Electric vehicle battery dynamic safety monitoring method and system based on Internet of Things

Abstract

The invention discloses an electric vehicle battery dynamic safety monitoring method and system based on the Internet of Things, and relates to the technical field of battery safety monitoring. Firstly, low-frequency initial battery data of a target vehicle is obtained, and an initial sample is generated through rapid preprocessing and time sequence division; inputting a hybrid prediction model to obtain initial two-dimensional prediction characteristics of the highest battery cell voltage and the probe temperature, detecting through a target algorithm to obtain an initial fault probability, and adaptively adjusting the sampling frequency according to the initial fault probability; then high-frequency update data are collected and uploaded to the cloud together with the initial fault probability, and the cloud updates parameters of the hybrid prediction model by using an optimization algorithm; updating data are substituted into the optimized model to obtain two-dimensional features of the target, the accurate fault probability is output through anomaly detection, and a hierarchical monitoring scheme is formulated according to the accurate fault probability. The problem that the safety monitoring efficiency of the electric vehicle battery is low due to dual solidification of parameters and models in the existing safety monitoring of the electric vehicle battery is solved.

Description

Technical Field

[0001] This invention belongs to the field of battery safety monitoring technology, specifically relating to a method and system for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things. Background Technology

[0002] With the rapid development of the electric vehicle industry, the power battery, as the core power source of electric vehicles, directly affects vehicle driving safety, the personal safety of passengers, and property safety, and is one of the key constraints on the large-scale development of the electric vehicle industry. Currently, electric vehicle battery safety monitoring systems based on Internet of Things (IoT) technology have become the mainstream technology direction in the industry. Their core requirement is to collect battery operating status data and analyze battery operating trends to achieve early warning and dynamic monitoring of battery faults, thereby preventing safety accidents such as battery thermal runaway and fires.

[0003] Existing monitoring solutions generally adopt a monitoring mode with fixed sampling frequency and fixed model parameters. The core contradiction of this mode is that it cannot simultaneously meet the requirements of prediction accuracy, terminal computing power consumption, and IoT transmission bandwidth. It falls into a dilemma of either sacrificing safety or wasting resources. Specifically, if a low initial sampling frequency is used, although it can reduce vehicle-side power consumption, save IoT bandwidth, and reduce the computing power pressure on the edge, the data collection of battery operating status is not dense, and subtle voltage and temperature anomalies are easily missed. As a result, the hybrid prediction model cannot capture early fault characteristics, the fault false alarm rate increases, and the core requirement of dynamic safety monitoring cannot be achieved. If a high initial sampling frequency is used, although it can capture subtle battery operating anomalies, improve model prediction accuracy, and achieve early fault warning, continuous high-frequency sampling at the vehicle end will lead to a surge in power consumption, which will seriously affect the electric vehicle's range. Uploading a large amount of sampled data to the cloud will consume too much IoT bandwidth and increase transmission costs. At the same time, the edge needs to process massive amounts of data, resulting in excessive computing power pressure. Moreover, after the cloud receives massive amounts of data, the efficiency of model training iteration will decrease, resulting in a waste of computing resources.

[0004] Furthermore, existing solutions also suffer from secondary risks due to model fixation, which is a major reason for the long-term decline in monitoring accuracy. Once the vehicle model is deployed, its parameters remain fixed and cannot adapt to changes in battery operating status caused by battery aging, environmental changes, and differences in usage scenarios. During battery operation, the patterns of voltage and temperature changes will change with usage time and environmental changes. The fixed-parameter hybrid prediction model will gradually experience model aging problems with increased prediction deviations, which in turn leads to a decrease in the accuracy of fault detection. Early faults that could originally be detected will gradually become missed faults, ultimately losing the core significance of monitoring.

[0005] In summary, existing IoT-based electric vehicle battery safety monitoring systems generally suffer from the core defect of dual fixed parameters and models. Both the sampling frequency and model parameters are set to fixed values. Fixed sampling frequency directly leads to a triangular contradiction of data redundancy, power waste, and insufficient effective information. Fixed model parameters, on the other hand, cause the model to be unable to adapt to battery aging and changes in operating conditions, and the accuracy to continuously decline. These two types of fixed problems restrict and affect each other, resulting in low efficiency in electric vehicle battery safety monitoring. Summary of the Invention

[0006] The purpose of this invention is to solve the problem that existing electric vehicle battery safety monitoring generally suffers from the problem of dual fixed parameters and models, resulting in low efficiency of electric vehicle battery safety monitoring. Therefore, this invention proposes an electric vehicle battery dynamic safety monitoring method and system based on the Internet of Things.

[0007] In a first aspect of this invention, a method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things is first proposed, the method comprising: After acquiring battery operating status data of the target vehicle at an initial sampling frequency and performing rapid preprocessing, the preprocessed battery operating status data is divided into an initial input sample set at a preset time step. Substituting the initial input sample set into the hybrid prediction model yields the initial two-dimensional features; the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps. Based on the target algorithm, the initial two-dimensional features are used to perform fault detection to obtain a preliminary fault probability; Based on the initial failure probability, a target sampling frequency is determined according to preset rules, and the battery operating status of the target vehicle is collected at the target sampling frequency to obtain updated battery sampling data; The updated battery sampling data and the preliminary failure probability are uploaded to the cloud server so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain optimized parameters. After updating the hybrid prediction model according to the optimization parameters, the updated battery sampling data is substituted into the updated hybrid prediction model to obtain the target two-dimensional features; Based on the target algorithm, anomaly detection is performed on the two-dimensional features of the target to obtain the target failure probability, and a monitoring scheme is determined based on the target failure probability.

[0008] This technical solution adopts a closed-loop design of dynamic sampling adjustment and real-time model optimization. First, an initial fault probability is obtained through initial sampling, preprocessing, and model calculation. Then, the sampling frequency is dynamically adjusted according to this probability to balance accuracy and resource consumption. Subsequently, the real-time sampling data and the initial fault probability are uploaded to the cloud. The cloud optimizes the model parameters through a preset optimization algorithm and sends them to the terminal to update the model. Finally, the updated model is used to accurately detect the dynamic sampling data to obtain the target fault probability and determine the monitoring scheme. The entire solution focuses on the core operating parameters of the battery and optimizes the model structure. It not only gets rid of the dilemma of fixed sampling frequency, but also solves the problem of accuracy decay due to model solidification, realizing long-term, efficient, and economical dynamic safety monitoring of the battery.

[0009] Optionally, the hybrid prediction model includes a TCN feature extraction submodule, an LSTM temporal modeling submodule, and a parallel prediction output layer connected in sequence; The TCN feature extraction submodule includes, from input to output, a linear embedding layer, three causal dilated convolutional layers, a residual connection layer, a Dropout layer, and a linear compression layer; the kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. The LSTM time series modeling submodule includes a first LSTM layer, a second LSTM layer, and a Dropout layer from input to output. The hidden layer dimension of the first LSTM layer is 128, and the return sequence mode is enabled. The hidden layer dimension of the second LSTM layer is set to 64, and the return sequence mode is disabled. The dropout rate of the Dropout layer is set to 0.5. The parallel prediction output layer consists of two fully connected layers.

[0010] Optionally, the preliminary fault probability is obtained by performing anomaly detection on the initial two-dimensional features based on the target algorithm, including: Each element in the initial two-dimensional feature matrix is ​​used as a prediction data point; The distance between each predicted data point and other data points is calculated based on Euclidean distance, and the k-neighborhood of each data point is determined. For each predicted data point, calculate the reachability distance of the predicted data point to every data point in its k-neighborhood, and determine the local reachability density of the predicted data point based on the reachability distance; Based on the local reachability density of each predicted data point and the local reachability density of its corresponding k-neighborhood, calculate the local anomaly factor of the predicted data point. Through formula The local anomaly factor is converted into the initial failure probability; where A is the local anomaly factor, and the initial failure probability is zero when the local anomaly factor is less than or equal to 1.

[0011] By employing a local anomaly factor correlation algorithm, the algorithm can accurately capture the gradual and subtle changes in the highest single-cell voltage and the highest probe temperature in two-dimensional features, enabling precise identification of early battery faults. Simultaneously, by setting the fault probability to 0 when the local anomaly factor is less than or equal to 11, it filters out minor fluctuations in normal data, significantly reducing false positives and false negatives, and improving the reliability and specificity of fault detection. Furthermore, the detection process has low computational complexity, requires no large number of samples for training, effectively simplifies the terminal's computational burden, reduces vehicle-side power consumption, and avoids consuming excessive IoT transmission bandwidth.

[0012] Compared to traditional threshold detection methods and complex deep learning anomaly detection algorithms, this detection process overcomes the shortcomings of traditional threshold detection methods, such as their inability to adapt to battery aging, environmental changes, and the tendency to miss detections. It also avoids the drawbacks of complex algorithms, such as high computational load and high power consumption. This achieves a balance between detection accuracy, power consumption control, and bandwidth saving, providing key technical support for long-term, efficient, and economical dynamic safety monitoring of batteries and significantly improving the practicality and applicability of the entire battery safety monitoring system.

[0013] Optionally, the target sampling frequency, determined based on preset rules for the initial failure probability, includes: If there exists a predicted data point with a local reachability density greater than 1, then the data point is recorded as an anomaly detection point. The anomaly percentage is obtained by calculating the proportion of anomaly detection points among all predicted data points. If the abnormality percentage is greater than a preset threshold, then calculate the preliminary average failure probability of all abnormal detection points; The target sampling frequency is determined based on the initial average failure probability and the preset failure threshold.

[0014] By defining anomaly detection points as those with a local reachability density greater than 1, the criteria for judging abnormal data are further clarified. Combined with the quantitative analysis of the proportion of abnormal data, the overall degree of abnormality in battery operation can be accurately reflected, avoiding misjudgment caused by a single abnormal data point and improving the rationality and comprehensiveness of anomaly identification.

[0015] Optionally, the updated battery sampling data and the preliminary failure probability are uploaded to a cloud server, so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model to obtain optimized parameters, including: Using updated battery sampling data as model input samples and the corresponding preliminary failure probabilities as supervision labels, a training sample set for model optimization is constructed. Based on the deviation between the predicted value output by the hybrid prediction model and the true value in the updated battery sampling data, a loss function for model training is constructed. The network coefficients of the hybrid prediction model are iteratively adjusted based on the gradient of the loss function using a preset optimization algorithm. When the loss function converges to a preset threshold or the number of iterations reaches a preset number, parameter iteration is stopped, and the model parameters that meet the convergence condition after iteration are determined as optimization parameters.

[0016] The parameter update process of the hybrid prediction model, which determines the optimization parameters through dual stopping conditions, can effectively solve the pain point of accuracy decay caused by model solidification in the background technology. It enables the model to dynamically adapt to battery aging, environmental and scene changes to maintain prediction accuracy and reduce false negative rate. At the same time, it improves the optimization targeting by building training samples with real-time data. Local updates on the terminal cannot support large-scale computing and will increase power consumption. Without using real-time updated data to build training samples, the parameters will become out of touch with the actual state of the battery. Therefore, by using this solution for parameter updates, existing defects can be accurately solved and the reliability of the monitoring system can be guaranteed.

[0017] In a second aspect of this invention, an Internet of Things-based dynamic safety monitoring system for electric vehicle batteries is proposed, comprising: The data partitioning module is used to acquire battery operating status data collected by the target vehicle at an initial sampling frequency, perform rapid preprocessing, and then partition the preprocessed battery operating status data at a preset time step to obtain an initial input sample set. An initial two-dimensional feature determination module is used to substitute the initial input sample set into the hybrid prediction model to obtain initial two-dimensional features; the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps. The preliminary fault probability determination module is used to perform fault detection on the initial two-dimensional features based on the target algorithm to obtain the preliminary fault probability. The target sampling frequency determination module is used to determine the target sampling frequency based on preset rules for the initial failure probability, and to collect the battery operating status of the target vehicle at the target sampling frequency to obtain updated battery sampling data. The model optimization parameter determination module is used to upload the updated battery sampling data and the preliminary failure probability to the cloud server, so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain the optimized parameters. The target two-dimensional feature determination module is used to update the hybrid prediction model according to the optimization parameters, and then substitute the updated battery sampling data into the updated hybrid prediction model to obtain the target two-dimensional features. The monitoring scheme generation module is used to perform anomaly detection on the two-dimensional features of the target based on the target algorithm to obtain the target failure probability, and to determine the monitoring scheme based on the target failure probability.

[0018] Optionally, the hybrid prediction model includes a TCN feature extraction submodule, an LSTM temporal modeling submodule, and a parallel prediction output layer connected in sequence; The TCN feature extraction submodule includes, from input to output, a linear embedding layer, three causal dilated convolutional layers, a residual connection layer, a Dropout layer, and a linear compression layer; the kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. The LSTM time series modeling submodule includes a first LSTM layer, a second LSTM layer, and a Dropout layer from input to output. The hidden layer dimension of the first LSTM layer is 128, and the return sequence mode is enabled. The hidden layer dimension of the second LSTM layer is set to 64, and the return sequence mode is disabled. The dropout rate of the Dropout layer is set to 0.5. The parallel prediction output layer consists of two fully connected layers.

[0019] Optionally, the preliminary fault probability determination module includes: A prediction data point generation module is used to take each element in the initial two-dimensional feature matrix as a prediction data point. The neighborhood partitioning module is used to calculate the distance between each predicted data point and other data points based on Euclidean distance, and to determine the k-neighborhood of each data point; The local reachability density determination module is used to calculate the reachability distance of each predicted data point in the k-neighborhood for each predicted data point, and determine the local reachability density of the predicted data point based on the reachability distance. The local anomaly factor determination module is used to calculate the local anomaly factor of each predicted data point based on the local reachability density of each predicted data point and the local reachability density of its corresponding k-neighborhood. The local anomaly factor transformation module is used to transform factors through formulas. The local anomaly factor is converted into the initial failure probability; where A is the local anomaly factor, and the initial failure probability is zero when the local anomaly factor is less than or equal to 1.

[0020] Optionally, the target sampling frequency determination module includes: The anomaly detection point marking module is used to mark a data point as an anomaly detection point when there is a local reachability density greater than 1 corresponding to a predicted data point. The anomaly percentage determination module is used to calculate the percentage of anomaly detection points among all predicted data points to obtain the anomaly percentage value. The preliminary fault probability average determination module is used to calculate the preliminary fault probability average of all abnormal detection points if the abnormality ratio value is greater than a preset threshold. The target sampling frequency generation module is used to determine the target sampling frequency based on the preliminary average fault probability and the preset fault threshold.

[0021] Optionally, the model optimization parameter determination module includes: The training sample set generation module is used to construct a training sample set for model optimization by using updated battery sampling data as model input samples and the corresponding preliminary failure probability as supervision labels. The loss function construction module is used to construct the loss function for model training based on the deviation between the predicted value output by the hybrid prediction model and the true value in the updated battery sampling data. The network coefficient optimization module is used to iteratively adjust the network coefficients of the hybrid prediction model based on the gradient of the loss function using a preset optimization algorithm. The optimization parameter generation module is used to stop parameter iteration when the loss function converges to a preset threshold or the number of iterations reaches a preset number, and to determine the model parameters that meet the convergence conditions after iteration as optimization parameters.

[0022] The beneficial effects of this invention are: This invention proposes an IoT-based dynamic safety monitoring method for electric vehicle batteries. Through low-frequency initial sampling, rapid preprocessing, and time-series sample construction, a hybrid prediction model outputs two-dimensional look-ahead features of battery voltage and temperature. Combined with a target anomaly detection algorithm, a preliminary fault probability is obtained, enabling adaptive dynamic adjustment of the sampling frequency. Then, based on high-frequency updated sampling data and fault probabilities, model parameters are iteratively optimized in the cloud to further improve prediction and anomaly detection accuracy. Finally, a hierarchical monitoring strategy is formulated based on the precise target fault probability. This method ensures accurate early-stage battery fault identification, effectively reduces missed detections and false positives, and significantly reduces sampling and transmission overhead under normal conditions. It balances real-time monitoring, system reliability, and on-board energy consumption optimization, solving the problem of low efficiency in existing electric vehicle battery safety monitoring due to the dual fixation of parameters and models. This method achieves intelligent battery safety monitoring. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 A flowchart illustrating an IoT-based dynamic safety monitoring method for electric vehicle batteries, provided as an embodiment of the present invention. Figure 2 A flowchart for quantifying battery failure probability is provided in an embodiment of the present invention; Figure 3 This is a framework diagram of an IoT-based dynamic safety monitoring system for electric vehicle batteries, provided as an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0026] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention provides a method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things (IoT). See also... Figure 1 , Figure 1 A flowchart illustrating an IoT-based dynamic safety monitoring method for electric vehicle batteries, provided as an embodiment of the present invention. The method includes the following steps: S101: Acquire the battery operating status data of the target vehicle at the initial sampling frequency and perform rapid preprocessing. Then, divide the preprocessed battery operating status data into an initial input sample set with a preset time step. S102, Substitute the initial input sample set into the hybrid prediction model to obtain the initial two-dimensional features; S103, Based on the target algorithm, fault detection is performed on the initial two-dimensional features to obtain the preliminary fault probability; S104, Based on the initial failure probability, determine the target sampling frequency according to the preset rules, and collect the battery operating status of the target vehicle at the target sampling frequency to obtain updated battery sampling data; S105, the updated battery sampling data and preliminary failure probability are uploaded to the cloud server so that the cloud server can use the preset optimization algorithm to update the parameters of the hybrid prediction model and obtain the optimized parameters. S106. After updating the hybrid prediction model according to the optimized parameters, the updated battery sampling data is substituted into the updated hybrid prediction model to obtain the target two-dimensional features. S107, Based on the target algorithm, perform anomaly detection on the two-dimensional features of the target to obtain the target fault probability, and determine the monitoring scheme according to the target fault probability; Among them, the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps.

[0028] In one implementation, the battery operating status data includes total battery voltage, individual battery voltage, maximum and minimum voltage, total current, charging status, temperature from multiple sensors, and recorded maximum and minimum temperatures. Taking an initial sampling frequency of 0.1Hz as an example, the collected 8-dimensional real-time runtime sequence data is sequentially subjected to forward interpolation of missing values, outlier correction based on the national standard GB / 32960, and wavelet sparse denoising to obtain rapidly preprocessed data. The preset time step is determined by technicians, and taking 2 minutes as an example, several 12-step × 8-dimensional input samples are obtained, with each input sample corresponding to the end time t of the window. Based on the end time t of the window, the true values ​​of the highest single-cell voltage and the highest probe temperature in the next 24 time steps are extracted to form a 24-step × 2-dimensional label matrix. The 8-dimensional features of the input samples and the 2-dimensional target of the label matrix are respectively subjected to Min-Max normalization processing according to their dimensions, mapping the data to the [0,1] interval. The 8-dimensional features and label matrix of the normalized samples are used as the initial input samples.

[0029] In one implementation, when the battery's operating status is determined to be normal and without significant abnormalities, data is always collected at the initial sampling frequency. When an abnormal trend or fault risk is detected in the battery, the sampling frequency is increased accordingly based on the degree of abnormality to collect data more intensively and improve monitoring accuracy. When the abnormality disappears and the status returns to safety in subsequent cyclic detections, the sampling frequency automatically returns to the initial sampling frequency.

[0030] In one implementation, the monitoring scheme is determined based on the target failure probability as follows: if the target failure probability is less than or equal to a preset threshold of 0.4, a normal monitoring instruction is output; if the target failure probability exceeds the preset threshold of 0.4 but does not exceed the danger threshold of 0.8, an early warning monitoring instruction is output; if the target failure probability exceeds the danger threshold of 0.8, an emergency shutdown or fault investigation instruction is output. At the same time, all monitoring data and fault prediction results are stored for subsequent analysis, model optimization, and traceability.

[0031] In one embodiment, the hybrid prediction model includes a TCN feature extraction submodule, an LSTM temporal modeling submodule, and a parallel prediction output layer connected in sequence. The TCN feature extraction submodule consists of a linear embedding layer, three causal dilated convolutional layers, a residual connection layer, a dropout layer, and a linear compression layer, from input to output. The kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. The LSTM time series modeling submodule consists of a first LSTM layer, a second LSTM layer, and a Dropout layer, from input to output. The first LSTM layer has a hidden layer dimension of 128 and is in sequence return mode. The second LSTM layer has a hidden layer dimension of 64 and is in sequence return mode. The dropout rate of the Dropout layer is set to 0.5. The parallel prediction output layer consists of two fully connected layers.

[0032] In one implementation, the linear embedding layer maps the initial 12-step × 8-dimensional input sample to a 12-step × 128-dimensional high-dimensional feature sequence; the kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. During convolution, zeros are padded to the left of the feature sequence to ensure causality, and the output dimension of each layer remains 12-step × 128-dimensional; the residual connection layer adds the output of the linear embedding layer and the output of the three causal dilated convolutional layers across layers. When the dimensions are inconsistent, the dimensions are adjusted by 1×1 convolution before fusion; the dropout rate of the Dropout layer is set to 0.5; and the linear compression layer compresses the fused feature sequence into a 12-step × 64-dimensional long temporal feature sequence.

[0033] In one implementation, the long temporal feature sequence output by the TCN feature extraction submodule is substituted into the LSTM temporal modeling submodule. The LSTM temporal modeling submodule, from input to output, includes a first LSTM layer, a second LSTM layer, and a Dropout layer. The hidden layer dimension of the first LSTM layer is set to 128, and the return sequence mode is enabled. After modeling the long temporal feature sequence, a 12-step × 128-dimensional temporal hidden state sequence is output. The hidden layer dimension of the second LSTM layer is set to 64, the return sequence mode is disabled, and the temporal hidden state sequence is extracted. The hidden state at the last time step is listed, and a 1×64-dimensional global feature vector is output. The dropout rate of the Dropout layer is set to 0.5. The parallel prediction output layer includes two fully connected layers. The first fully connected layer expands the 1×64-dimensional global feature vector into a 1×128-dimensional feature vector. The second fully connected layer maps the 1×128-dimensional feature vector into a 1×48-dimensional feature vector, and then reshapes it into a 24-step × 2-dimensional prediction matrix to obtain the look-ahead prediction values ​​of voltage and temperature. The 2-dimensional values ​​are the look-ahead prediction values ​​of the highest single cell voltage and the highest probe temperature, respectively.

[0034] In one embodiment, see Figure 2 , Figure 2 A flowchart for quantifying battery failure probability is provided in this embodiment of the invention, wherein obtaining the preliminary failure probability by anomaly detection of initial two-dimensional features based on the target algorithm includes: Each element in the initial two-dimensional feature matrix is ​​used as a prediction data point; The distance between each predicted data point and other data points is calculated based on Euclidean distance, and the k-neighborhood of each data point is determined. For each predicted data point, calculate the reachability distance of the predicted data point to every data point in its k-neighborhood, and determine the local reachability density of the predicted data point based on the reachability distance; Based on the local reachability density of each predicted data point and the local reachability density of its corresponding k-neighborhood, calculate the local anomaly factor of the predicted data point. Through formula The local anomaly factor is converted into the initial failure probability; where A is the local anomaly factor, and the initial failure probability is zero when the local anomaly factor is less than or equal to 1.

[0035] In one implementation, the nearest neighbor number k=12 (half of the total number of prediction time steps) is set. The distance between each predicted data point and other data points is calculated based on Euclidean distance to determine the k-neighborhood of each data point. The k-neighborhood is the set of the 12 nearest data points to that data point. For each predicted data point p, the reachable distance of each data point o within its k-neighborhood is calculated. in, For reachability distance, k-distance(o) is the distance from data point o to its k-th nearest neighbor, and d(p,o) is the Euclidean distance between data point p and o. The local reachability density of data point p is calculated based on the reachability distance using the formula... ,in Let p be the number of data points in the k-neighborhood of data point p; for each predicted data point p, calculate its local anomaly factor using the formula... Set the anomaly detection threshold to 1. If If the value is greater than 1, then the data point is considered an outlier. The larger the value, the higher the degree of anomaly of the data point.

[0036] In one implementation, early signs of electric vehicle battery failure (such as a sudden increase in single cell voltage or a sudden rise in local temperature) often manifest as normal global data but outliers in local time steps or dimensions. This embodiment can accurately identify such local outliers, avoiding the problem of missing early failures due to the fixed threshold method. Moreover, compared with the shortcomings of clustering algorithms in being insensitive to sparse outliers and isolated forests ignoring temporal neighborhood correlations, this embodiment can distinguish between occasional interference and real early signs of failure by associating temporally adjacent predicted data points through k-neighborhood.

[0037] In one implementation, by deeply adapting the core logic of the LOF algorithm to the electric vehicle battery monitoring scenario, accurate anomaly detection and fault probability quantification of two-dimensional time-series prediction features of voltage and temperature are achieved, combining detection accuracy, scenario adaptability, engineering practicality, and algorithm application innovation. By relying on the core logic of LOF, which utilizes Euclidean distance to determine k-neighborhood, calculates reachable distance point by point, local reachability density, and local anomaly factors, it accurately captures local outlier features of battery fault precursors, solving the problems of traditional thresholding methods and clustering algorithms failing to detect early local anomalies and having poor adaptability to unsupervised detection. Furthermore, by transforming local anomaly factors into preliminary fault probabilities of 0-1... This breakthrough overcomes the limitation of the general LOF algorithm's binary output, providing a quantifiable and practical core judgment basis for subsequent dynamic sampling frequency adjustment. Simultaneously, the entire detection process is a linear and standardized operation without complex calculations, allowing for lightweight deployment at the edge of the fog layer, meeting the real-time requirements of battery monitoring. Furthermore, the point-by-point detection design covers all data points in the two-dimensional feature matrix, accommodating both single-dimensional and combined anomalies in voltage and temperature. Combined with scenario-customized k-neighborhood settings, the detection accuracy is further improved, effectively distinguishing between occasional interference and true fault precursors. Ultimately, this forms a precise, efficient, and practical anomaly detection link for dynamic battery safety monitoring.

[0038] In one embodiment, determining the target sampling frequency based on a preset rule for the initial failure probability includes: If there exists a predicted data point with a local reachability density greater than 1, then the data point is recorded as an anomaly detection point. The anomaly percentage is obtained by calculating the proportion of anomaly detection points among all predicted data points. If the percentage of abnormalities is greater than the preset threshold, then calculate the preliminary average failure probability of all abnormal detection points. The target sampling frequency is determined based on the initial average failure probability and the preset failure threshold.

[0039] In one implementation, determining the target sampling frequency based on the preliminary average fault probability and a preset fault threshold includes the following steps: Step 1: When a predicted data point has a local reachability density greater than 1, the data point is recorded as an anomaly detection point. Taking a 24-step × 2-dimensional prediction matrix as an example, the total number of anomaly detection points among the 48 predicted data points is counted. The preset anomaly percentage threshold is 15%. If the anomaly percentage is >15%, the preliminary fault probability average of all anomaly detection points (denoted as P_avg) is calculated as the core basis for determining the target sampling frequency. If the anomaly percentage is ≤15%, the initial sampling frequency is maintained at 0.1Hz without frequency adjustment, and the process returns to the preliminary fault prediction process for loop monitoring. Step 2: Based on the fault classification requirements of the battery monitoring scenario, a 3-level fault threshold range is preset (maintaining logical consistency with the subsequent safety monitoring command threshold). The threshold range is divided as follows: slight anomaly range (P_avg∈(0,0.4]), moderate anomaly range (P_avg∈(0.4,0.8]), and high anomaly range (P_avg∈(0.8,1.0]). Each range corresponds to a unique target sampling frequency to achieve accurate matching between the degree of anomaly and the sampling frequency. Step 3: Accurately match the calculated preliminary average fault probability (P_avg) with the preset 3-level fault threshold range mentioned above. After matching, determine the corresponding target sampling frequency. The specific matching rules are as follows: If P_avg∈(0,0.4] (minor anomaly), the corresponding target sampling frequency is 0.12Hz, which is suitable for the slight anomaly trend of the battery. The sampling frequency is moderately increased to control energy consumption while continuously monitoring the changes in the anomaly trend. If P_avg∈(0.4,0.8] (moderate anomaly), the corresponding target sampling frequency is 0.15Hz. For the moderate anomaly trend of the battery, the sampling frequency is significantly increased to capture more detailed data on battery operation anomalies, providing support for subsequent accurate fault determination. If P_avg∈(0.8,1.0] (high anomaly), the corresponding target sampling frequency is 0.2Hz. For the high anomaly risk of the battery, data is collected at the highest sampling frequency to ensure that the battery fault deterioration trend can be captured in real time, providing accurate data support for the generation of emergency response instructions. The initial sampling frequency is 0.1Hz, and the target sampling frequency adjustment range is limited to 0.12Hz~0.2Hz. This avoids the anomaly omission caused by too low a sampling frequency and also prevents energy waste and data redundancy caused by too high a sampling frequency, balancing monitoring accuracy and engineering practicality.

[0040] In one embodiment, updated battery sampling data and preliminary failure probabilities are uploaded to a cloud server, enabling the cloud server to apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain optimized parameters, including: Using updated battery sampling data as model input samples and the corresponding preliminary failure probabilities as supervision labels, a training sample set for model optimization is constructed. Based on the deviation between the predicted values ​​output by the hybrid prediction model and the true values ​​in the updated battery sampling data, a loss function for model training is constructed. By using a pre-defined optimization algorithm, the network coefficients of the hybrid prediction model are iteratively adjusted based on the gradient of the loss function; When the loss function converges to a preset threshold or the number of iterations reaches a preset number, parameter iteration is stopped, and the model parameters that meet the convergence condition after iteration are determined as optimization parameters.

[0041] In one implementation, the loss function is determined by a formula. ; in, The highest single-cell voltage prediction value output by the model. And the actual value of the highest measured single cell voltage in the updated battery sampling data. deviation, The highest probe temperature prediction value output by the model. Compared with the actual highest probe temperature measured in the updated battery sampling data The deviation.

[0042] In one implementation, the network coefficients are the weights and biases of the linear embedding layer, the three-layer causal dilated convolutional layer, the residual connection layer, and the linear compression layer in the TCN feature extraction submodule; the input weights and hidden layer weights of the two LSTM layers in the LSTM temporal modeling submodule; the weights and biases of the two fully connected layers in the parallel prediction output layer; and the preset threshold and preset number of iterations are determined by technical personnel.

[0043] In one implementation, a supervised training sample set containing updated battery sampling data and preliminary fault probabilities is constructed. A loss function is built based solely on the deviation between the predicted and actual voltage and temperature values. A preset optimization algorithm is used to iteratively adjust the model network coefficients. When the loss function converges or a preset number of iterations is reached, the optimization parameters are determined. This approach enables online iterative optimization of the hybrid prediction model using real-world data collected at the edge, continuously improving the model's prediction accuracy for battery voltage and temperature. Simultaneously, it ensures that the model parameter update process is stable, efficient, and easy to implement in engineering, providing a reliable model foundation for more accurate anomaly detection and safety monitoring in the future.

[0044] Based on the same inventive concept, this invention also provides an IoT-based dynamic safety monitoring system for electric vehicle batteries. See also Figure 3 , Figure 3 A framework diagram of an IoT-based dynamic safety monitoring system for electric vehicle batteries, provided for embodiments of the present invention, includes: The data partitioning module is used to acquire battery operating status data collected by the target vehicle at an initial sampling frequency, perform rapid preprocessing, and then partition the preprocessed battery operating status data at a preset time step to obtain an initial input sample set. The initial two-dimensional feature determination module is used to substitute the initial input sample set into the hybrid prediction model to obtain the initial two-dimensional features; the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps. The preliminary fault probability determination module is used to perform fault detection on the initial two-dimensional features based on the target algorithm to obtain the preliminary fault probability. The target sampling frequency determination module is used to determine the target sampling frequency based on preset rules for the initial failure probability, and to collect updated battery sampling data by collecting the battery operating status of the target vehicle at the target sampling frequency. The model optimization parameter determination module is used to upload updated battery sampling data and preliminary failure probability to the cloud server, so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain optimized parameters. The target two-dimensional feature determination module is used to update the hybrid prediction model according to the optimization parameters, and then substitute the updated battery sampling data into the updated hybrid prediction model to obtain the target two-dimensional features. The monitoring scheme generation module is used to perform anomaly detection on the two-dimensional features of the target based on the target algorithm to obtain the target failure probability, and to determine the monitoring scheme based on the target failure probability.

[0045] This invention provides an IoT-based dynamic safety monitoring method for electric vehicle batteries. Through low-frequency initial sampling, rapid preprocessing, and time-series sample construction, a hybrid prediction model outputs two-dimensional look-ahead features of battery voltage and temperature. Combined with a target anomaly detection algorithm, a preliminary fault probability is obtained, enabling adaptive dynamic adjustment of the sampling frequency. Then, based on high-frequency updated sampling data and fault probabilities, model parameters are iteratively optimized in the cloud to further improve prediction and anomaly detection accuracy. Finally, a hierarchical monitoring strategy is formulated based on the precise target fault probability. This method ensures accurate early-stage battery fault identification, effectively reduces missed detections and false positives, and significantly reduces sampling and transmission overhead under normal conditions. It balances real-time monitoring, system reliability, and on-board energy consumption optimization, solving the problem of low efficiency in existing electric vehicle battery safety monitoring due to the dual fixation of parameters and models. This method achieves intelligent battery safety monitoring.

[0046] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things, characterized in that, The method includes: After acquiring battery operating status data of the target vehicle at an initial sampling frequency and performing rapid preprocessing, the preprocessed battery operating status data is divided into an initial input sample set at a preset time step. Substituting the initial input sample set into the hybrid prediction model yields the initial two-dimensional features; the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps. Based on the target algorithm, the initial two-dimensional features are used to perform fault detection to obtain a preliminary fault probability; Based on the initial failure probability, a target sampling frequency is determined according to preset rules, and the battery operating status of the target vehicle is collected at the target sampling frequency to obtain updated battery sampling data; The updated battery sampling data and the preliminary failure probability are uploaded to the cloud server so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain optimized parameters. After updating the hybrid prediction model according to the optimization parameters, the updated battery sampling data is substituted into the updated hybrid prediction model to obtain the target two-dimensional features; Based on the target algorithm, anomaly detection is performed on the two-dimensional features of the target to obtain the target failure probability, and a monitoring scheme is determined based on the target failure probability.

2. The method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things according to claim 1, characterized in that, The hybrid prediction model includes a TCN feature extraction submodule, an LSTM temporal modeling submodule, and a parallel prediction output layer connected in sequence. The TCN feature extraction submodule includes, from input to output, a linear embedding layer, three causal dilated convolutional layers, a residual connection layer, a Dropout layer, and a linear compression layer; the kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. The LSTM time series modeling submodule includes a first LSTM layer, a second LSTM layer, and a Dropout layer from input to output. The hidden layer dimension of the first LSTM layer is 128, and the return sequence mode is enabled. The hidden layer dimension of the second LSTM layer is set to 64, and the return sequence mode is disabled. The dropout rate of the Dropout layer is set to 0.

5. The parallel prediction output layer consists of two fully connected layers.

3. The method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things according to claim 1, characterized in that, Based on the target algorithm, anomaly detection is performed on the initial two-dimensional features to obtain the preliminary fault probability, including: Each element in the initial two-dimensional feature matrix is ​​used as a prediction data point; The distance between each predicted data point and other data points is calculated based on Euclidean distance, and the k-neighborhood of each data point is determined. For each predicted data point, calculate the reachability distance of the predicted data point to every data point in its k-neighborhood, and determine the local reachability density of the predicted data point based on the reachability distance; Based on the local reachability density of each predicted data point and the local reachability density of its corresponding k-neighborhood, calculate the local anomaly factor of the predicted data point. Through formula The local anomaly factor is converted into the initial failure probability; where A is the local anomaly factor, and the initial failure probability is zero when the local anomaly factor is less than or equal to 1.

4. The method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things according to claim 3, characterized in that, Based on preset rules, the target sampling frequency is determined according to the initial failure probability, including: If there exists a predicted data point with a local reachability density greater than 1, then the data point is recorded as an anomaly detection point. The anomaly percentage is obtained by calculating the proportion of anomaly detection points among all predicted data points. If the abnormality percentage is greater than a preset threshold, then calculate the preliminary average failure probability of all abnormal detection points; The target sampling frequency is determined based on the initial average failure probability and the preset failure threshold.

5. The method for dynamic safety monitoring of electric vehicle batteries based on the Internet of Things according to claim 1, characterized in that, The updated battery sampling data and the preliminary failure probability are uploaded to the cloud server so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain optimized parameters, including: Using updated battery sampling data as model input samples and the corresponding preliminary failure probabilities as supervision labels, a training sample set for model optimization is constructed. Based on the deviation between the predicted value output by the hybrid prediction model and the true value in the updated battery sampling data, a loss function for model training is constructed. The network coefficients of the hybrid prediction model are iteratively adjusted based on the gradient of the loss function using a preset optimization algorithm. When the loss function converges to a preset threshold or the number of iterations reaches a preset number, parameter iteration is stopped, and the model parameters that meet the convergence condition after iteration are determined as optimization parameters.

6. A dynamic safety monitoring system for electric vehicle batteries based on the Internet of Things, characterized in that, The system includes: The data partitioning module is used to acquire battery operating status data collected by the target vehicle at an initial sampling frequency, perform rapid preprocessing, and then partition the preprocessed battery operating status data at a preset time step to obtain an initial input sample set. An initial two-dimensional feature determination module is used to substitute the initial input sample set into the hybrid prediction model to obtain initial two-dimensional features; the initial two-dimensional features are the predicted values ​​of the highest single cell voltage and the highest probe temperature at different time steps. The preliminary fault probability determination module is used to perform fault detection on the initial two-dimensional features based on the target algorithm to obtain the preliminary fault probability. The target sampling frequency determination module is used to determine the target sampling frequency based on preset rules for the initial failure probability, and to collect the battery operating status of the target vehicle at the target sampling frequency to obtain updated battery sampling data. The model optimization parameter determination module is used to upload the updated battery sampling data and the preliminary failure probability to the cloud server, so that the cloud server can apply a preset optimization algorithm to update the parameters of the hybrid prediction model and obtain the optimized parameters. The target two-dimensional feature determination module is used to update the hybrid prediction model according to the optimization parameters, and then substitute the updated battery sampling data into the updated hybrid prediction model to obtain the target two-dimensional features. The monitoring scheme generation module is used to perform anomaly detection on the two-dimensional features of the target based on the target algorithm to obtain the target failure probability, and to determine the monitoring scheme based on the target failure probability.

7. The IoT-based dynamic safety monitoring system for electric vehicle batteries according to claim 6, characterized in that, The hybrid prediction model includes a TCN feature extraction submodule, an LSTM temporal modeling submodule, and a parallel prediction output layer connected in sequence. The TCN feature extraction submodule includes, from input to output, a linear embedding layer, three causal dilated convolutional layers, a residual connection layer, a Dropout layer, and a linear compression layer; the kernel size of the three causal dilated convolutional layers is 3, and the dilation factors are set to 1, 2, and 4 respectively. The LSTM time series modeling submodule includes a first LSTM layer, a second LSTM layer, and a Dropout layer from input to output. The hidden layer dimension of the first LSTM layer is 128, and the return sequence mode is enabled. The hidden layer dimension of the second LSTM layer is set to 64, and the return sequence mode is disabled. The dropout rate of the Dropout layer is set to 0.

5. The parallel prediction output layer consists of two fully connected layers.

8. The electric vehicle battery dynamic safety monitoring system based on the Internet of Things according to claim 6, characterized in that, The preliminary fault probability determination module includes: A prediction data point generation module is used to take each element in the initial two-dimensional feature matrix as a prediction data point. The neighborhood partitioning module is used to calculate the distance between each predicted data point and other data points based on Euclidean distance, and to determine the k-neighborhood of each data point; The local reachability density determination module is used to calculate the reachability distance of each predicted data point in the k-neighborhood for each predicted data point, and determine the local reachability density of the predicted data point based on the reachability distance. The local anomaly factor determination module is used to calculate the local anomaly factor of each predicted data point based on the local reachability density of each predicted data point and the local reachability density of its corresponding k-neighborhood. The local anomaly factor transformation module is used to transform factors through formulas. The local anomaly factor is converted into the initial failure probability; where A is the local anomaly factor, and the initial failure probability is zero when the local anomaly factor is less than or equal to 1.

9. The electric vehicle battery dynamic safety monitoring system based on the Internet of Things according to claim 8, characterized in that, The target sampling frequency determination module includes: The anomaly detection point marking module is used to mark a data point as an anomaly detection point when there is a local reachability density greater than 1 corresponding to a predicted data point. The anomaly percentage determination module is used to calculate the percentage of anomaly detection points among all predicted data points to obtain the anomaly percentage value. The preliminary fault probability average determination module is used to calculate the preliminary fault probability average of all abnormal detection points if the abnormality ratio value is greater than a preset threshold. The target sampling frequency generation module is used to determine the target sampling frequency based on the preliminary average fault probability and the preset fault threshold.

10. The IoT-based dynamic safety monitoring system for electric vehicle batteries according to claim 6, characterized in that, The model optimization parameter determination module includes: The training sample set generation module is used to construct a training sample set for model optimization by using updated battery sampling data as model input samples and the corresponding preliminary failure probability as supervision labels. The loss function construction module is used to construct the loss function for model training based on the deviation between the predicted value output by the hybrid prediction model and the true value in the updated battery sampling data. The network coefficient optimization module is used to iteratively adjust the network coefficients of the hybrid prediction model based on the gradient of the loss function using a preset optimization algorithm. The optimization parameter generation module is used to stop parameter iteration when the loss function converges to a preset threshold or the number of iterations reaches a preset number, and to determine the model parameters that meet the convergence conditions after iteration as optimization parameters.

Images