Design of wireless battery management system monitoring and automated alarm system based on improved long short-term memory neural network

Network model construction based on LSTM

Neither the traditional Kalman filter for SOC estimation nor the support vector machine regression method considers that SOC, as a battery state, cannot change and often depends on the previous state of charge. As a variant of the recurrent neural network, the LSTM model can effectively deal with nonlinear time series problems and have long-term memory ability. LSTM network has a forgetting gate, an input gate and an output gate, which can input valid or invalid information to its kernel state unit through the threshold.

First, by converting the voltage, temperature, current and estimated SOC characteristics of the single battery in the previous state into time series characteristic data, a SOC prediction network based on the short-term memory network LSTM is shown in Fig. 2 is constructed and realized.

According to the operating state of the energy storage battery, such as voltage, current, temperature, etc., collected by BMS in real-time and superimposed with the time characteristics of a moment, an input data format conforming to the timing characteristics is constructed. The size of the matrix is 1*7. After the input data is input into the LSTM network, it first passes through a layer of the LSTM network, and the activation function is RELU. After the dropout layer, it is learned by a layer of LSTM, Then connect the 1*50 dimension fully connected neural network, and through the learning and dimension reduction of the following layers, finally obtain the 1-dimension output value, that is, the predictive value of the energy storage battery SOC.

Optimization and improvement based on batch normalization algorithm

Although the method of random gradient descent has simplified the training depth network, it still has some drawbacks, such as more severe dependence on the initial set learning rate, parameter initialization, weight attenuation coefficient, dropout ratio and other parameters, so this topic uses BN (batch normalization) algorithm to optimize it.

As shown in Fig. 3, the BN algorithm first normalizes the dimensions output after each full connection layer. (1)${\displaystyle {\hat{x}}^{\left(k\right)}=\frac{x^k-E\left[x^k\right]}{\sqrt{var\left[x^k\right]}}\left(1\right)}$

After the above normalization, to prevent the distribution shift of features, it is necessary to transform and reconstruct the normalized data to restore the original features. (2)${\displaystyle y^k={\gamma }^k{\hat{x}}^k+{\beta }^k\left(2\right)}$

When ${\gamma }^{\mathrm{k}}=\sqrt{\mathrm{var}\left[{\mathrm{x}}^{\mathrm{k}}\right]},{\beta }^{\mathrm{k}}=\mathrm{E}\left[{\mathrm{x}}^{\mathrm{k}}\right]$, the transformed and reconstructed features are consistent with the original features, by learning two parameters, the network realizes batch normalization, solves the problems of gradient disappearance and gradient explosion, and increases the robustness and error performance of the model.

Model optimization of network structure combined with Denoising AutoEncoder

To enhance the model’s robustness and stability under complex working conditions, this study proposes to use the Denoising AutoEncoder (DAE) to extract features and denoise the input data. This algorithm is based on the input data to denoise and extract features. The extracted features do not discard some features but fuse and represent them as high-level abstract low-dimensional features that reflect the nature of the data and have more robustness and generalization capabilities.

DAE adds a certain amount of noise to the input data and then realizes the noise removal function through the degradation process in the learning process. The DAE obtains pure noise-free samples for input data with particular noise after encoding, decoding and other operations.

Assume that the input sample data is x ∈ R^n×l, and DAE randomly sets it to 0 according to a certain proportion, v, to obtain the degenerated input eigenvector $\overline{\mathrm{x}}$. Code them based on degenerate input samples based on linear mapping and the nonlinear activation function to complete the coding phase. (3)${\displaystyle \mathrm{h}=\mathrm{g}\left(\mathrm{W}\overline{\mathrm{x}}+{\mathrm{b}}_1\right)}$

Next, the decoder decodes the encoded feature data to obtain the optimal abstract representation of the input data and complete the decoding phase. (4)${\displaystyle \mathrm{z}=\mathrm{g}\left({\mathrm{W}}^{\mathrm{T}}\mathrm{h}+{\mathrm{b}}_2\right)}$

In the above formula, z is the reconstruction value of the input data; b₂ is the node offset term of the decoding layer; g(⋅) is the activation function set by the node.

X and Z are the input training data and the output reconstruction data, respectively. Then the reconstruction error function between the reconstructed data samples and the original input data is: (5)${\displaystyle \mathrm{L}\left(\mathrm{X},\mathrm{Z}\right)=\frac{1}{2}\sum _{\mathrm{i}=1}^{\mathrm{m}}{∥{\mathrm{z}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}}∥}_2^2}$

Implementation of the business layer

The business layer of the BMS monitoring and alarm system is mainly responsible for realizing the interaction and request distribution among the system modules, distributing the algorithm business to the algorithm layer, and realizing the corresponding functions by interacting with the data layer modules. This study uses the Spring MVC framework to learn low-coupling system development. Spring Boot is used for automatic configuration, which assists in simplifying the steps of project construction, and uses the Tomcat server as a lightweight application server.

The business layer of this system uses the Spring MVC framework to realize the construction of the business model. It implements a request-driven web framework through the MVC design pattern implemented by Java.

Implementation of algorithm layer

The algorithm layer preprocesses the request message to extract valuable data features. This system uses Tornado lightweight framework to complete. This technology’s asynchronous, non-blocking IO processing method can use the HTTP protocol to pass parameters to the server, including query strings and data sent in the request body (form data, JSON, XML, etc.), URL regular expression interception, etc. The parameters are preprocessed and transmitted to TensorFlow Serving. The TensorFlow Serving system is well suited to serving multiple models that can dynamically change based on real-world data. It presents a solution to apply the model to actual production. After receiving the parameter data preprocessed by Tornado, this study needs to train and deploy the deep learning algorithm implemented in Chapter 3 to estimate the SOC of the energy storage battery.

The TensorFlow Serving model is mainly used for deploying deep learning algorithm modules, a service system with flexibility, high performance, and use in production environments. New algorithms can be easily deployed through this service system, and multi-model services and distributed APIs can be provided simultaneously. Using the continuous integration framework based on TensorFlow Serving in the algorithm layer is mainly divided into three steps: (1) Model training: including data collection and cleaning, model training, evaluation and optimization; (2) Model online: import the model trained in the previous step into TensorFlow Serving and go online; (3) Service usage: The client communicates with the TensorFlow Serving side through gRPC and RESTful API and obtains the service.

For TensorFlow serving multi-model deployment, you only need to modify the model configuration file to serve the new model. Config, add model_version_policy, place the model file of the new model in the corresponding path, and then start the corresponding service.

Implementation of the communication layer

Usually, Socket communication consists of two parts: the server side and the client side. The server and client sides can send and receive data from each other by establishing a network connection. The specific process is shown in Fig. 4.

The Server Socket class is the primary mechanism for implementing server-side TCP/IP communication programs. When a Server Socket object is created, the program provides services on the specified port on the server side and starts listening for possible service requests from clients. When a client requests a connection and is accepted, the server program will create a Socket object to connect with the remote client to implement read-and-write communication operations. The accept() of this class is used to wait and accept a connection from the client. When accept() is called, the server process or thread will be blocked until it listens to a client process that makes a service request and succeeds with its connection. Then accept() will return a newly created server-side Socket object.

Implementation of the data layer

The data layer is mainly responsible for storing the real-time operation status parameters of the BMS monitoring and alarm system, alarm warning information, user logs and other file data. This research uses the MySQL database management system to save the state parameters of different single cells of the battery pack in separate tables, which increases the operating speed and flexibility of the system. At the same time, the database is mapped and configured using the MyBatis framework under the Spring framework.

This research uses the MyBatis framework to automatically generate the SQL statement that meets the needs by providing the MyBatis mapping method. It can input parameters into the Prepared Statement to achieve automatic input mapping. The query result set can also be mapped to Java objects to meet the output requirements map. Figure 5 displays the workflow. The SqlSession Factory Builder is used by the data layer’s MyBatis application to create the SqlSession Factory from the basis-config.xml configuration file. Then, create a SqlSession instance using a SqlSession Factory instance that was previously constructed. The final step involves making a Mapper object with the SqlSession instance established in the previous step and using the SQL database operation statement that the object has been assigned to complete database operations, such as transaction commit operations or CRUD row operations. Finally, close the SqlSession instance after the procedure is complete.

Performance analysis of battery SOC estimation algorithm

This study first completed the training based on the extended short-term memory network on the preprocessed data set and obtained the training model LSTM-origin. Afterward, based on batch normalization, the learning and convergence capabilities of the deep learning network were improved and optimized. The model LSTM-BN was obtained by retraining. Finally, the DAE compression autoencoder is used to optimize the generalization ability and robustness of the model, and the model LSTM-DAE is obtained by retraining. Models after training using the same training set are tested for performance metrics on the same tests.

The three models—the original LSTM-origin model, the improved LSTM-BN model, and the LSTM-DAE model—can be trained and tested using the same data set, and by using the same initial and iteration round numbers, it is possible to compare the number of iterations and error reduction of each model in the training and test sets. In Fig. 6, this is displayed. It can be seen from Fig. 6 that under the same training set and test set, the loss functions of the three model algorithms designed in this study for the SOC of the energy storage battery tend to be stable after multiple rounds of iterations. Because there is no longer a need to rely on the network initialization settings and learning rate, the training set and test set errors based on the original long short-term memory network are at their highest positions among the three models when the error tends to be stable. The iteration speed of the LSTM-BN model is substantially faster than that of the LSTM-origin model because of the model’s increased iteration efficiency, which also results in a decrease in error. The LSTM-BN technique is then further optimized using the DAE algorithm. The error is decreased to the lowest level among the three models with additional feature extraction and denoising optimization, and the fluctuation is also reduced to increase the model’s stability. The model’s resilience and generalizability are also gradually increased due to its optimization.

After the training set is finished and the test set is anticipated, Fig. 7 compares the accurate SOC with the SOC predicted by the model for the three models created in this study. Due to the significant number of samples in the test set, 115 sample points were chosen for this image to illustrate the true SOC value and the SOC value projected by each model to visually and unmistakably demonstrate the model’s predictive power. The expected value of the model’s LSTM origin may be seen in the Fig. 7 to have a similar trend to the actual value of SOC, but there is still a distinct inaccuracy. Between them, there is a substantial amount of variability. Last but not least, the DAE algorithm-optimized LSTM-DAE model can more closely track the real value with the least amount of error. Because of the sharply decreased fluctuation, the LSTM-DAE model provides the best resilience and predictive power. Figure 8 compares the values of the four evaluation indicators of the three models designed in this study (including mean absolute error, maximum error, mean square error, and R square). It can be seen that the performance indicators of the LSTM-DAE model are significantly better than other models, and the prediction accuracy and robustness have been substantially improved.

Function test of wireless BMS monitoring alarm system

This research tests and verifies the functions of the wireless BMS monitoring and alarm system to ensure the BMS’s stable, reliable and accurate operation. The test plan mainly includes the safety performance test, system working condition test and routine performance test of the battery management system.

Safety performance test

The safety performance test of the battery management system is mainly aimed at the simulation test of insulation resistance. According to GB/T 18384-2015 Safety Requirements for Electric Vehicles, the insulation resistance of the BMS system shall not be less than 500 Ω/V (General Administration of Quality Supervision, 2015). This research monitors the change of insulation resistance by testing the insulation performance of the positive and negative poles of the high-voltage bus of the battery system. The results in Fig. 9 show that the system has high safety performance.

System working condition test

Through the SOC estimation test, state quantity acquisition accuracy test, and fault diagnostic test, this research evaluates the system’s operating condition performance in an effort to decrease failure rates, enhance product quality and performance, and accelerate the development cycle of BMS. Figure 10 shows the SOC estimation test and SOC error chart. The test results show that when the SOC is>80%, the SOC estimation error can be maintained within 4%. When the SOC is<80%, the estimation error can be maintained within 5%, meeting the relevant requirements of the BMS design specification.

Routine performance test

Static and discharge equalization tests are the critical components of the BMS hardware’s standard performance test. After the BMS is turned on, without starting the car model, static equalization sets various initial battery voltages and waits for automatic equalization. The BMS is configured to perform discharge testing under constant working conditions thanks to discharge equalization. Due to the equalization approach, the battery management system will carry out discharge equalization. This investigation chooses twelve batteries in the same BMS port for testing. In Fig. 11, the equalization image is displayed. Figure 11 shows the static equalization test and discharge equalization test. According to the test results, the voltage difference between each cell decreases with time, totally realizing the equalization function test and satisfying the equalization requirements.