the cumulative cost matrix: calculate the local cost of each cell in the matrix: d(i, j) = |Y[i] - X[j]|; calculate another matrix C, where each cell C[i, j] represents the minimum cumulative cost among all possible paths from the starting point to (i, j): C[i, j] = d(i, j) + min( C[i-1, j], C[i, j-1], C[i-1, j-1]). The final cumulative cost matrix C is obtained, and the bottom right cell of matrix C is the DTW value: Dtw = C[m,n]. The function C1 = ... Dtw is converted to a confidence level C1 in the range [0, 1].

[0068] (2) The confidence level C2 based on future state prediction is directly calculated using the HMM model, which is the predicted probability that the battery will be in the most dangerous state after k steps. The calculation formula is: C2 = P(S t+k = Si |O1, O2,…, O t (3) Based on the confidence level C3 of the current parameter exceeding the standard, assess the degree of deviation and duration of the current real-time parameter value relative to the previously calculated dynamic threshold. The greater the deviation and the longer the duration, the higher the confidence level.

[0069] Define the magnitude factor as Magnitude. The measured value of the current parameter x, the current dynamic threshold T(μ + 3σ) calculated in stage S1, and the relative magnitude exceeding the threshold are also considered. ,therefore: Where k is the steepness factor, which controls the shape of the function. The larger k is, the faster the value will approach 1 if there is a small overshoot.

[0070] Define the duration factor as Duration, and the parameter as the length of time t during which the parameter continuously exceeds a threshold. Then: Where τ is the time constant. The smaller τ is, the faster the confidence increases over time, then: C3 = Magnitude * Duration Step 3: Make decisions based on multi-source confidence fusion: Perform weighted fusion of pattern matching confidence, state prediction confidence and threshold exceedance confidence to calculate the comprehensive risk confidence. Define the weight vector Ω = [ω1, ω2, ω3]^T, and satisfy ω1 + ω2 + ω3 = 1. Static weight preset: (1) Early stage of hot abuse: The parameters have not exceeded the standard and the spectral features are not obvious, but the HMM may have detected a weak trend. This stage focuses on sensitivity. The HMM is most sensitive to small trend changes, so it is given the highest weight to try to capture the earliest signs of abnormality, where ω2 > ω1 > ω3.

[0071] (2) Late stage of heat abuse and heat diffusion stage: Parameters begin to fluctuate and may slightly exceed the limit. The matching degree of the spectrum increases and the prediction probability of HMM increases. Accuracy is the key at this stage. It is necessary to cross-validate by combining trend prediction and pattern matching to avoid false alarms from a single algorithm. The threshold weight is still low because a single point exceeding the limit may only be a misjudgment, where ω1 ≈ ω2 > ω3.

[0072] (3) Thermal runaway stage: Parameters continue to exceed limits, and the spectral matching characteristics are obvious. At this stage, the fault mode is very clear, and the spectral matching is the most reliable. The prediction task of HMM has been completed, and the weights are reduced. The evidence weight of exceeding the threshold increases because it provides a direct and strong danger signal, ω1 > ω3 > ω2.

[0073] In summary, in this embodiment, the weighting coefficients are dynamically adjusted according to different stages of battery fault development: during the fault latency period (early stage of thermal abuse), ω2>ω1>ω3 is set to enhance sensitivity to subtle trends; during the fault development period (late stage of thermal abuse and thermal diffusion period), ω1=ω2>ω3 is set to perform cross-validation by combining trend prediction and pattern matching; and during the late stage of the fault (thermal runaway period), ω1>ω3>ω2 is set to rely on clear fault modes and direct evidence of exceeding limits.

[0074] The weight vector Ω = [ω1, ω2, ω3]^T needs to be selected based on the actual application, or it can be dynamically optimized and adjusted based on the accuracy of historical alarms using a reinforcement learning algorithm. The formula for calculating the comprehensive risk confidence level is: C final = ω1 * C1 + ω2 * C2 + ω3 * C3 where, C final To represent the overall risk confidence level, C1, C2, and C3 are the pattern matching confidence level, state prediction confidence level, and threshold exceedance confidence level, respectively, and ω1, ω2, and ω3 are the corresponding weight coefficients, satisfying ω1+ω2+ω3=1.

[0075] Step 4: Tiered Early Warning and Response: Based on the preset numerical range of the comprehensive risk confidence level, trigger the corresponding level of early warning signal and emergency response measures.

[0076] In this embodiment, the preset numerical range includes at least four levels, each corresponding to a different response measure: when 0.1 <C final When the value is ≤ 0.3, it is defined as Level 1 monitoring level, and only exception logs are recorded; when it is 0.3... <C final When the value is ≤ 0.5, it is defined as Level 2 alert level, and a warning notification is sent to the user terminal; when it is 0.5... <C final When C is ≤ 0.8, it is defined as Level 3 warning level, triggering the local audible and visual alarm on the data acquisition terminal and sending an emergency alarm signal to the cloud; when C final When the value is greater than 0.8, it is defined as Level 4 hazard level, and the highest level emergency response is implemented, which includes linking the fire protection system and emergency circuit disconnection.

[0077] In summary, compared with existing monitoring and early warning methods for electrochemical energy storage power stations, the real-time monitoring and early warning method for electrochemical energy storage power stations in this embodiment has the following advantages: 1. This real-time monitoring and early warning method for electrochemical energy storage power stations can monitor the power station batteries in real time, collect and store multi-timescale sensor data related to the safety of batteries and key circuits in the electrochemical energy storage power station in real time, and can promptly capture the changing trends of abnormal data, effectively improving the safety of the energy storage power station and ensuring its stable operation. Through four innovative modules—dynamic threshold, multimodal knowledge graph, hidden Markov model state prediction, and multi-source fusion decision-making—the accuracy, robustness, and adaptability of early warning for batteries are significantly improved. It is especially suitable for battery health management in electrochemical energy storage power stations with multiple models and operating conditions, solving the technical problems of incomplete data collection, poor timeliness, and incomplete data processing in existing monitoring and early warning methods for electrochemical energy storage power stations.

[0078] 2. This real-time monitoring and early warning method for electrochemical energy storage power stations achieves dynamic self-adaptation. Thresholds are smoothly adjusted as the battery ages without manual intervention, making it suitable for operation in resource-constrained embedded data acquisition terminals. It consumes minimal memory and computing resources, ensuring system real-time performance and low power consumption. This method draws on the ideas of information fusion and ensemble learning, greatly improving the system's robustness by integrating multiple independent and complementary information sources, enabling more reliable and accurate final decisions than any single information source.

[0079] Example 2 (Refer to Figures 2 and 3) provides a real-time monitoring and early warning device for an electrochemical energy storage power station. This device applies the real-time monitoring and early warning method for electrochemical energy storage power stations (i.e., the battery early warning algorithm) from Example 1. The real-time monitoring and early warning device includes a data acquisition terminal, a cloud server, and a user terminal, and may also include a power module.

[0080] The data acquisition terminal includes a main control center 9, a wireless communication module, a display module, an audible and visual alarm module 8, and a sensor array (sensor data acquisition module) for collecting temperature, characteristic gas concentration, voltage / current, and internal resistance. The data acquisition terminal collects and uploads multi-parameter time-series data and executes local early warnings. Specifically, the data acquisition terminal collects charging and discharging hazard factors and risk characteristic parameters, processes these parameters centrally through a cloud server, and provides users with scientific predictions after calculation and judgment by a battery early warning algorithm. The data acquisition terminal is equipped with a communication interface for acquiring monitoring parameter information and sending the parameter information to the cloud server for analysis and calculation.

[0081] In this embodiment, the sensor array includes a DHT11 temperature and humidity sensor 1, a TSP-BAT24-120 voltage / current sensor 2, an MQ136 hydrogen sulfide concentration sensor 3, an MQ7 CO concentration sensor 4, and an MQ4 methane concentration sensor 5. These sensors are arranged around the battery under test, forming a comprehensive parameter monitoring network. The sensor array can use a PCF8591+MQ7, MQ4, and MQ136 converter to monitor the concentrations of carbon monoxide, methane, and hydrogen sulfide gases, respectively. The data acquisition terminal adopts a cylindrical structure design, which facilitates installation and arrangement around the battery.

[0082] The display module is used to display the collected parameter information. In this embodiment, the display module is an LED intelligent display module 6. The device is equipped with an OLED intelligent display module 6 on the side, which displays the data collected in real time by the sensor array, making it easy for on-site personnel to intuitively understand the monitoring status. After the system is running, the OLED displays the DHT11 temperature and humidity, MQ136 hydrogen sulfide concentration, MQ4 methane concentration, MQ7 CO concentration, and TSP-BAT24-120 voltage and current values. The threshold setting interface can be accessed by pressing the K3 key. Initially, it is the lower limit of the temperature. Cyclicly pressing the K3 key can switch the upper and lower limit threshold positions of the corresponding sensor. K1 and K2 are used to adjust the values. After setting, press the K4 key to save and return to the main interface.

[0083] The wireless communication module is used to upload the collected parameter information to the cloud server. This communication module uses the NB-IoT module 7, specifically the BC260Y-CN module. This module has a PSM pin for low-power wake-up, supports the MQTT protocol and multiple cloud platforms, and sends AT commands via UART for control. The NB-IoT module 7 has three power-saving modes: Power Saving Mode (PSM), Discontinuous Reception Mode (DRX), and Extended Discontinuous Reception (eDRX) mode. Under normal conditions, PSM or e-DRX is used for low-power monitoring; when an anomaly is triggered, DRX mode is activated through sensors such as acceleration and temperature surge sensors to transmit data in real time and receive control commands. The data acquisition terminal has a communication interface for acquiring monitoring parameter information and sending this information to the cloud server.

[0084] The audible and visual alarm module 8 is used to detect whether parameter information exceeds a set threshold. When the parameter is at a dangerous value, it triggers an audible and visual alarm and transmits the alarm information to the main control center 9. Specifically, upon startup, the audible and visual alarm module 8 initializes the sensor and NB-IoT communication module, and initializes the alarm thresholds (e.g., voltage: 58V, temperature: >60°C, CO, CH4, H2S: volume fraction >0.05%). Different thresholds are set via the main control center 9 according to the application scenario. When the collected data exceeds the set threshold, an audible and visual warning is triggered. If a reset operation is not performed within half a minute, the main control center 9 sends an alarm signal via the communication module. When the monitored parameter information is at a dangerous value, an audible and visual alarm is triggered, and the stepper motor reverses. The sensor array wakes up from sleep mode, increases the data acquisition frequency, activates the audible and visual alarm, and sends an alarm signal.

[0085] The main control center 9 is used to summarize the parameter information collected by the sensor array, upload it via the wireless communication module, and adjust the working mode of the data acquisition terminal when an alarm is received. After the system is running, the OLED displays the DHT11 temperature and humidity, MQ136 hydrogen sulfide concentration, MQ4 methane concentration, and MQ7 CO concentration values. The threshold setting interface can be accessed by pressing the K3 key. Initially, it is the lower temperature limit. Cyclicly pressing the K3 key can switch the upper and lower threshold positions of the corresponding sensors. K1 and K2 are used to adjust the values. After setting, press the K4 key to save and return to the main interface.

[0086] When the detected temperature, characteristic gas, voltage, or current exceeds the upper limit of the set threshold (e.g., temperature, humidity, smoke, methane, CO exceeding the upper limit of the set threshold), an audible and visual alarm is triggered, and the stepper motor rotates forward. When the detected temperature, characteristic gas, voltage, or current falls below the lower limit of the set threshold (e.g., temperature, humidity, smoke, methane, CO falling below the lower limit of the set threshold), an audible and visual alarm is triggered, and the stepper motor rotates in reverse. When the monitored parameter information is at a dangerous value, an audible and visual alarm is triggered, and the stepper motor rotates in reverse; the sensor array wakes up from sleep mode, increases the data acquisition frequency, activates the audible and visual alarm, and sends an alarm signal. When the monitored parameter information is within the threshold range, the audible and visual alarm is canceled, the sensor array operates at low power and enters sleep mode, the NB-IoT module 7 switches to PSM or e-DRX mode, and the RTC periodically wakes up the system for detection.

[0087] The cloud server is used to receive and store data uploaded by the data acquisition terminal, and to perform fusion analysis and state prediction on time series data, executing fusion analysis, state prediction, and hierarchical early warning decisions. Here, the cloud server is used to analyze the battery state. The basic principle of the model is as follows: the signal acquisition module collects and statistically analyzes three sets of data in real time: the internal temperature of the lithium iron phosphate battery under normal operating conditions, different gas concentrations, and battery voltage and current obtained from the BMS, establishing reliable sample values ​​for each of the three sets of data under normal operating conditions of the lithium battery.

[0088] The user terminal receives and displays battery status and warning information from the cloud server. The power module 10 provides voltage to the data acquisition terminal. This module uses a 7805 voltage regulator chip, an input filter capacitor, and an output filter capacitor, connected to the chip's input and output terminals respectively, to power the data acquisition terminal and ensure stable system operation.

[0089] The following embodiment will conduct simulation and result analysis. The original simulation data is generated based on the thermal runaway characteristic parameter evolution model in the published literature. Based on the statistical 3σ principle, the normal and abnormal parts of the data are selected as the original simulation data, totaling 7200 data points. The total experiment duration is 2 hours. The first 30 minutes are the normal operation state, and after 30 minutes, the thermal runaway fault is simulated. Continuous sampling at 1Hz is used to ensure data integrity. The collected data includes CO concentration, H2S concentration, CH4 concentration, temperature, battery internal resistance, voltage and current, etc. The table below shows the predicted values ​​of each monitoring data of the device.

[0090] Table 1. Numerical Table of Monitoring Data First, based on the simulation objectives, a continuous time series data segment with a total duration of T (set to 2 hours) is extracted and constructed from the basic data source. This sequence is clearly divided into two stages: the first 30 minutes represent the healthy operating state of the battery, and the data in this stage is used to initialize the dynamic adaptive threshold learning model of the S1 stage of the algorithm, establishing the baseline (μ, σ) of each parameter and the initial safety threshold; after 30 minutes, it represents the process of fault induction and development to thermal runaway. The data in this stage simulates the complete process of the battery evolving from an initial fault such as an internal short circuit or overheating to thermal abuse, massive gas release, voltage drop, and finally thermal runaway.

[0091] Before inputting data into the algorithm, the following standardized preprocessing is performed: time alignment and resampling are conducted to ensure that all sensor data channels have strictly synchronized timestamps, and resampling is performed according to the basic time step required by the algorithm to form a uniform time series. Parameters of different dimensions are normalized to a similar numerical range to facilitate subsequent model calculations, simulate the output of real data acquisition terminals, and improve the robustness of the algorithm.

[0092] The simulation uses the aforementioned time-series data as input, and advances the simulation time step by step using a time-series sliding window to simulate the real-time online operation of the algorithm. The system calculates the initial mean and standard deviation of each target parameter, such as temperature, gas concentration, and internal resistance. Based on the "3σ principle," an initial dynamic safety threshold is set for each parameter.

[0093] Simultaneously, the sliding window and its update mechanism are initialized, and the initial internal resistance value is recorded for subsequent SOH estimation (characteristic gas data are collected as shown in Figures 4, 5, and 6). New data points are included in the sliding window, and the mean and variance of the data within the window are updated, thereby updating the dynamic safety threshold. Simultaneously, the current SOH is estimated based on the data within the window, and the sliding window size is dynamically adjusted accordingly to achieve adaptive behavior. A fixed-length observation subsequence is traced back from the current moment. This subsequence is compared with typical fault modes in a "multimodal fault evolution knowledge graph" pre-stored in the cloud and constructed from laboratory data. The morphological similarity between the current observation sequence and each historical fault sequence is calculated and converted into a pattern matching confidence score C1. The same observation subsequence is input into the pre-trained HMM model. The model uses a forward algorithm to calculate the most likely hidden state of the battery at the current moment and further predicts the probability that the battery will be in a high-risk state in the next 30 seconds, as the state prediction confidence score C2. It is determined in real time whether the current parameter values ​​exceed their respective latest dynamic safety thresholds. The threshold exceedance confidence score C3 is calculated by comprehensively evaluating the magnitude and duration of the exceedance.

[0094] At each time step, the algorithm receives the confidence output from the previous step and calculates the final fusion confidence C. final = ω1 * C1 + ω2 * C2 + ω3 * C3, and the warning threshold [0.1, 0.3, 0.5, 0.8] is the level four warning threshold.

[0095] Simulation results: Total experiment duration: 2 hours; total data points: 7200; fault occurrence time: 30 minutes; highest warning level: 4; warning trigger time: approximately 40 minutes; system response time: 10.3 minutes; prediction accuracy: 94.1%. The results demonstrate that the method used in this embodiment achieves good prediction performance.

[0096] In summary, compared with existing monitoring and early warning devices for electrochemical energy storage power stations, the real-time monitoring and early warning device for electrochemical energy storage power stations in this embodiment has the following advantages: 1. This real-time monitoring and early warning device for electrochemical energy storage power stations operates simultaneously through multiple data sensors, coordinating with a cloud server to process data, reducing the time the processor core waits for data and improving system response speed. By intelligently controlling two working modes through the main control center 9—a low-power consumption mode when the battery is working normally, and a high-frequency acquisition mode when abnormal—power consumption and labor costs are reduced. This device can accurately determine the working status of the battery in the electrochemical energy storage power station, greatly avoiding the occurrence of safety accidents and providing strong technological protection for the safe operation of the energy storage power station.

[0097] 2. This real-time monitoring and early warning device for electrochemical energy storage power stations utilizes sensor data fusion technology to collect and store multi-timescale sensor data related to the safety of batteries and critical circuits in the electrochemical energy storage power station in real time. Regarding the data acquisition frequency, it is dynamically adjusted according to the operating status of the energy storage power station. During normal operation, a lower acquisition frequency is used to reduce power consumption and data transmission pressure; when an anomaly is detected, the acquisition frequency is automatically increased to ensure timely capture of abnormal data trends. This dynamic adjustment strategy ensures both data real-time performance and improves the overall efficiency of the system.

[0098] Example 3 This example provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the real-time monitoring and early warning method for electrochemical energy storage power stations described in Example 1.

[0099] The method in Example 1 can be applied in software form, such as by designing it as a standalone program and installing it on a computer terminal, which can be a computer, smartphone, control system, or other IoT device. Alternatively, the method in Example 1 can be designed as an embedded program and installed on a computer terminal, such as on a microcontroller.

[0100] Example 4 This example provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements the steps of the real-time monitoring and early warning method for an electrochemical energy storage power station described in Example 1.

[0101] When applying the method of Example 1, it can be applied in the form of software, such as by designing it as a program that can run independently on a computer-readable storage medium. The computer-readable storage medium can be a USB flash drive, designed as a USB security token, and the program can be designed to start the entire method through an external trigger.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements 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 real-time monitoring and early warning of electrochemical energy storage power stations, characterized in that, It includes the following steps: Real-time acquisition of battery temperature, concentration of various characteristic gases, voltage, current and internal resistance parameters to obtain multi-parameter time series data; The time series data is fused, analyzed, and its status is predicted: based on historical operating data under battery health conditions, adaptive safety thresholds for each parameter are established and updated online. The real-time observation sequence is matched with a pre-constructed multimodal fault evolution knowledge graph to obtain the pattern matching confidence; the real-time observation sequence is input into a pre-trained hidden Markov model to predict the probability that the battery will be in a dangerous state in the future, and this probability is used as the state prediction confidence. Calculate the magnitude and duration of the current parameter value exceeding the safety threshold to obtain the threshold exceedance confidence level; The pattern matching confidence, the state prediction confidence, and the threshold exceeding confidence are weighted and fused to calculate the comprehensive risk confidence. Based on the preset numerical range of the comprehensive risk confidence level, the corresponding level of early warning signal and emergency response measures will be triggered.

2. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 1, characterized in that, The adaptive safety threshold update method includes the following steps: based on the statistical 3σ principle, calculate the initial mean and standard deviation of each parameter data sampled over a long period of time under battery health conditions, and set an initial safety threshold; use an online statistical algorithm to dynamically update the mean and variance of the parameters using the latest data within the sliding window, and update the adaptive safety threshold in real time; introduce the battery health status value as feedback, and dynamically adjust the size of the sliding window according to the battery health status value to adapt to the degradation of battery performance throughout its entire life cycle.

3. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 1, characterized in that, The method for constructing the multimodal fault evolution knowledge graph includes the following steps: under controlled laboratory conditions, inducing multiple known faults in the battery and simultaneously collecting time series data from the initial stage of the fault to the entire process of thermal runaway; dividing the time series data into stages and labeling states based on characteristic inflection points; and based on the labeled data, mining frequently occurring state association rules through data mining methods to construct a graph representing the evolution law of parameters and the relationship between state transitions.

4. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 3, characterized in that, The hidden state set of the Hidden Markov Model is defined as S = { S1: normal, S2: initial stage of thermal abuse, S3: later stage of thermal abuse, S4: thermal diffusion stage, S5: thermal runaway stage}. The initialization parameters of the observation state set O are defined as λ = (A, B, π), π = [1, 0, 0, 0, 0]^T, A is the state transition matrix, B is the observation probability matrix, and the labeled battery fault sequence data is input to perform unsupervised training on λ. The method for calculating the state prediction confidence includes the following steps: dividing the real-time observation sequence O1, O2, ..., O t Input the Hidden Markov Model and use the forward algorithm to calculate the probability P(S) that the battery will be in a thermal runaway state at the k-th time step in the future. t+k = S5 | O1, O2, ..., O t ), where t represents time.

5. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 1, characterized in that, The formula for calculating the threshold exceedance confidence level is: C3 = amplitude factor * duration factor; where C3 is the threshold exceedance confidence level, the amplitude factor is obtained by mapping the relative amplitude of the current parameter value exceeding the adaptive safety threshold through an S-shaped function, and the duration factor is obtained by mapping the duration of the parameter continuously exceeding the adaptive safety threshold through an exponential decay function.

6. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 1, characterized in that, The formula for calculating the comprehensive risk confidence level is: C final = ω1 * C1 + ω2 * C2 + ω3 * C3; where C final Let C1, C2, and C3 be the pattern matching confidence, the state prediction confidence, and the threshold exceeding confidence, respectively, and let ω1, ω2, and ω3 be the corresponding weight coefficients, satisfying ω1+ω2+ω3=1.

7. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 6, characterized in that, The weighting coefficients are dynamically adjusted according to different stages of battery fault development: during the fault latency period, ω2 > ω1 > ω3 is set to improve sensitivity to weak trends; during the fault development period, ω1 = ω2 > ω3 is set to perform cross-validation by combining trend prediction and pattern matching. In the late stages of a failure, ω1 > ω3 > ω2 is set to rely on a clear failure mode and direct evidence of exceeding the limit.

8. The real-time monitoring and early warning method for electrochemical energy storage power stations as described in claim 6, characterized in that, The preset numerical range includes at least four levels, each corresponding to a different response measure: when 0.1 < C final When ≤ 0.3, it is defined as Level 1 monitoring level, only recording exception logs; when 0.3 < C final When ≤ 0.5, it is defined as Level 2 alert level, and a warning notification is sent to the user terminal; when 0.5 < C final When C is ≤ 0.8, it is defined as Level 3 warning level, triggering the local audible and visual alarm on the data acquisition terminal and sending an emergency alarm signal to the cloud; when C final When the value is >0.8, it is defined as Level 4 hazard level, and the highest level emergency response is executed, which includes linking the fire protection system and emergency circuit cut-off.

9. A real-time monitoring and early warning device for electrochemical energy storage power stations, characterized in that, The application is the real-time monitoring and early warning method for electrochemical energy storage power stations as described in any one of claims 1-8. The real-time monitoring and early warning device includes: a data acquisition terminal, which includes a main control center, a wireless communication module, a display module, an audible and visual alarm module, and a sensor array for collecting temperature, characteristic gas concentration, voltage / current, and internal resistance, and for collecting and uploading multi-parameter time series data and executing local early warning; the display module is used to display the collected parameter information; the audible and visual alarm module is used to detect whether the parameter information exceeds a set threshold, trigger an audible and visual alarm when the parameter is at a dangerous value, and transmit the alarm information to the main control center; the main control center is used to summarize the parameter information collected by the sensor array, upload it through the wireless communication module, and adjust the working mode of the data acquisition terminal when receiving alarm information; a cloud server, which is used to receive and store the data uploaded by the data acquisition terminal, and perform fusion analysis and state prediction on the time series data, and execute fusion analysis, state prediction, and hierarchical early warning decision; and a user terminal, which is used to receive and display battery status information and early warning information from the cloud server.

10. The real-time monitoring and early warning device for electrochemical energy storage power stations as described in claim 9, characterized in that, The sensor array includes a temperature and humidity sensor, a hydrogen sulfide concentration sensor, a methane concentration sensor, a carbon monoxide concentration sensor, and a voltage / current sensor arranged around the circumference of the battery under test; wherein the data acquisition terminal is generally cylindrical in shape.

Images