A real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems

Abstract

The application discloses a kind of real-time monitoring and fault response control device of commercial liquid cooling energy storage system, specifically relates to liquid cooling energy storage fault management technical field, by collecting the temperature, pressure, vibration and flow data of key components in liquid cooling system, and after pretreatment;Using time series analysis and signal processing technology, the characteristic vector reflecting the state of key components is extracted from multidimensional operation data set, the significance of weak abnormal signal is enhanced, and the key feature set containing weak abnormal characteristics is generated;An abnormality identification model is constructed using a long short-term memory network, the key feature set is input, the abnormal probability corresponding to the key component is output, and the warning threshold of the abnormal probability is set to trigger the warning;Based on the Weibull reliability model, the fault maintenance time window of the key component is predicted, and based on the fault maintenance time window prediction result and the abnormal probability, the priority of maintaining the key component is dynamically adjusted and preventive maintenance is performed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of liquid-cooled energy storage fault management, more specifically, the present application relates to a real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems. BACKGROUND

[0002] Industrial and commercial liquid-cooled energy storage systems, as an important part of modern energy management, are widely used in power peak shaving, renewable energy grid connection and industrial load balancing fields. The system controls the temperature of battery units and key components through liquid cooling technology to ensure efficient operation and long-term stability. Key components such as battery units, cooling pipes and pumps and valves operate under complex conditions and need to withstand multiple stresses such as high temperature, high pressure and vibration. Their state changes directly affect the safety and reliability of the system. To ensure system operation, real-time monitoring of temperature, pressure, vibration and flow and other multi-dimensional operating data is needed, and potential faults are predicted based on monitoring results to implement preventive maintenance. However, due to the dynamic nature of the operating environment and the diversity of fault modes, traditional monitoring and response mechanisms have significant limitations in dealing with weak abnormal signals and complex fault evolution.

[0003] In the prior art, the liquid-cooled energy storage system has the problem of inaccurate maintenance timing due to complex and variable working conditions. Specifically, traditional methods usually rely on static thresholds or a single data source to judge abnormalities, making it difficult to accurately capture early degradation of battery units or small fluctuations in cooling fluid flow, resulting in delayed abnormal detection; At the same time, the fault prediction model lacks dynamic adaptability to real-time operating data, resulting in large prediction bias of remaining useful life and difficulty in matching the actual fault process with the maintenance time window. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems, which integrates multi-source data acquisition and feature engineering technology to monitor the state of key components in real time, uses a long short-term memory network to accurately identify weak abnormal signals and output abnormal probability, combines a Weibull reliability model to dynamically predict the remaining useful life and generate a fault maintenance time window, and optimizes the maintenance priority based on the abnormal probability and time window prediction results, to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems, comprising:

[0006] Multi-source data acquisition and preprocessing module: by deploying sensors, real-time acquisition of temperature, pressure, vibration and flow data of key components in the liquid-cooled system, and preprocessing of the collected data to generate a denoised and standardized multi-dimensional operating data set;

[0007] Feature engineering and weak signal extraction module: time series analysis and signal processing technology are used to extract feature vectors reflecting the state of key components from multi-dimensional operation data set, enhance the significance of weak abnormal signals, and generate key feature set containing weak abnormal features;

[0008] Abnormal pattern recognition module: based on the key feature set, an abnormal recognition model is constructed using a long short-term memory network, a training set is constructed through historical operation data, and a trained abnormal recognition model is obtained; the abnormal probability corresponding to the key component is output, and the warning threshold of the abnormal probability is set to trigger the warning;

[0009] Fault warning and optimization module: based on the Weibull reliability model, the remaining useful life of the key component is predicted, the fault maintenance time window is generated combined with the confidence interval, and based on the prediction result of the fault maintenance time window and the abnormal probability, the priority of maintaining the key component is dynamically adjusted and preventive maintenance is performed; the operation data of the key component after maintenance is used to update the abnormal recognition model and the Weibull reliability model, forming a closed-loop optimization control.

[0010] Preferably, the fault maintenance time window is obtained in the following manner:

[0011] Data acquisition: acquire the average abnormal probability of the key component and the historical fault data;

[0012] Model initialization: the Weibull reliability model parameters are initialized, which are suitable for describing the failure law of mechanical components, and the parameters include shape parameter β representing failure mode and scale parameter η representing life; the Weibul model is fitted according to the historical data, and the model parameters are adjusted combined with the current abnormal probability; the parameters are optimized by maximum likelihood estimation or least squares method to ensure that the Weibull reliability model reflects the current state;

[0013] Life prediction: based on the adjusted Weibull reliability model and abnormal probability, the remaining useful life of the component is calculated by the following formula, and the fault maintenance time window is generated;

[0014]

[0015] Where t is the time, F(t) is the abnormal probability, t is the remaining useful life, and t is obtained by backstepping t according to the abnormal probability to obtain the remaining useful life;

[0016] Set the confidence interval of the remaining useful life, calculate the upper and lower limits of the remaining useful life by the inverse function of the Weibull reliability model, generate the fault maintenance time window, and verify that the remaining useful life is consistent with the trend of the abnormal probability.

[0017] Preferably, the training process of the anomaly identification model comprises the following steps:

[0018] Step S1, data set preparation and division: prepare the historical data set of the key components, extract the key feature set, and label the fault time point, and divide it into a training set, a validation set and a test set;

[0019] Step S2, time series data preprocessing and input construction: time series processing is performed on the historical data set to generate an input format suitable for long short-term memory network; based on the key feature set, a sliding window method is used to construct time series input to generate sequence data containing historical trends; the sequence data is normalized to map the feature values to the [0, 1] interval; the standardized time series input sequence is output to capture the cumulative effect of weak signals;

[0020] Step S3, model initialization and training: set the long short-term memory network structure, including the input layer, the hidden layer and the output layer, and output the anomaly probability; use mean square error as the loss function; train the short-term memory network through the Adam optimizer and the loss function to obtain the trained anomaly identification model;

[0021] Step S4, adjust the hyperparameters of the anomaly identification model through the validation set to reduce the loss function value; use the cross-validation method to evaluate the generalization ability of the short-term memory network to ensure the accuracy of the anomaly identification; determine the early warning threshold of the anomaly probability according to the verification result.

[0022] Preferably, the device further comprises a pre-warning threshold dynamic management module, which solves the problem of fixed threshold failure of the liquid cooling system caused by working condition changes through pre-warning threshold dynamic management, comprising:

[0023] Based on the multi-dimensional running data set, the multi-dimensional running data set of the key components is collected in real time to generate a time series data set containing multiple parameters;

[0024] After normalizing each parameter, the fluctuation characteristics of each parameter are calculated, the time-frequency analysis method based on wavelet transform is used to decompose the signal, the high-frequency fluctuation component and the low-frequency trend feature are extracted, a multi-dimensional fluctuation feature vector v is generated, and a dynamic threshold Y is constructed DT Formula:

[0025] Y DT = α·EWMA(v) + β·MAD(v) × (1 + γ·CV(v)) wherein, EWMA(v) represents a smooth trend index, which is the weighted average of the fluctuation trend of each parameter; MAD(v) represents the dispersion; CV(v) adjusts the fluctuation intensity, which is the ratio of the standard deviation of each parameter to the average of each parameter, and α, β, γ are adjustment coefficients.

[0026] Preferably, the feature engineering and weak signal extraction module comprises:

[0027] Wavelet transform is used to decompose the signal in the multidimensional running dataset to separate high-frequency noise and low-frequency trend components;

[0028] Based on empirical mode decomposition, the intrinsic mode functions of weak anomalous signals are extracted. The instantaneous frequency and instantaneous amplitude of each intrinsic mode function are calculated by Hilbert transform, and a multidimensional feature vector reflecting the anomalous trend is generated.

[0029] Principal component analysis is used to reduce the dimensionality of multidimensional feature vectors, forming a key feature set.

[0030] Preferably, the training process of the anomaly detection model further includes:

[0031] An attention mechanism is introduced into the Long Short-Term Memory Network. Based on the characteristics of weak anomalous signals, weights are assigned to each time step to highlight the impact of weak anomalous signals on the probability of anomalousness.

[0032] A bidirectional LSTM structure is set up to capture both historical and future trends, and the number of neurons in the hidden layer is dynamically adjusted.

[0033] Dropout regularization is used to prevent overfitting, combined with an early stopping strategy.

[0034] It should be explained that the attention mechanism is a computational method commonly used in neural networks, designed to enhance the model's ability to focus on important parts of the input sequence. In this invention, the attention mechanism is used in a Long Short-Term Memory (LSTM) network. By calculating the weights of the data at each time step, it highlights the contribution of weak anomalous signals (such as small changes in temperature or vibration) to the prediction of anomalous probability, and generates a context vector by weighted summation of the input features.

[0035] The specific steps for assigning weights to each time step are explained below:

[0036] The input is time series data of key feature sets (such as feature vectors for temperature, pressure, etc.), denoted as X = [x1, x2, ..., x...]. T ], where T is the total number of time steps;

[0037] The LSTM hidden layer outputs the hidden state ht (in vector form) at each time step;

[0038] Calculate attention weight α t Use a single-layer feedforward neural network (e.g., with tanh activation function), with inputs ht and the previous time-step context vector c. t-1 The formula is

[0039] e t =W a ·ht +b a ,

[0040] where, W a and b a are trainable parameters, and a t is the normalized weight; e t represents the attention score of any time step k in the time series, k is the index of the time step, ranging from 1 to T;

[0041] The weighted sum generates the context vector As the input of the anomaly probability prediction, the time step highlighting the weak anomaly signal (such as amplitude mutation) is improved, and the corresponding weight is improved; for example, if the temperature mutation from 25℃ to 28℃ at a certain time step, the calculated a t may increase from 0.05 to 0.15, and the specific weight value is determined by the training data, which ensures that the abnormal signal is amplified.

[0042] Explanation, bidirectional LSTM is an improved LSTM structure, the core is to input the input sequence into two independent LSTM networks at the same time, generate forward hidden state and reverse hidden state respectively, and finally splice into the comprehensive hidden state of each time step, which is used for anomaly probability prediction; In the present application, the complete dependence relationship of the sequence is captured by forward and reverse processing, such as the cause and effect of pump pressure anomaly.

[0043] Explanation, regarding how to capture historical and future trends: bidirectional LSTM captures historical trends from the past to the present (such as the cumulative effect of gradually rising temperature) through forward LSTM, and captures trends from the future to the present (such as abnormal fluctuations before failure) through reverse LSTM, historical trends from the past to the present and trends from the future to the present constitute the dependence relationship in the sequence, so as to more accurately predict the anomaly probability; for example, if the pump vibration suddenly increases at t = 10, the reverse LSTM can trace back to the pressure anomaly influence at t = 11, and the forward LSTM captures the trend before t = 9, which improves the prediction accuracy.

[0044] Explanation, regarding how to dynamically adjust the number of hidden layer neurons, dynamically adjusting the number of hidden layer neurons means adjusting the network capacity of LSTM according to the performance of the verification set, so as to balance the model complexity and the risk of overfitting, and the specific implementation method can be:

[0045] Initially set two hidden layers, and the number of neurons in each layer is 64;

[0046] During training, the loss function (e.g. mean squared error) is evaluated using the validation set. If the loss decreases slowly (less than 0.01 for 5 consecutive iterations), increase the number of neurons in each layer (e.g. to 128). If there are signs of overfitting (e.g. the loss on the validation set increases), reduce the number of neurons (e.g. to 32). Adjustments are made through grid search, with a range of [32, 64, 128], and retrain and validate after each adjustment. For example, if the loss on the validation set stagnates after 50 iterations, try increasing the number of neurons to 128 and observe the change in loss. The adjustment process can be automated through a Python script.

[0047] Explanation: How to prevent overfitting with Dropout regularization: By randomly dropping neurons during training (dropout rate set to 0.2, i.e. 20% of neurons are set to zero), reduce the model's over-reliance on specific features. Implementation:

[0048] Add a Dropout layer after each hidden layer of LSTM;

[0049] During training, randomly drop neurons with a rate of 0.2 and their connections, only 80% of neurons are retained for calculation;

[0050] During testing, all neurons participate in calculation, but the output is multiplied by 0.8 to maintain consistency. This method reduces parameter redundancy and avoids overfitting.

[0051] Explanation: How to combine the early stopping strategy, Early Stopping is a general regularization technique to prevent overtraining, implementation:

[0052] During training, calculate the validation set loss (e.g. mean squared error) after each iteration;

[0053] Set the patience value to 10, i.e. if the validation set loss does not decrease for 10 consecutive times (with a decrease threshold of 0.001), stop training;

[0054] Record the current best model parameters (e.g. weights when the loss is lowest), and use them as the final anomaly identification model; for example, if the loss is 0.05 at the 80th iteration and does not decrease for the next 10 iterations, save the 80th model parameters to ensure the model does not overfit.

[0055] Preferably, based on the anomaly probability output by the anomaly pattern recognition module, adjust the sampling frequency of the sensor.

[0056] Preferably, the early warning threshold dynamic management module further comprises:

[0057] Based on Markov chain, predict the short-term fluctuation trend of multi-dimensional running data set, calculate the state transition probability of each parameter in the future time;

[0058] The dynamic threshold formula is adjusted in combination with the prediction result, and a trend factor δ is introduced, when the fluctuation trend rises, δ < 1, the early warning threshold is lowered, and when it falls, δ > 1, the early warning threshold is raised.

[0059] Preferably, the device further comprises a fault maintenance time window optimization module for obtaining an optimized fault maintenance time window, comprising:

[0060] By real-time monitoring, the abnormal probability of key components in the liquid cooling system is obtained, and combined with historical operation data and current multi-dimensional operation data set, a comprehensive input data set reflecting the state of the components is generated;

[0061] Using a multi-physical field coupling model, based on the comprehensive input data set and the abnormal probability, the thermal, mechanical and fluid mechanical behavior of the key components under the current abnormal probability is simulated;

[0062] Through finite element analysis, a stress-strain distribution map is generated, combined with the abnormal probability to predict the fatigue crack propagation time, a virtual life curve of the component state changing with time is generated, which is used to represent the dynamic degradation process and potential failure characteristics of the key components;

[0063] The virtual life curve is combined with the Weibull reliability model, and the real-time monitoring data and historical operation data are fused through the Bayesian updating method to correct the Weibull reliability model parameters, and the optimized Weibull reliability model is output;

[0064] Based on the optimized Weibull reliability model, an optimized fault maintenance time window is generated.

[0065] Explanations, the comprehensive input data set refers to the integration of real-time monitoring data, historical operation data and abnormal probability, which is used for the input data set of the multi-physical field coupling model, the specific content includes:

[0066] Real-time monitoring data: temperature, pressure, vibration and flow values (such as temperature 25℃, pressure 2bar) of key components (such as pumps, valves) collected by sensors;

[0067] Historical operation data: state records of components in the past operation (such as average life 10000 hours, failure frequency distribution);

[0068] Abnormal probability: a value between 0 and 1 (such as 0.85) output by the abnormal pattern recognition module, reflecting the current fault risk;

[0069] Data integration method: Align the above data by timestamp to form a multi-dimensional vector (e.g., [t, T, P, V, F, P_anomaly]), and map it to the [0, 1] interval through normalization processing (e.g., Min-Max normalization) as model input.

[0070] Explanation: How to couple the multi-physical field model: The multi-physical field coupling model refers to the simultaneous analysis of the interaction of thermal, mechanical and fluid mechanical behavior through numerical simulation methods (such as COMSOL Multiphysics software), and the specific coupling steps are:

[0071] Model definition: Establish a three-dimensional geometric model of the key components (such as the impeller of the pump), and set the material properties (such as density, thermal conductivity);

[0072] Physical field setting: Build a thermal field model based on the heat conduction equation, input temperature boundary conditions; build a mechanical field based on the elasticity equation, input stress caused by vibration or pressure; build a fluid mechanics field model based on the Navier-Stokes equation, input flow and pressure conditions;

[0073] Coupling mechanism: Coupling is achieved through shared variables, such as temperature changes affecting material stress (thermal expansion coefficient) and fluid pressure affecting stress distribution;

[0074] Solution: Discretize using the finite element method, iteratively solve the steady-state or transient solution of each physical field, and output the temperature field, stress field and flow field.

[0075] Explanation: How to generate a virtual life curve through finite element analysis, including the following steps:

[0076] Stress-strain distribution: Based on the multi-physical field coupling model output (such as stress 100 MPa), generate a stress-strain distribution map of the component through FEA (such as ANSYS software);

[0077] Fatigue crack propagation prediction: Adjust the initial defect assumption (such as crack length 0.1 mm) combined with the anomaly probability (such as 0.85), and use Paris law to calculate the crack propagation rate;

[0078] Life calculation: Calculate the number of cycles N from the initial crack length to the critical length (such as 1 mm) by integrating the crack propagation equation;

[0079] Curve generation: Plot the virtual life curve (such as decreasing from 1000 hours to 0 hours) with time as the horizontal axis (convert the number of cycles to hours according to the operating conditions) and remaining life as the vertical axis, representing the dynamic degradation process.

[0080] Preferably, the device further comprises a fault maintenance trigger condition management module for managing the trigger condition of fault maintenance, which is the basis for determining the start of the fault maintenance time window, including:

[0081] Obtaining dynamic simulation data of battery cells and key components generated based on a multi-physical field coupling model;

[0082] Analyzing the fault propagation characteristics in the dynamic simulation data, constructing an adaptive fault evolution network, and identifying the fault propagation path and propagation strength between components through a graph neural network to generate an evolution topology reflecting the synergistic effect of multiple fault points;

[0083] According to the evolution topology, fuse real-time sensor data and historical fault patterns, and use a contrast learning mechanism to judge the matching degree of the current running state and the potential fault pattern, and generate a time series probability distribution of fault evolution;

[0084] If the time series probability distribution indicates that the potential fault pattern exceeds a preset threshold, a synergistic optimization strategy based on Monte Carlo simulation and reinforcement learning is used to dynamically adjust the trigger condition of the fault maintenance time window.

[0085] Explanation, what physical quantities are included in the dynamic simulation data is unclear, dynamic simulation data refers to the time series data output by the multi-physical field coupling model, which includes:

[0086] Battery cell: temperature, pressure, charge distribution;

[0087] Key components (such as pumps): stress, strain, flow rate, vibration acceleration;

[0088] Data form: multi-dimensional vector recorded by time step (such as every minute), reflecting the dynamic behavior of the component.

[0089] Explanation, the meaning of fault propagation characteristics is unclear, fault propagation characteristics refer to the rules of fault transmission between components, such as pump failure causing abnormal valve pressure, which is manifested as:

[0090] Propagation path: the causal relationship of fault from the source component (such as pump) to the affected component (such as valve);

[0091] Propagation strength: the degree of fault influence (such as a 10% increase in pressure);

[0092] Extraction method: analyze the time correlation of physical quantities in dynamic simulation data (such as stress mutation after temperature rise).

[0093] Explanation, the meaning of fault evolution network is unclear, fault evolution network is a graph-based model that represents the dynamic process of fault propagation between components, which is:

[0094] Node: Key components in the liquid cooling system (e.g., pumps, valves);

[0095] Edge: Failure propagation relationship between components (e.g., pump failure affects valve);

[0096] Construction method: Use graph neural network to process dynamic simulation data, learn propagation path and strength.

[0097] Explanation: The evolutionary topology is a graph representation of the fault evolution network over time, reflecting the synergistic effect of multiple point failures. The specific generation steps are as follows:

[0098] Input: Dynamic simulation data and real-time sensor data;

[0099] GNN processing: Update node features (e.g., pump stress) and edge weights (e.g., propagation strength) through graph convolution, formula:

[0100]

[0101] where, is the adjacency matrix, H is the node feature, W is the weight, and σ is the activation function;

[0102] Output: Topology graph evolving over time (e.g., pump-valve edge weight increases from 0.3 to 0.8);

[0103] Explanation: How to use the contrastive learning mechanism to match the current running state with the potential fault mode is unclear; the contrastive learning mechanism matches the current running state with the potential fault mode by judging similarity, which includes:

[0104] Input: Current running state (real-time sensor data, such as temperature 30℃) and historical fault mode (e.g., failure at temperature 40℃);

[0105] Feature extraction: Generate embedding vectors (e.g., 128 dimensions) for the current state and fault mode through GNN;

[0106] Loss function: Use InfoNCE loss to calculate similarity: output matching degree score (0 to 1), generate fault probability distribution.

[0107] Explanation: How to optimize based on Monte Carlo simulation and reinforcement learning is unclear, Monte Carlo simulation (Monte Carlo) and reinforcement learning (RL) are combined to optimize the triggering conditions of the fault maintenance time window, which balances cost and risk by simulating fault distribution and dynamically adjusting threshold, the specific implementation method is:

[0108] First, Monte Carlo simulation is based on the probability distribution of failure, randomly sampling multiple times (such as 1000 times) to simulate the time of failure, generating a high-risk time window distribution (such as 10-14 hours), the specific implementation steps include:

[0109] Input the current failure probability distribution, for example, the discrete distribution of probability changes over time (such as t=10 hours probability is 0.7, t=12 hours is 0.9);

[0110] Random sampling multiple times (such as 1000 times), each sampling generates a failure time point from the probability distribution;

[0111] Statistical sampling results, generate a time window distribution of failure, for example, the results show that the failure may occur within 10-14 hours, the mean is 12 hours, and the standard deviation is 1 hour;

[0112] Output the time window distribution as a reference for reinforcement learning, for example, the time window [10, 14] hours represents the high-risk interval of potential failure;

[0113] Then, reinforcement learning uses Q-learning algorithm to optimize the trigger condition, and finds the optimal threshold through iterative learning, the specific steps are:

[0114] Define the state as the current abnormal probability and time window, the action is to adjust the high and low of the trigger threshold, and the reward is to balance the reduction of maintenance cost and the reduction of failure risk;

[0115] Initialize the Q table to zero, through multiple iterations, each time according to the current state to select the action, calculate the reward, and update the Q table value according to the Q-learning formula; in each update, use the latest state and reward information to repeat the iteration until the Q table converges, output the trigger threshold with the maximum reward in each state as the optimal result;

[0116] Finally, select the action with the maximum reward in each state through the Q table, and output the optimized trigger threshold; for example, if the Q table shows that the action 0.75 has the highest Q value in the state (probability 0.85, window [10, 14]), then set the trigger threshold to 0.75; this trigger threshold can be dynamically adjusted according to the system state, for example, when the probability rises to 0.9, it may be further optimized to 0.7 to ensure timely response.

[0117] Explanation, it is not clear how to dynamically adjust the trigger condition of the failure maintenance time window, based on the updated Q table to dynamically adjust the trigger condition, the specific implementation method is as follows

[0118] Adjustment when risk increases:

[0119] If the Monte Carlo simulation shows that the failure risk is rising (e.g., the mean of the probability distribution rises from 0.7 to 0.9), and the time window shrinks from [11, 15] hours to [9, 12] hours, then the trigger threshold is lowered to respond earlier; for example, the original threshold of 0.8 is adjusted to 0.75, and the decision is based on maximizing the Q value of the state (0.9, [9, 12]) in the Q table;

[0120] Adjustment when the cost is too high:

[0121] If reinforcement learning optimization shows that the current threshold is too low, causing maintenance costs to increase, for example, the threshold of 0.75 triggers too frequently, and the cost rises from 80 yuan to 120 yuan, the reward value decreases, then the threshold is increased to 0.85; specific to the Q table evaluation, for example, the Q value of action 0.85 in state (0.85, [10, 14]) is higher than 0.75, reflecting a better balance between cost and risk;

[0122] Every hour, based on the latest sensor data and abnormal probability, the Monte Carlo simulation is re-run, the time window distribution is updated, and the Q table is adjusted through Q-learning iteration to ensure that the trigger threshold adapts to the system state.

[0123] Technical effects and advantages of the present application:

[0124] (1) The real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems provided by the present application extracts weak abnormal features through multi-source data acquisition and feature engineering, outputs abnormal probability using long short-term memory network and attention mechanism, dynamically adjusts the warning threshold combining Markov chain, optimizes the maintenance time window through Weibull model and realizes closed-loop control, effectively solves the problems of difficult detection of weak abnormal signals, failure of fixed threshold and inaccurate maintenance timing.

[0125] (2) The real-time monitoring and fault response control device for industrial and commercial liquid-cooled energy storage systems provided by the present application analyzes thermal runaway risk using multi-physical field coupling model and generates virtual life curve, optimizes maintenance time window through Weibull model and Bayesian update; uses graph neural network, contrastive learning, Monte Carlo simulation and reinforcement learning to dynamically adjust the trigger condition, outputs maintenance decisions for unknown faults, realizes accurate fault prediction and dynamic response, and improves system reliability and maintenance efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0126] Figure 1 The structure block diagram of the liquid-cooled energy storage system monitoring control device of the present application.

[0127] Figure 2 The flow chart of the fault maintenance time window optimization method of the present application. DETAILED DESCRIPTION

[0128] Exemplary embodiments of the present disclosure will be described herein below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0129] Meanwhile, it should be understood that the dimensions of the various portions shown in the accompanying drawings are not drawn to scale for the sake of convenience in description.

[0130] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the application or its use to the embodiments set forth herein.

[0131] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered part of the specification.

[0132] Embodiment 1, refer to Figure 1 the liquid-cooled energy storage system monitoring and control device structure diagram, the present application provides a kind of real-time monitoring and fault response control device of commercial liquid-cooled energy storage system as shown in Figure 1 Including:

[0133] Multi-source data acquisition and preprocessing module: by deploying high-precision sensors, real-time acquisition of temperature, pressure, vibration and flow data of key components in liquid cooling system, and preprocessing of collected data, generating denoised and standardized multi-dimensional operation data set;

[0134] Feature engineering and weak signal extraction module: using time series analysis and signal processing technology, feature vectors reflecting the state of key components are extracted from the multi-dimensional operation data set, the significance of weak abnormal signals is enhanced, and a key feature set containing weak abnormal features is generated;

[0135] Explain, in liquid-cooled energy storage system, key components (such as pump, valve) may only show a small change in temperature, pressure or vibration before failure, which is usually masked by noise or below the traditional monitoring threshold, called "weak abnormal signal", and the significance of which is enhanced by highlighting the features of these signals through feature engineering means; For example: difference analysis, high-order statistics, frequency domain transformation;

[0136] Abnormal pattern recognition module: based on the key feature set, an abnormal recognition model is constructed using long short-term memory network, a training set is constructed by historical operation data, and a trained abnormal recognition model is obtained; output the abnormal probability (value between 0 and 1) corresponding to the key component, and set the warning threshold of the abnormal probability to trigger the warning;

[0137] Failure warning and optimization module: based on Weibull reliability model to predict the remaining service life of key components, combined with confidence interval to generate failure maintenance time window, based on failure maintenance time window prediction results and abnormal probability, dynamically adjust the priority of maintaining key components and perform preventive maintenance; use the running data of key components after maintenance to update the abnormal recognition model and Weibull reliability model, form a closed-loop optimization control.

[0138] It is explained that by comprehensively considering the abnormal probability and the starting time of the failure maintenance time window, the priority of each component is determined; the specific way is to combine the high and low of the abnormal probability with the urgency of the maintenance time window, so that the components with high risk and time urgency obtain higher priority, and this comprehensive way can flexibly adjust the relative importance of the two according to actual needs, but does not need to limit the specific calculation method; according to the calculated priority, all key components are sorted, and the components with higher priority are arranged for preventive maintenance in priority; with the update of system running data (such as abnormal probability rising or time window shortening), the priority will be dynamically adjusted to ensure that the maintenance decision always adapts to the current state.

[0139] It needs to be further explained in the present application that the acquisition method of the failure maintenance time window is:

[0140] Acquisition data: acquire the average abnormal probability (such as 0.85) and historical failure data (such as component average life, failure distribution) of key components;

[0141] Model initialization: the initialized Weibull reliability model parameters, because it is suitable for describing the failure law of mechanical components, the parameters include shape parameter β, which represents failure mode; and scale parameter η, which represents life; according to historical data fitting Weibul model, adjust the model parameters combined with current abnormal probability; optimize the parameters by maximum likelihood estimation or least squares method, to ensure that the Weibull reliability model reflects the current state;

[0142] For example, if the abnormal probability is high, shorten the η value, which means that the life attenuation is accelerated; if the abnormal probability is low (such as <0.3), keep or extend η, which means that the component state is stable;

[0143] Life prediction: based on the adjusted Weibull reliability model and abnormal probability, the remaining service life of the component is calculated by the following formula, and the failure maintenance time window is generated;

[0144]

[0145] Where, t is time, F(t) is abnormal probability, t is remaining service life, t is calculated according to abnormal probability (such as 0.85), and the remaining service life is obtained;

[0146] Set the confidence interval of the remaining useful life, calculate the upper and lower limits of the remaining useful life by the inverse function of the Weibull reliability model, and generate the fault maintenance time window; for example, if the central value of the remaining useful life is 12 hours and the confidence interval is ±2 hours, the generated fault maintenance time window is 10-14 hours;

[0147] Verify that the remaining useful life is consistent with the abnormal probability trend (e.g., the higher the probability, the shorter the remaining useful life), and if abnormal, timely alarm.

[0148] Further explained in the present application is that the training process of the abnormality identification model includes the following steps:

[0149] Step S1, data set preparation and division: prepare the historical data set of the key components, extract the key feature set, and label the fault time point, and divide it into a training set (70%), a verification set (20%), and a test set (10%);

[0150] Step S2, time series data preprocessing and input construction: time series processing is performed on the historical data set to generate an input format suitable for long short-term memory network; based on the key feature set, a sliding window method is used to construct time series input, for example, with a 30-minute window and a 5-minute step, to generate sequence data containing historical trends; the sequence data is normalized (e.g., Min-Max standardization) to map the feature values to the [0, 1] interval; the standardized time series input sequence is output to capture the cumulative effect of weak signals;

[0151] Step S3, model initialization and training: set the long short-term memory network structure, including the input layer, two hidden layers (64 neurons per layer), and the output layer, which outputs the abnormal probability (a value between 0 and 1); use mean square error as the loss function, which is used to measure the difference between the model's predicted abnormal probability and the true label; train the short-term memory network through the Adam optimizer and the loss function, with 100 iterations, and use the abnormal event labels in the historical running data for supervised learning to obtain the trained abnormality identification model;

[0152] Step S4, adjust the hyperparameters of the abnormality identification model through the verification set to reduce the loss function value; use the cross-validation method to evaluate the generalization ability of the short-term memory network to ensure the accuracy of abnormality identification; determine the warning threshold of the abnormal probability according to the verification result, for example, set it to 0.8, and trigger the warning when the abnormal probability exceeds this value.

[0153] Further explained in the present application is that the device further includes a warning threshold dynamic management module, which solves the problem of fixed threshold failure of the liquid cooling system caused by working condition changes through dynamic management of the warning threshold, including:

[0154] Based on the multi-dimensional operation data set, the multi-dimensional operation data set (including temperature, pressure, vibration and flow) of the key components is collected in real time, and a time series data set containing multiple parameters is generated;

[0155] After each parameter is normalized, the fluctuation characteristics of each parameter (such as temperature, pressure, vibration, and flow) are calculated, a time-frequency analysis method based on wavelet transform is used to decompose the signal, high-frequency fluctuation components and low-frequency trend characteristics are extracted, and a multi-dimensional fluctuation feature vector v is generated, for example, v = [v w , w p , v z , v f ]; when the fluctuation feature extraction is completed, a dynamic threshold Y DT is constructed.

[0156] Y DT = α·EWMA(v) + β·MAD(v) × (1 + γ·CV(v))

[0157] Wherein, wherein EWMA(v) represents a smooth trend index, which is a weighted average of the fluctuation trend of each parameter; MAD(v) measures the dispersion; CV(v) represents the adjustment fluctuation intensity, which is the ratio of the standard deviation of each parameter to the average of each parameter, and α, β, γ are adjustment coefficients;

[0158] For ease of understanding, the acquisition process of MAD(v) is illustrated as follows:

[0159] After normalizing each parameter, v = [0.1, 0.2, 0.15, 0.3] is obtained:

[0160] The median median(v) = 0.175;

[0161] The deviation |v i -0.175| = [0.075, 0.025, 0.025, 0.125], i is the serial number;

[0162] MAD(v) = median(0.075, 0.025, 0.025, 0.125) = 0.05, indicating that the distribution is stable;

[0163] When the real-time abnormal probability exceeds the dynamic threshold Y DT , a warning signal is triggered.

[0164] Further explained in the present application is that the early warning threshold dynamic management module further comprises:

[0165] Based on the Markov chain, the short-term fluctuation trend of the multi-dimensional operation data set is predicted, and the state transition probability of each parameter in the future time (such as 1 hour) is calculated;

[0166] The dynamic threshold formula is adjusted in combination with the prediction result, and a trend factor δ (range 0.8-1.2) is introduced, when the fluctuation trend rises, δ < 1, the early warning threshold is reduced, and when the fluctuation trend falls, δ > 1, the early warning threshold is increased.

[0167] It needs to be further explained in the application that the application does not make specific limitation on the threshold updating method: the sliding window data is updated every M hours (M is a predetermined time interval), and the dynamic early warning threshold is recalculated to adapt to the system change;

[0168] Or, set an adaptive updating frequency (such as updating the threshold once every 10 minutes when the abnormal probability change rate is greater than 0.1).

[0169] It needs to be further explained in the application that the feature engineering and weak signal extraction module comprises:

[0170] Wavelet transform is used to decompose the signal of the multi-dimensional running data set, and high-frequency noise and low-frequency trend components are separated;

[0171] The intrinsic mode function (IMF) of the weak abnormal signal is extracted based on empirical mode decomposition (EMD) (the intrinsic mode function is used to adaptively decompose the signal, separate features of different scales, and thus extract the weak abnormality hidden in the signal), the instantaneous frequency and the instantaneous amplitude of each intrinsic mode function are calculated through Hilbert transform, and a multi-dimensional feature vector reflecting the abnormal trend is generated; the instantaneous frequency: represents the local frequency of the signal at a certain time, and can reveal the change of the signal frequency with time. For example, if an abnormality occurs in the system, the instantaneous frequency of a certain IMF may suddenly increase or decrease;

[0172] It is explained that, in the embodiment of the application, the wavelet transform uses Daubechies wavelet to decompose the signal to level 3, separates high-frequency noise and low-frequency trend, empirical mode decomposition (EMD) extracts three intrinsic mode functions (IMFs), calculates the instantaneous frequency and the instantaneous amplitude through Hilbert transform, and principal component analysis retains the features of the first 90% variance;

[0173] The instantaneous amplitude: reflects the intensity or energy of the signal at a certain time. If the abnormality causes a sudden change in the signal amplitude (such as sudden amplification or weakening), the instantaneous amplitude can capture such changes.

[0174] The multi-dimensional feature vector is reduced in dimension through principal component analysis, and a key feature set is formed. The accuracy and robustness of abnormality detection are enhanced.

[0175] It needs to be further explained in the application that the training process of the abnormality recognition model further comprises:

[0176] Introduce attention mechanism in long short-term memory network, assign weight to each time step based on the characteristics of weak abnormal signals, and highlight the influence of weak abnormal signals on abnormal probability;

[0177] Explain that attention mechanism will calculate a weight value (usually between 0 and 1) for each time step in the sequence, representing the relative importance of the time step to the final task (here, abnormal probability prediction);

[0178] Set up a bidirectional LSTM structure to capture historical and future trends and dynamically adjust the number of hidden layer neurons;

[0179] Use Dropout regularization (dropout rate 0.2) to prevent overfitting, combined with early stopping strategy (stop training when validation set loss does not decrease for 10 consecutive times).

[0180] In one possible embodiment, according to the abnormal probability output by the abnormal pattern recognition module, the sampling frequency of the sensor is adjusted; for example, when the abnormal probability exceeds 0.7, the sampling frequency is increased to 10 times per second, otherwise it is kept at 1 time per minute, generating raw multi-dimensional operating data with adaptive frequency.

[0181] Summary: The embodiment of the present application acquires multi-dimensional operating data of key components in the liquid-cooled energy storage system in real time by deploying a multi-source data acquisition and preprocessing module, extracts weak abnormal features using wavelet transform and empirical mode decomposition technology through the feature engineering and weak signal extraction module, and generates a key feature set; The long short-term memory network and attention mechanism of the abnormal pattern recognition module are used to accurately output the abnormal probability, and the pre-warning threshold dynamic management module is used to predict the fluctuation trend based on Markov chain and dynamically adjust the threshold; Finally, through the Weibull reliability model prediction of the fault pre-warning and optimization module, the remaining life is predicted and the maintenance time window is optimized, realizing closed-loop optimization control; effectively solving the problems of difficult detection of weak abnormal signals, failure of fixed threshold and inaccuracy of maintenance timing caused by complex and variable working conditions of liquid-cooled energy storage system;

[0182] In addition, through multi-level signal processing, intelligent prediction and dynamic optimization, the robustness of system monitoring and the timeliness of fault response are improved, the fault risk and maintenance cost are significantly reduced, and an efficient solution is provided for stable operation of industrial and commercial liquid-cooled energy storage systems.

[0183] Embodiment 2, the difference between the embodiment of the present application and embodiment 1 is that the device further comprises a fault maintenance time window optimization module for obtaining an optimized fault maintenance time window, refer to Figure 2 the fault maintenance time window optimization method flowchart, including:

[0184] The abnormal probability of the key components in the liquid cooling system is obtained by real-time monitoring, and a comprehensive input data set reflecting the state of the components is generated by combining historical operation data and the current multi-dimensional operation data set;

[0185] The thermal, mechanical and fluid mechanics behaviors of the key components under the current abnormal probability are simulated by using a multi-physical field coupling model based on the comprehensive input data set and the abnormal probability.

[0186] In one possible embodiment, the electrochemical behavior is introduced to simulate the temperature-pressure-charge distribution coupling effect of the battery cell, and the influence of the non-uniform distribution of the cooling liquid flow on the thermal runaway risk is evaluated in combination with the liquid flow dynamic effect.

[0187] The stress-strain distribution map is generated by finite element analysis, the fatigue crack propagation time is predicted in combination with the abnormal probability, the simulation accuracy is adjusted by using dynamic mesh self-adaptation, the virtual life curve of the component state changing with time is generated, and the virtual life curve is used to represent the dynamic degradation process and potential failure characteristics of the key components.

[0188] The virtual life curve is combined with the Weibull reliability model, the real-time monitoring data and the historical operation data are fused by the Bayesian updating method, the Weibull reliability model parameters (shape parameter β and scale parameter η) are corrected, and the fitting error of the virtual life curve is controlled to be within 5%, and the optimized Weibull reliability model is output.

[0189] The optimized failure maintenance time window is generated based on the optimized Weibull reliability model, and an optimized time basis is provided for maintenance decision.

[0190] It is explained that the comprehensive input data set is used as the input of the multi-physical field model, and defines the starting point and boundary of the simulation; for example, if the abnormal probability is high, the model may assume a higher initial temperature or pressure; the dynamic physical behavior of the component is simulated by using the multi-physical field model, the simulation is not static but dynamic, and reflects the possible degradation process of the component from the current state (abnormal probability 0.85) to the future; the degradation index is extracted from the simulation result, the time series data is generated, the virtual life curve is drawn, and the degradation and failure characteristics are represented; the optimized failure maintenance time window is generated, and an optimized time basis is provided for maintenance decision.

[0191] It needs to be further explained in the embodiments of the present application that the device further comprises a failure maintenance trigger condition management module for managing the trigger condition of the failure maintenance, and the trigger condition is the basis for starting the failure maintenance time window, which is usually a threshold or a rule. For example, when the abnormal probability exceeds 0.8, the calculation or execution of the maintenance window is triggered, including:

[0192] It is explained that the fault maintenance time window is not always active, but needs a condition to determine when to calculate or apply the window. The dynamic adjustment trigger condition directly affects the timing of the window generation;

[0193] Obtain dynamic simulation data of battery cells and key components generated based on a multi-physical field coupling model;

[0194] Analyze the fault propagation characteristics in the dynamic simulation data, construct an adaptive fault evolution network, and identify the fault propagation path and propagation strength between components through a graph neural network to generate an evolution topology reflecting the synergistic effect of multiple faults;

[0195] According to the evolution topology, fuse real-time sensor data and historical fault patterns, and use a contrast learning mechanism to judge the matching degree of the current running state and the potential fault pattern, and generate a time series probability distribution of fault evolution;

[0196] If the time series probability distribution indicates that the potential fault pattern exceeds a preset threshold, a synergistic optimization strategy based on Monte Carlo simulation and reinforcement learning is used to dynamically adjust the trigger condition of the fault maintenance time window, and an optimized maintenance decision for unknown fault patterns is output.

[0197] In one possible embodiment, the device further comprises a maintenance strategy fusion decision module, which dynamically generates a cross-component collaborative maintenance scheme based on the trigger condition of the fault maintenance time window, the virtual life curve generated by the multi-physical field coupling model, and the fault propagation topology between components through an adaptive hybrid learning algorithm; The maintenance strategy fusion decision module performs the following operations:

[0198] (1) Input the residual life prediction value of the virtual life curve, the real-time anomaly probability, and the edge weight of the fault propagation topology into a graph convolutional neural network to generate a component maintenance urgency score;

[0199] (2) Based on the current operating load of the liquid cooling system, the availability of external maintenance resources, and the preset downtime tolerance threshold, construct a multi-objective optimization constraint condition containing an economic penalty function and a risk diffusion suppression function;

[0200] (3) When a sudden change in component degradation rate or an increase in cross-regional fault correlation is detected, the latest failure mode features are fused through an online incremental learning mechanism to reconstruct the decision boundary of the maintenance strategy and update the weight parameters.

[0201] It is explained that the graph convolutional neural network is used to generate a component maintenance urgency score, and the specific operation is as follows:

[0202] Step 1, input data definition:

[0203] Node features: Each component as a node in the graph structure, its attributes include:

[0204] Remaining life: The remaining effective working time of the component predicted by the virtual life curve (e.g. pump body remaining life 1200 hours);

[0205] Real-time anomaly probability: The current failure risk value output by the anomaly identification model (e.g. 0.85 indicates high risk);

[0206] Edge weight: Reflects the connection relationship between components (e.g. the propagation probability from pump to valve is 0.6), provided by the fault propagation topology of claim 9;

[0207] Step 2, graph convolution operation:

[0208] Neighborhood information aggregation: For each node, calculate the weighted average of the features of all adjacent nodes, and the weight is obtained by normalizing the edge weight; for example, when the pump node aggregates the features of the valve node, the contribution of the valve feature value is determined by the ratio of the pump→valve edge weight (0.6) to the sum of all adjacent edge weights;

[0209] Feature transformation and activation: Superimpose the aggregated neighborhood features and the node's own features, perform linear transformation through a trainable parameter matrix, and apply a nonlinear activation function (such as ReLU) to generate updated node features;

[0210] Step 3, output layer processing: After multiple layers of graph convolution operations, finally map the node features to the interval [0, 1] through the Softmax function to generate the maintenance urgency score of each component.

[0211] Explanation, the economic penalty function includes maintenance labor cost, spare parts consumption cost and downtime loss cost, the weight is dynamically set according to the running scene (for example, in industrial and commercial scenarios, the weight proportion of labor cost is 60%, spare parts consumption is 30%, and downtime loss is 10%); the risk diffusion suppression function includes the correlation evaluation of component anomaly probability and fault propagation intensity, the calculation method is: multiply the real-time anomaly probability of each component by the edge weight propagated to the key node, then multiply the accumulated value by the preset risk suppression coefficient to generate the system-level risk score; the economic penalty function prioritizes maintenance cost control, the risk diffusion suppression function ensures that high-risk components are prioritized, both dynamically balanced through a preset weight coefficient, combined with real-time resource constraints and fault propagation path optimization decision boundary, to achieve global reliability optimization and economic benefit maximization of the liquid-cooled energy storage system maintenance scheme.

[0212] When a sudden change in component degradation rate or an increase in cross-region fault correlation is detected, the latest failure mode features are integrated through an online incremental learning mechanism to reconstruct the decision boundary of the maintenance strategy and update the weight parameters.

[0213] The adaptive hybrid learning algorithm includes deep reinforcement learning and transfer learning, specifically including:

[0214] Deep reinforcement learning branch: based on virtual life curve data, the long-term value return of maintenance actions (such as maintenance cost reduction rate, risk suppression gain) is calculated through deep reinforcement learning algorithm;

[0215] Transfer learning branch: using similar failure modes in historical maintenance records (such as pump valve linkage failure case library), through pre-training model parameter transfer, accelerating policy convergence in new scenarios;

[0216] Dynamic generation logic: input data: component remaining life (virtual life curve), abnormal probability (LSTM output), fault propagation topology edge weight;

[0217] Output scheme: generate maintenance sequence (such as "replace pump body first -> delay repair cooling pipeline"), and quantify time window (such as 12±2 hours) and execution granularity (such as local maintenance or system downtime).

[0218] Summary: the embodiment of the present application adds a fault maintenance time window optimization module and a fault maintenance trigger condition management module based on embodiment 1, uses real-time monitoring to obtain abnormal probability and combines historical and current multi-dimensional operation data to generate comprehensive input data set; through the multi-physical field coupling model to simulate the thermal, mechanical and fluid mechanical behavior of key components, introduce electrochemical and liquid flow dynamic effect analysis battery thermal runaway risk, use finite element analysis to generate virtual life curve; based on Weibull reliability model and Bayesian update optimization maintenance time window. At the same time, through the graph neural network to build adaptive fault evolution network, fusion comparative learning and Monte Carlo simulation and reinforcement learning, dynamically adjust the trigger condition, output the maintenance decision for unknown failure mode. Realize the accurate prediction and dynamic response of liquid-cooled energy storage system under complex working conditions, solve the problem that traditional method is difficult to deal with unknown failure mode and maintenance time is not optimal, improve the reliability and maintenance efficiency of the system.

[0219] Finally: the above only for the preferred embodiment of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A real-time monitoring and fault response control device for commercial liquid-cooled energy storage systems, characterized in that, Comprise: Multi-source data acquisition and preprocessing module: real-time acquisition of temperature, pressure, vibration and flow data of key components in liquid cooling system, and denoising and standardization processing of collected data to generate multi-dimensional running data set; Feature engineering and weak signal extraction module: using wavelet transform and empirical mode decomposition technology, separate high-frequency noise and low-frequency trend component from the multi-dimensional running data set, extract instantaneous frequency and instantaneous amplitude features reflecting the state of key components, and generate key feature set through principal component analysis dimension reduction; Abnormal pattern recognition module: based on long short-term memory network, build an abnormality recognition model, input the key feature set and output the abnormal probability of each key component, and set the warning threshold of abnormal probability to trigger the warning; The long short-term memory network introduces an attention mechanism to assign weights to time steps, and combines a bidirectional LSTM structure to capture historical and future trends; Fault warning and optimization module: based on Weibull reliability model to predict the remaining useful life of key components, combined with confidence interval to generate fault maintenance time window, and dynamically adjust the maintenance priority according to the abnormal probability; Fault maintenance time window optimization module, for obtaining the optimized fault maintenance time window, comprising: Obtain the abnormal probability of key components in the liquid cooling system through real-time monitoring, and generate a comprehensive input data set reflecting the state of the components by combining historical operation data and current multi-dimensional running data set; Using multi-physical field coupling model, based on the comprehensive input data set and the abnormal probability, simulate the thermal, mechanical and fluid mechanics behavior of the key components under the current abnormal probability; Generate stress-strain distribution map through finite element analysis, combined with abnormal probability to predict fatigue crack propagation time, generate virtual life curve of component state changing with time; Combine the virtual life curve with the Weibull reliability model, and fuse real-time monitoring data and historical operation data through Bayesian updating method to correct the Weibull reliability model parameters, output the optimized Weibull reliability model; Based on the optimized Weibull reliability model, generate the optimized fault maintenance time window.

2. A real-time monitoring and fault response control device for industrial liquid-cooled energy storage system as claimed in claim 1, wherein, The acquisition method of the fault maintenance time window is: Obtain the average abnormal probability and historical fault data of the key components; initializing weibull reliability model parameters, including shape parameter β ; and scale parameters η ; fitting Weibul model according to historical data, adjusting model parameters combined with current anomaly probability; Based on the adjusted Weibull reliability model and abnormal probability, calculate the remaining useful life t of the component through the following formula, and generate the fault maintenance time window; ； where t is time, is an abnormal probability, t is a remaining service life, and the remaining service life is obtained by backstepping t according to the abnormal probability. Set the confidence interval of the remaining useful life, calculate the upper and lower limits of the remaining useful life through the inverse function of the Weibull reliability model, generate the fault maintenance time window; verify that the remaining useful life is consistent with the trend of abnormal probability.

3. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 1 wherein, The training process of the abnormality recognition model includes the following steps: Step S1, prepare the historical data set of key components, extract the key feature set, and label the fault time point, divide into training set, validation set and test set; Step S2, time series processing is performed on the historical data set to generate an input format suitable for a long short-term memory network; based on the key feature set, a sliding window method is used to construct a time series input to generate sequence data containing historical trends; the sequence data is normalized to map the feature values to the [0, 1] interval; and a standardized time series input sequence is output to capture the cumulative effect of weak signals; Step S3, a long short-term memory network structure is set, including an input layer, a hidden layer, and an output layer, and an abnormal probability is output; a mean square error is used as a loss function; and the short-term memory network is trained through an Adam optimizer and the loss function to obtain a trained abnormality recognition model.

4. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 1 wherein, Further comprising: An early warning threshold dynamic management module, comprising: Based on the multi-dimensional operation data set, the multi-dimensional operation data set of the key components is collected in real time to generate a time series data set containing multiple parameters; After normalizing each parameter, the fluctuation characteristics of each parameter are calculated, a time-frequency analysis method based on wavelet transform is used to decompose the signal, high-frequency fluctuation components and low-frequency trend characteristics are extracted, a multi-dimensional fluctuation characteristic vector v is generated, and a dynamic threshold is constructed Equation: ； Wherein, wherein EWMA(v) represents a smoothing trend index, which is a weighted average of the fluctuation trend of each parameter; MAD(v) represents the dispersion; CV(v) adjusts the fluctuation intensity, which is the ratio of the standard deviation of each parameter to the mean of each parameter, and a, β, γ are adjustment coefficients.

5. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 3 wherein, The training process of the abnormality recognition model further comprises: An attention mechanism is introduced into the long short-term memory network, a weight is assigned to each time step based on the characteristics of the weak abnormal signal, and the influence of the weak abnormal signal on the abnormal probability is highlighted; A bidirectional LSTM structure is set to capture historical and future trends simultaneously and dynamically adjust the number of hidden layer neurons; Dropout regularization is used to prevent overfitting, combined with an early stopping strategy.

6. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 1 wherein, According to the abnormal probability output by the abnormal pattern recognition module, the sampling frequency of the sensor is adjusted.

7. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 4 wherein, The early warning threshold dynamic management module further comprises: Based on the Markov chain, the short-term fluctuation trend of the multi-dimensional operation data set is predicted, and the state transition probability of each parameter in the future time is calculated; The prediction result is combined to adjust the dynamic threshold formula, and a trend factor δ is introduced, when the fluctuation trend rises, δ<1, the early warning threshold is reduced, and when it falls, δ>1, the early warning threshold is increased.

8. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 1 wherein, The device further comprises a fault maintenance trigger condition management module for managing the trigger conditions of fault maintenance, which is the basis for determining the start of the fault maintenance time window, comprising: Obtaining dynamic simulation data of battery cells and key components generated based on a multi-physical field coupling model; Analyzing the fault propagation characteristics in the dynamic simulation data, constructing an adaptive fault evolution network, and using a graph neural network to identify the fault propagation path and propagation strength between components to generate an evolution topology structure reflecting the synergistic effect of multiple faults; According to the evolution topology structure, the real-time sensor data and the historical fault mode are fused, the matching degree between the current operating state and the potential fault mode is judged by using a contrast learning mechanism, and a time series probability distribution of fault evolution is generated; If the time series probability distribution indicates that the potential fault mode exceeds a preset threshold, a synergistic optimization strategy based on Monte Carlo simulation and reinforcement learning is used to dynamically adjust the trigger conditions of the fault maintenance time window.

9. A real time monitoring and fault response control device for industrial liquid cooled energy storage system as claimed in claim 8, wherein, Also include maintenance policy fusion decision module, based on the trigger condition of fault maintenance time window, virtual life curve and fault propagation topology between components, through adaptive hybrid learning algorithm to dynamically generate cross-component collaborative maintenance scheme, including: Input the residual life prediction value of the virtual life curve, real-time anomaly probability and edge weight of the fault propagation topology to the graph convolutional neural network to evaluate the maintenance urgency of each component; Based on the current operating load of the liquid cooling system, the availability of external maintenance resources and the preset downtime tolerance threshold, build multi-objective optimization constraints including economic penalty function and risk diffusion suppression function; When detecting component degradation rate mutation or cross-region fault correlation enhancement, reconstruct the decision boundary of the maintenance policy and update the weight parameters by online incremental learning mechanism to integrate the latest failure mode features.

Images