Lithium battery sensor data monitoring system based on optimized RNN (Recurrent Neural Network) and operation method thereof

Abstract

The invention discloses an optimized RNN-based lithium battery sensor data monitoring system and an operation method thereof, and belongs to the technical field of sensor detection, the optimized RNN-based lithium battery sensor data monitoring system comprises a data acquisition and preprocessing module, a model construction module, a model training module, a model deployment module and a continuous learning module, the capturing capability of the model on a time sequence local mode is enhanced through rolling window feature extraction, the precision and generalization capability of the model are improved through an optimized RNN network structure and Dropout regularization, the problem of class imbalance is solved through a loss function based on class weight, the training efficiency and performance are improved through an early stop mechanism and dynamic learning rate adjustment, and the training efficiency and performance are improved. According to the system and the operation method thereof, the precision, the recall rate and the efficiency of sensor data anomaly detection can be remarkably improved, and the system and the operation method thereof have engineering deployment and continuous learning capabilities and are suitable for the fields of industrial Internet of Things, equipment monitoring and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor detection, and particularly to a lithium battery sensor data monitoring system based on optimized RNN and its operation method. Background Art

[0002] The installed capacity of electrochemical energy storage and the market ownership of electric vehicles are increasing at a high speed year by year. New types of electrochemical energy storage batteries such as lithium-ion and sodium-ion are widely used in energy storage systems and electric vehicles due to their advantages of high energy density, light weight, and long life. The sensor data in new types of electrochemical energy storage batteries contains important information on the operating state of equipment. Therefore, real-time and accurate anomaly detection of this data is crucial for ensuring the safe and stable operation of the system, preventing faults, and reducing operating costs.

[0003] Traditional methods for anomaly detection of sensor data, such as threshold methods, statistical methods, and methods based on simple machine learning models (such as SVM, RNN), face many challenges when dealing with complex, high-dimensional, and time-series sensor data:

[0004] Threshold methods and statistical methods: They rely on manual experience to set fixed thresholds, are difficult to adapt to dynamic data environments and complex anomaly patterns, are prone to false alarms and missed detections, and have poor robustness.

[0005] Support Vector Machine (SVM): For high-dimensional time-series data, feature engineering is complex, it is difficult to effectively capture the time-dependent relationships in the data, and it is sensitive to parameters, with limited generalization ability.

[0006] Recurrent Neural Network (RNN): Although it can handle time-series data, it is prone to problems such as vanishing gradients or exploding gradients, is difficult to capture long-term dependencies, has insufficient ability to detect long-term anomaly patterns, and in practical applications, the RNN model often shows a low recall rate for anomaly classes and is difficult to effectively identify abnormal data, especially in the case of scarce anomaly samples.

[0007] In addition, in the scenario of anomaly detection of industrial sensor data, the proportion of abnormal data is usually extremely low, and the data categories are severely imbalanced. Traditional machine learning models tend to be biased towards the majority class (normal class), resulting in severely insufficient recognition ability for the minority class (abnormal class). For example, preliminary experiments show that in the case of an extremely low proportion of anomaly class samples, the recall rate of the anomaly class of an anomaly detection system based on a traditional LSTM model is close to zero, so its practical application value is limited. Therefore, a lithium battery sensor data monitoring system based on optimized RNN and its operation method are proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a lithium battery sensor data monitoring system based on optimized RNN and its operation method to solve the problems raised in the above background art.

[0009] To achieve the above object, the present invention provides the following technical solutions: A lithium battery sensor data monitoring system based on an optimized RNN, including a data acquisition and preprocessing module, a model construction module, a model training module, a model deployment module, and a continuous learning module;

[0010] Among them, the data acquisition and preprocessing module is used to read sensor data, perform data cleaning and format conversion, and for the time-series characteristics of sensor data, a rolling window feature extraction method is used for data preprocessing work;

[0011] The model construction module uses an optimized recurrent neural network model for sensor data anomaly detection;

[0012] The model training module is used to improve the model training efficiency and performance;

[0013] The model deployment module is used to reduce the model volume and inference latency, so as to facilitate deployment on edge devices or cloud platforms to meet the needs of real-time detection in industrial fields;

[0014] The continuous learning module is used to adapt to the dynamic changes in the distribution of sensor data and the emergence of new anomaly patterns.

[0015] As a further preference of this technical solution: The rolling window feature extraction in the data acquisition and preprocessing module includes rolling average and rolling standard deviation. Through the rolling average and rolling standard deviation, noise can be effectively smoothed, data trends can be highlighted, and time-local context information can be captured, thereby enhancing the model's perception ability of short-term time-series patterns, and more effectively identifying sensor data anomalies based on local anomaly patterns.

[0016] As a further preference of this technical solution: The standard rolling window size in rolling window feature extraction is 5, and in specific implementations, the size of the rolling window can be flexibly set according to the characteristics of sensor data and anomaly patterns.

[0017] As a further preference of this technical solution: The optimized recurrent neural network model in the model construction module includes an RNN layer, a feature compression layer, and a classification output layer;

[0018] Among them, the RNN layer: As the core time-series feature extraction layer, it captures the time-dependent relationships and patterns in the original sensor signals;

[0019] The feature compression layer: A dense layer with 16 neurons is used, and the activation function is ReLU, which is used for feature space compression to enhance the model's non-linear expression ability;

[0020] Classification output layer: A single-neuron dense layer is used, and the activation function is Sigmoid, which is used to output the anomaly probability.

[0021] As a further optimization of this technical solution: In the RNN layer, RNN cells are used to balance the feature expression ability and computational efficiency;

[0022] In the feature compression layer, an information bottleneck is formed to force the network to learn key discriminative features, thereby improving the generalization ability of the model;

[0023] In the classification output layer, the Sigmoid activation function constrains the output value within the range of [0, 1], which directly corresponds to the probability of an anomaly occurring, providing an interpretable decision basis.

[0024] As a further optimization of this technical solution: The number of RNN cells is preferably 64 or adjusted according to experiments. The key to the operation effect of the RNN layer lies in selecting the appropriate number of RNN cells and whether to use multiple layers of RNN. If multiple layers are used, the role of each layer needs to be described;

[0025] Among them, in order to enhance the generalization ability of the model and prevent overfitting, the Dropout regularization technique is introduced in the RNN layer to randomly discard neuron connections and improve the robustness of the model.

[0026] As a further optimization of this technical solution: An optimized training strategy is adopted in the model training module, which includes optimizer selection, initial learning rate, batch size, early stopping mechanism, dynamic learning rate adjustment, class weights, and regularization strategy;

[0027] Among them, optimizer selection: The Adam optimizer is adopted;

[0028] Initial learning rate: The initial learning rate is set to 0.001;

[0029] Batch size: The batch size is selected as 64;

[0030] Early stopping mechanism: An early stopping mechanism based on the validation set loss is adopted;

[0031] Dynamic learning rate adjustment: The ReduceLROnPlateau dynamic learning rate adjustment strategy is adopted;

[0032] Class weights: For data class imbalance, class weights are automatically calculated and set according to the sample ratio of the training set;

[0033] Regularization strategy: The Dropout regularization technique is adopted.

[0034] As a further optimization of this technical solution: in the continuous learning module, through the feedback closed-loop mechanism and the Elastic Weight Consolidation strategy, the continuous optimization and performance maintenance of the model are realized.

[0035] The operation method of the lithium battery sensor data monitoring system based on the optimized RNN includes the following steps:

[0036] S1. Collect sensor data;

[0037] S2. Preprocess the sensor data;

[0038] S3. Construct a recurrent neural network model, input the preprocessed data into the anomaly detection model for inference, and obtain the anomaly probability;

[0039] S4. Adopt an optimized training strategy and determine whether to trigger an alarm according to the anomaly probability;

[0040] S5. Construct a continuous learning framework and continuously optimize the model according to the feedback results.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. The purpose of the present invention is to provide a sensor data anomaly monitoring system and method based on rolling window features and an optimized recurrent neural network model, aiming to solve the problems of insufficient accuracy, low recall rate, and poor robustness in the prior art for sensor data anomaly detection. By introducing a series of technical means such as rolling window feature extraction, optimizing the network structure of the RNN, adopting class weights, integrating an early stopping mechanism, and dynamic learning rate adjustment, the performance and efficiency of sensor data anomaly detection are significantly improved, and it has good engineering deployment and continuous learning capabilities, meeting the requirements of industrial applications;

[0043] 2. The present invention has high-precision and high-recall anomaly detection. Through rolling window feature extraction and an optimized RNN model, the accuracy and anomaly class recall rate of sensor data anomaly detection are significantly improved, enabling more effective identification of various anomaly patterns and reducing the false negative rate;

[0044] 3. The present invention can effectively process time series data. The RNN architecture is naturally suitable for processing time series data and can capture the time-dependent relationships in the data, which is crucial for sensor data anomaly detection;

[0045] 4. The present invention effectively enhances the robustness and generalization ability. Dropout regularization and an optimized training strategy improve the robustness and generalization ability of the model, enabling it to adapt to complex and changing industrial environments and reducing the risk of overfitting;

[0046] 5. The present invention has high inference performance and deployment capabilities. The model lightweight deployment solution effectively reduces the model size and inference latency, meets the requirements of industrial real-time detection, and is convenient to be deployed on resource-constrained edge devices;

[0047] 6. The present invention has strong sustainable learning capabilities. The continuous learning framework enables the model to continuously adapt to new data distributions and abnormal patterns, maintaining long-term effective anomaly detection performance;

[0048] 7. The present invention has high engineering application value. The proposed system and its operation method have good engineering application value and can be widely applied to fields such as industrial equipment monitoring, intelligent manufacturing, and environmental monitoring, effectively improving production efficiency, reducing operating costs, and ensuring system security. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the architecture of the lithium battery sensor data monitoring system based on the optimized RNN of the present invention;

[0050] Figure 2 It is a schematic diagram of the flow of the operation method of the lithium battery sensor data monitoring system based on the optimized RNN of the present invention;

[0051] Figure 3 It is a schematic diagram of the architecture of the optimized recurrent neural network model in the lithium battery sensor data monitoring system based on the optimized RNN of the present invention;

[0052] Figure 4 It is a schematic diagram of the architecture of the optimized training strategy in the lithium battery sensor data monitoring system based on the optimized RNN of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0054] Embodiment 1

[0055] Please refer to Figures 1-4 , the present invention provides a technical solution: a lithium battery sensor data monitoring system based on an optimized RNN, including a data acquisition and preprocessing module, a model construction module, a model training module, a model deployment module, and a continuous learning module;

[0056] Among them, the data acquisition and preprocessing module is used to read sensor data, perform data cleaning and format conversion, and perform data preprocessing work by adopting the method of rolling window feature extraction in view of the time series characteristics of sensor data;

[0057] The model construction module adopts an optimized recurrent neural network model for sensor data anomaly detection;

[0058] A model training module, which is used to improve the efficiency and performance of model training;

[0059] A model deployment module, which is used to reduce the model size and inference latency, so as to facilitate deployment on edge devices or cloud platforms to meet the requirements of real-time detection in industrial sites;

[0060] A continuous learning module, which is used to adapt to the dynamic changes of sensor data distribution and the emergence of new abnormal patterns.

[0061] In this embodiment, specifically: the rolling window feature extraction in the data acquisition and preprocessing module includes rolling average and rolling standard deviation. Through the rolling average and rolling standard deviation, noise can be effectively smoothed, data trends can be highlighted, and time-local context information can be captured, thereby enhancing the model's perception ability of short-term time series patterns, and more effectively identifying sensor data anomalies based on local abnormal patterns.

[0062] In this embodiment, specifically: the standard rolling window size in the rolling window feature extraction is 5, and in the specific implementation, the size of the rolling window can be flexibly set according to the characteristics of sensor data and the characteristics of abnormal patterns.

[0063] In this embodiment, specifically: the optimized recurrent neural network model (RNN) in the model construction module includes an RNN layer, a feature compression layer (Dense-16 layer), and a classification output layer (Dense-1 layer);

[0064] Among them, the RNN layer: as the core time series feature extraction layer, it captures the time-dependent relationships and patterns in the original sensor signals;

[0065] The feature compression layer (Dense-16 layer): uses a dense layer with 16 neurons, and the activation function is ReLU, which is used for feature space compression to enhance the non-linear expression ability of the model;

[0066] The classification output layer (Dense-1 layer): uses a single-neuron dense layer, and the activation function is Sigmoid, which is used to output the anomaly probability.

[0067] In this embodiment, specifically: in the RNN layer, RNN units are used to balance the feature expression ability and computational efficiency;

[0068] In the feature compression layer (Dense-16 layer), an information bottleneck is formed to force the network to learn key discriminant features, thereby improving the generalization ability of the model;

[0069] In the classification output layer (Dense-1 layer), the Sigmoid activation function constrains the output value in the interval [0,1], which directly corresponds to the probability of anomaly occurrence and provides an interpretable decision basis.

[0070] In this embodiment, specifically: the number of RNN units is appropriately 64 or adjusted according to experiments. The key to the operation effect of the RNN layer lies in selecting the appropriate number of RNN units and whether to use multiple layers of RNN. If multiple layers are used, the functions of each layer need to be described.

[0071] In this embodiment, specifically: to enhance the generalization ability of the model and prevent overfitting, the Dropout regularization technique is introduced in the RNN layer, randomly discarding neuron connections to improve the robustness of the model.

[0072] In this embodiment, specifically: an optimized training strategy is adopted in the model training module, which includes optimizer selection, initial learning rate, batch size, early stopping mechanism, dynamic learning rate adjustment, class weights, and regularization strategy;

[0073] Among them, optimizer selection: The Adam optimizer is adopted, which has an adaptive learning rate to accelerate model convergence;

[0074] Initial learning rate: The initial learning rate is set to 0.001 to ensure a reasonable initial state of the network;

[0075] Batch size: The batch size is selected as 64 to balance the GPU memory utilization rate and training stability;

[0076] Early stopping mechanism: An early stopping mechanism based on the validation set loss is adopted, and the patience parameter is set to 10 epochs to prevent overfitting;

[0077] Dynamic learning rate adjustment: The ReduceLROnPlateau dynamic learning rate adjustment strategy is adopted to dynamically adjust the learning rate according to the validation set loss to improve model performance;

[0078] Class weights: In response to data class imbalance, class weights are automatically calculated and set according to the training set sample ratio to increase the model's attention to abnormal class samples and improve the abnormal class recognition ability;

[0079] Regularization strategy: The Dropout regularization technique is adopted, and the Dropout rate is set to 0.2 in the RNN layer to prevent overfitting.

[0080] In this embodiment, specifically: the model deployment module includes technologies such as model quantization, layer fusion, and TensorRT acceleration.

[0081] In this embodiment, specifically: in the continuous learning module, through the feedback closed-loop mechanism and the Elastic Weight Consolidation (EWC) strategy, the continuous optimization and performance maintenance of the model are achieved.

[0082] Operation method of a lithium battery sensor data monitoring system based on an optimized RNN, comprising the following steps:

[0083] S1. Collect sensor data;

[0084] S2. Preprocess the sensor data;

[0085] S3. Construct a recurrent neural network model, input the preprocessed data into the anomaly detection model for inference, and obtain the anomaly probability;

[0086] S4. Adopt an optimized training strategy and determine whether to trigger an alarm according to the anomaly probability;

[0087] S5. Construct a continuous learning framework and continuously optimize the model according to the feedback results.

[0088] Embodiment 2

[0089] To more clearly illustrate the technical solution, the following embodiments are implemented in the Python language and the TensorFlow / Keras deep learning framework:

[0090] A1. Data reading and preprocessing:

[0091] In the data preprocessing stage, first read the sensor data, perform data cleaning and format conversion. The key step is rolling window feature extraction. During operation, the rolling average TotalPressure_rolling_mean and rolling standard deviation TotalPressure_rolling_std of the total pressure TotalPressure are calculated. The window size is set to 5. These rolling window features will be used as part of the model input;

[0092] A2. Threshold determination and label generation:

[0093] In this embodiment, the generation of anomaly labels is based on the total pressure threshold. First, according to the preset threshold time point, obtain the total pressure value at this time point as the threshold threshold_total_pressure;

[0094] If the threshold time point does not exist in the data, find the total pressure value at the closest time point as the threshold. Then, mark the data points where the total pressure exceeds the threshold as abnormal (Label = 1), otherwise mark them as normal (Label = 0);

[0095] A3. Dataset division and feature standardization:

[0096] To perform model training and evaluation, divide the dataset into a training set and a test set;

[0097] Stratified sampling stratify=y is adopted to ensure that the proportion of normal and abnormal samples in the training set and test set is the same as that in the original data, thus solving the impact brought by class imbalance;

[0098] Then, StandardScaler is used to standardize the features, eliminate the differences in feature dimensions, and accelerate the convergence of the model;

[0099] Finally, the feature data is reshaped into the 3D input shape (samples, timesteps, features) required by the LSTM model;

[0100] A4. Construct and train an optimized recurrent neural network model (RNN):

[0101] A two-layer LSTM model based on the RNN architecture is constructed, and Dropout regularization dropout=0.2, recurrent_dropout=0.2 is added to the RNN layer to improve the generalization ability of the model;

[0102] The model is compiled using the Adam optimizer and the binary_crossentropy loss function. The key optimization points are the calculation and application of class weights class_weight=class_weight, and the integration of the EarlyStopping and ReduceLROnPlateau callback functions callbacks=[early_stopping, reduce_lr]. These strategies together improve the training efficiency and performance of the model and solve the class imbalance problem;

[0103] A5. Model evaluation and performance metrics:

[0104] In the model evaluation stage, the test set data is used to evaluate the performance of the trained model. During the operation, comprehensive evaluation metrics such as accuracy, precision, recall, F1-score, AUC value, and confusion matrix are calculated to evaluate the model performance from multiple perspectives;

[0105] Among them, special attention is paid to the recall rate and AUC value to evaluate the model's ability to identify abnormal samples;

[0106] A6. Model saving and deployment:

[0107] The trained model and the StandardScaler object are saved to local files for convenient subsequent model deployment and application;

[0108] In the model deployment stage, techniques such as model quantization, layer fusion, and TensorRT acceleration can be adopted to further optimize the model performance and deployment efficiency, and build a real-time anomaly detection system;

[0109] The construction of the continuous learning framework needs to be designed and implemented according to the actual application scenarios and data characteristics.

[0110] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lithium battery sensor data monitoring system based on optimized RNN, characterized by: It includes data collection and preprocessing module, model building module, model training module, model deployment module and continuous learning module; The data acquisition and preprocessing module is used to read sensor data, perform data cleaning and format conversion, and use rolling window feature extraction to perform data preprocessing in view of the temporal characteristics of sensor data. The model building module uses an optimized recurrent neural network model for sensor data anomaly detection; The model training module is used to improve the model training efficiency and performance; The model deployment module is used to reduce the model size and inference latency, so as to facilitate deployment on edge devices or cloud platforms to meet the needs of real-time detection in industrial sites; The continuous learning module is used to adapt to the dynamic changes in sensor data distribution and the emergence of new abnormal patterns.

2. The lithium battery sensor data monitoring system based on optimized RNN according to claim 1 is characterized in that: The rolling window feature extraction in the data acquisition and preprocessing module includes a rolling mean and a rolling standard deviation. The rolling mean and the rolling standard deviation can effectively smooth noise, highlight data trends, and capture local context information, thereby enhancing the model's perception of short-term time series patterns, thereby more effectively identifying sensor data anomalies based on local abnormal patterns.

3. The lithium battery sensor data monitoring system based on optimized RNN according to claim 1, characterized in that: The optimized recurrent neural network model in the model building module includes an RNN layer, a feature compression layer and a classification output layer; Among them, the RNN layer: as the core temporal feature extraction layer, captures the temporal dependencies and patterns in the original sensor signals; Feature compression layer: A dense layer with 16 neurons and ReLU activation function is used to compress the feature space and enhance the nonlinear expression ability of the model; Classification output layer: A single neuron dense layer is used with Sigmoid as the activation function to output abnormal probability.

4. The lithium battery sensor data monitoring system based on optimized RNN according to claim 3 is characterized in that: In the RNN layer, RNN units are used to balance feature expression capabilities and computational efficiency; In the feature compression layer, an information bottleneck is formed, forcing the network to learn key discriminative features, thereby improving the generalization ability of the model; In the classification output layer, the Sigmoid activation function constrains the output value to the interval [0,1], which directly corresponds to the probability of anomaly occurrence and provides an explainable decision basis.

5. The lithium battery sensor data monitoring system based on optimized RNN according to claim 1, characterized in that: The model training module adopts an optimized training strategy, which includes optimizer selection, initial learning rate, batch size, early stopping mechanism, dynamic learning rate adjustment, class weight and regularization strategy; Among them, the optimizer selection: Adam optimizer is used; Initial learning rate: Set the initial learning rate to 0.001; Batch size: Select a batch size of 64; Early stopping mechanism: adopts an early stopping mechanism based on the validation set loss; Dynamic learning rate adjustment: Use ReduceLROnPlateau dynamic learning rate adjustment strategy; Category weight: To address data category imbalance, the category weight is automatically calculated and set based on the sample ratio of the training set; Regularization strategy: Use Dropout regularization technology.

6. The lithium battery sensor data monitoring system based on optimized RNN according to claim 1, characterized in that: The continuous learning module achieves continuous optimization and performance maintenance of the model through a feedback closed-loop mechanism and an Elastic Weight Consolidation strategy.

7. The operation method of the lithium battery sensor data monitoring system based on the optimized RNN is characterized in that: The following steps are involved: S1, collect sensor data; S2, preprocessing sensor data; S3, build a recurrent neural network model, input the preprocessed data into the anomaly detection model for reasoning, and obtain the anomaly probability; S4, adopting an optimized training strategy and determining whether to trigger an alarm based on the abnormal probability; S5. Build a continuous learning framework and continuously optimize the model based on feedback results.

Images