Driving system and method of electric vehicle battery relay based on deep learning

Abstract

The invention discloses a driving system and method of an electric vehicle battery relay based on deep learning, and relates to the technical field of data processing, real-time operation data of the electric vehicle battery relay are obtained, time-frequency domain features of the battery relay are extracted from the real-time operation data through a pre-trained time-frequency domain joint feature extraction network, and the time-frequency domain features of the battery relay are extracted; generating probability distribution of an electromagnetic interference state based on a pre-trained state judgment model, if it is judged that the relay is in a weak electromagnetic interference environment, dynamically adjusting time sequence parameters of a relay driving signal through a reinforcement learning model, and if it is judged that the relay is in a strong electromagnetic interference environment, dynamically adjusting time sequence parameters of the relay driving signal according to time-frequency domain characteristics. Predicting a multi-dimensional contact wear state of the relay through a deep residual diagnosis network, and generating an optimization parameter of a relay driving signal according to the contact wear state; and the intelligence and the self-adaptation of relay driving are realized.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a driving system and method for an electric vehicle battery relay based on deep learning. Background Art

[0002] In the battery system of electric vehicles, relays, as key electrical switching components, are responsible for controlling the on and off of high-voltage circuits. Their working status directly affects the safe operation of the battery system.

[0003] In the actual operating environment of electric vehicles, battery relays often face complex and changing electromagnetic environments. During driving, the start-stop of high-power devices such as the engine, motor, and air conditioning compressor, as well as external electromagnetic interference, can have varying degrees of impact on the normal operation of the relay. Especially in environments with strong electromagnetic interference, relay contacts may experience jitter, adhesion, or abnormal wear. In severe cases, this can even lead to relay failure, causing battery system failure or safety accidents.

[0004] Therefore, during the operation of electric vehicle battery relays, real-time monitoring and judgment of electromagnetic interference status and adoption of corresponding driving strategies according to different interference environments are of great significance for ensuring the reliable operation of the relays, extending their service life and improving the safety of the entire battery system. Summary of the Invention

[0005] The present invention aims to address at least one of the technical problems existing in the prior art. To this end, the present invention proposes a deep learning-based drive system and method for an electric vehicle battery relay. This system can adjust the drive strategy in real time based on environmental changes and relay status, achieving intelligent and adaptive relay drive.

[0006] To achieve the above objectives, a driving method for an electric vehicle battery relay based on deep learning is proposed, which includes the following steps: Step 1: Obtain real-time operating data of the electric vehicle battery relay; Step 2: extracting the time-frequency domain features of the battery relay from the real-time operation data using a pre-trained time-frequency domain joint feature extraction network; Step 3: Input the time-frequency domain features into a pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state. Based on the electromagnetic interference state, determine whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment; Step 4: If the relay is in a strong electromagnetic interference environment, the multi-dimensional contact wear status of the relay is predicted using a deep residual diagnosis network based on the time-frequency domain characteristics. Step 5: generating optimized parameters of the relay drive signal according to the contact wear state; After determining whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment, the method further includes: If the current operating state is in the weak electromagnetic interference environment, dynamically adjusting the timing parameters of the relay drive signal through the reinforcement learning model; The real-time operation data includes current signals, voltage signals and temperature signals.

[0007] Among them, the current signal includes the main circuit current and the coil current; the voltage signal includes the main contact voltage and the coil voltage; the temperature signal includes the contact temperature, the coil temperature and the ambient temperature.

[0008] The extracting time-frequency domain features of the battery relay from the real-time running data through the pre-trained time-frequency domain joint feature extraction network comprises the following steps: Step 21: Step 21: Organize the main circuit current, coil current, main contact voltage, coil voltage, contact temperature, coil temperature and ambient temperature in the real-time operation data into time series data of 7 signal channels; Step 22: Perform frequency domain transformation on the time series data of each signal channel by combining short-time Fourier transform and wavelet transform, extract spectrum statistical features, and form a time-frequency spectrum diagram; Step 23: Extract and fuse the time series features and frequency domain features of the time series data and the time-frequency spectrum through a pre-trained time-frequency domain joint feature extraction network to generate a comprehensive time-frequency domain feature representation of the relay; the time-frequency domain joint feature extraction network adopts a multimodal fusion architecture with an attention mechanism. The time-frequency domain joint feature extraction network includes three parts: a time domain branch, a frequency domain branch, and a fusion module; the extraction and fusion of the time series features and frequency domain features of the time series data and the time-frequency spectrum through a pre-trained time-frequency domain joint feature extraction network to generate a time-frequency domain feature representation of the relay includes: Through the time domain branch, a multi-layer bidirectional long short-term memory network is used to extract time domain features from the time series data to obtain multi-channel time series features; The frequency domain branch is used to process the time-frequency spectrum using a two-dimensional convolutional neural network, and the output of the two-dimensional convolutional neural network is converted into a multi-dimensional frequency domain feature vector through global average pooling and a fully connected layer; Through the fusion module, the multi-channel time series features and the frequency domain feature vectors are enhanced and spliced ​​using a cross-attention mechanism, and the output results after the enhancement and splicing are subjected to multi-layer perceptron dimensionality reduction to generate a multi-dimensional joint time-frequency domain feature vector as the time-frequency domain feature representation of the relay.

[0009] Inputting the time-frequency domain features into a pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state includes the following steps: Step 31: Construct a state judgment model combining a multi-layer perceptron and an attention mechanism, process the time-frequency domain features through the state judgment model and generate the electromagnetic interference state; Step 32: Constructing a sample data set of electromagnetic interference status, wherein the sample data set includes relay operation data under weak electromagnetic interference and strong electromagnetic interference environments; Step 33: Based on the sample data set, a batch gradient descent algorithm is used to train the state judgment model to obtain a trained state judgment model; Step 34: Input the time-frequency domain features extracted in real time into the trained state judgment model to generate a probability distribution of the electromagnetic interference state; The method of judging whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment according to the electromagnetic interference state is: Calculating an electromagnetic interference intensity index based on a probability distribution of the electromagnetic interference state, and determining whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment based on the electromagnetic interference intensity index; The electromagnetic interference intensity index is calculated by weighted summation.

[0010] The method of dynamically adjusting the timing parameters of the relay drive signal by using a reinforcement learning model includes the following steps: Step 41: Construct a time series parameter space model of the relay drive signal for the reinforcement learning model, and define a state space, an action space, and a reward function, wherein the state space is used to indicate the electromagnetic interference state and the current working state of the battery relay, the action space is used to indicate the set of time series parameters of the relay drive signal, and the reward function is used to reward or punish based on the performance of the relay; Step 42: Build the policy network and value network of the reinforcement learning model based on the deep deterministic policy gradient algorithm; Step 43: Build an experience replay buffer for the reinforcement learning model, and store the interaction data during the reinforcement learning model training and application process through a priority sampling mechanism; Step 44: Implementing a gradient descent-based network parameter update algorithm for the reinforcement learning model, optimizing the policy network and the value network, and completing the construction of the reinforcement learning model; Step 45: Build a laboratory test platform to simulate the dynamic response characteristics of the relay under electromagnetic interference conditions to train the reinforcement learning model; Step 46: Using a phased training strategy, execute the reinforcement learning model training process, and optimize the timing parameter strategy of the relay drive signal by interacting with the laboratory test platform to obtain a trained reinforcement learning model; Step 47: Based on the current electromagnetic interference state and the current working state of the relay, the timing parameters of the relay drive signal adapted to the current electromagnetic interference state are generated in real time by the trained reinforcement learning model; The method of predicting the multi-dimensional contact wear state of a relay by using a deep residual diagnosis network based on time-frequency domain characteristics includes the following steps: Step 51: constructing a network structure of a deep residual diagnosis network, wherein the deep residual diagnosis network includes a feature input layer, a residual block stacking layer, and a contact wear state output layer; Step 52: Collect sample data sets of relay contacts at different wear levels, wherein each data in the sample data set includes the time-frequency domain features of the relay and the corresponding contact wear state label Step 53: Combining regression loss and classification loss, designing a loss function of a deep residual diagnosis network including the regression loss function and the classification loss function to optimize the prediction accuracy of the contact wear state; Step 54: Based on the sample training set, a phased training strategy and a learning rate scheduling mechanism are adopted to train the deep residual diagnosis network to obtain a trained deep residual diagnosis network; Step 56: Based on the trained deep residual diagnosis network, according to the time-frequency domain features, predict the contact wear state of the relay in real time and generate a wear feature representation; Generating optimized parameters of the relay drive signal according to the contact wear state comprises the following steps: For each dimension of the contact wear state, based on the physical characteristics of the dimension, a preset optimization strategy is used to generate the corresponding optimized parameters of the driving signal; The deep learning-based driving system for an electric vehicle battery relay includes a real-time data collection module, a time-frequency domain feature extraction module, an interference environment judgment module, a weak interference parameter optimization module, and a strong interference parameter optimization module; wherein each module is electrically connected; A real-time data collection module acquires the real-time operating data of the electric vehicle battery relay and sends the real-time operating data to the time-frequency domain feature extraction module; A time-frequency domain feature extraction module extracts the time-frequency domain features of the battery relay using the real-time operation data of the pre-trained time-frequency domain joint feature extraction network, and sends the time-frequency domain features to the interference environment judgment module and the strong interference parameter optimization module; The interference environment judgment module inputs the time-frequency domain features into the pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state. Based on the electromagnetic interference state, it determines whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment. If it is in a strong electromagnetic interference environment, it switches to the strong interference parameter optimization module; The strong interference parameter optimization module predicts the multi-dimensional contact wear status of the relay based on the time-frequency domain characteristics through a deep residual diagnosis network, and generates the optimized parameters of the relay drive signal based on the contact wear status.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The present invention first collects the operating data of the battery relay in real time, uses a pre-trained time-frequency domain joint feature extraction network to process the collected data, and simultaneously extracts time domain and frequency domain features, so as to more comprehensively characterize the working state of the relay. The extracted time-frequency domain features are then input into a pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state, and based on this, it is judged whether the current environment is weak electromagnetic interference or strong electromagnetic interference. In a weak electromagnetic interference environment, the reinforcement learning model is started. When it is judged to be in a strong electromagnetic interference environment, the deep residual diagnosis network is started, and then the optimized parameters of the relay drive signal are generated according to the predicted contact wear state. Therefore, the drive strategy can be adjusted in real time according to environmental changes and relay status, realizing the intelligent and adaptive relay drive. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of a method for driving an electric vehicle battery relay based on deep learning in Example 1 of the present invention; Figure 2 This is a module connection diagram of the deep learning-based electric vehicle battery relay drive system in Example 2 of the present invention. DETAILED DESCRIPTION

[0013] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0014] Example 1 like Figure 1 As shown, the driving method of the electric vehicle battery relay based on deep learning includes the following steps: Step 1: Obtain real-time operating data of the electric vehicle battery relay; Step 2: extracting the time-frequency domain features of the battery relay from the real-time operation data using a pre-trained time-frequency domain joint feature extraction network; Step 3: Input the time-frequency domain features into a pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state. Based on the electromagnetic interference state, determine whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment; Step 4: If the relay is in a strong electromagnetic interference environment, the multi-dimensional contact wear status of the relay is predicted using a deep residual diagnosis network based on the time-frequency domain characteristics. Step 5: Generate optimized parameters of the relay drive signal according to the contact wear state.

[0015] In a further real-time example of the present invention, after determining whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment, the method further includes: If the current operating state is in the weak electromagnetic interference environment, dynamically adjusting the timing parameters of the relay drive signal through the reinforcement learning model; The real-time operation data includes current signals, voltage signals and temperature signals.

[0016] Current signals include the main circuit current and coil current; voltage signals include the main contact voltage and coil voltage; and temperature signals include contact temperature, coil temperature, and ambient temperature. A high-precision sensor array is used to achieve multi-channel synchronous acquisition, with a sampling frequency of no less than 10kHz.

[0017] Specifically, in the process of collecting real-time operating data of electric vehicle battery relays, a multimodal sensor network architecture is adopted. The multimodal sensor network architecture consists of three types of sensors: current sensor, voltage sensor and temperature sensor, which collect current signals, voltage signals and temperature signals respectively.

[0018] During the current signal acquisition phase, a dual measurement solution combining a Hall-effect current sensor and a Rogowski coil is employed. The main loop current is measured using a Hall-effect sensor with a range of ±500A, boasting linearity better than 0.1% and a temperature drift coefficient less than 50ppm / °C. The coil current is measured using a miniature Rogowski coil with an accuracy of 0.5% to capture even small changes in the coil current.

[0019] During the voltage signal acquisition phase, the main contact voltage is isolated and measured using a high-voltage differential probe with a ±1000V range and an input impedance greater than 10MΩ to ensure that the voltage measurement process does not interfere with the normal operation of the relay. The coil voltage is measured using a precision resistor divider network and an operational amplifier for signal conditioning. The voltage signal is sampled synchronously with the current signal at the same acquisition frequency to eliminate phase differences between channels and ensure synchronized timestamps between the voltage and current signals.

[0020] During the temperature signal acquisition phase, the contact temperature and coil temperature are measured using a miniature PT100 platinum resistance temperature sensor with a temperature measurement range of -50°C to 200°C. The temperature sensor is fixed to the key part of the relay with heat conduction glue to ensure the efficiency of temperature transfer. The second ambient temperature is measured using a digital temperature and humidity sensor, which also monitors the ambient humidity to provide auxiliary parameters for subsequent analysis. Similarly, the temperature signal is sampled synchronously with the current signal and voltage signal using the same acquisition frequency to ensure the time synchronization of the temperature signal with the current signal and voltage signal. Furthermore, the extracting of time-frequency domain features of the battery relay from the real-time operation data using the pre-trained time-frequency domain joint feature extraction network includes the following steps: Step 21: Organize the main circuit current, coil current, main contact voltage, coil voltage, contact temperature, coil temperature and ambient temperature in the real-time operation data into time series data of 7 signal channels.

[0021] Specifically, the current signals of the main circuit and coil currents, the voltage signals of the main contact and coil voltages, and the temperature signals of the contact, coil, and ambient temperatures are organized into seven channels of time series data, with each channel corresponding to a specific signal. The time series data, consisting of chronologically ordered real-time operating data, is then segmented using a fixed-length sliding window with a window length of 1000 samples, corresponding to a 100ms time period at a 10kHz sampling rate, and a 50% overlap between adjacent windows.

[0022] Furthermore, the multi-layer bidirectional long short-term memory network comprises a three-layer network structure, with each layer containing 128 hidden units. The first layer receives the original 7-channel time series data and extracts low-level time series features; the second layer further captures mid-level time series dependencies; and the third layer outputs high-level time series abstract representations. To improve the robustness of the network, each layer can be followed by a dropout layer with a dropout rate set to 0.2. Ultimately, the multi-layer bidirectional long short-term memory network outputs a 256-dimensional time-domain feature vector as a multi-channel time series feature, encoding the temporal dynamics of the relay's operating state.

[0023] Step 22: Use a method combining short-time Fourier transform and wavelet transform to perform frequency domain transform on the time series data of each signal channel, extract frequency domain statistical features, and form a time-frequency spectrum diagram.

[0024] Specifically, for each signal channel, a short-time Fourier transform (STFT) was first applied with a 256-point window length, a 64-point step size, and a Hanning window function to generate a time-frequency spectrum. The STFT parameter settings ensured a balance between time and frequency resolution, enabling the capture of frequency features in the 10 Hz to 5 kHz range.

[0025] At the same time, a continuous wavelet transform is applied to each signal channel, using the Morlet wavelet as the mother wavelet. The scale range is set to 1 to 64, with a total of 32 scales, corresponding to a frequency range of 10 Hz to 5 kHz. As can be understood, the wavelet transform provides multi-resolution analysis capabilities, with high frequency resolution in low-frequency bands and high time resolution in high-frequency bands, making it particularly suitable for capturing transient characteristics during relay operation.

[0026] The time-frequency representation generated by the short-time Fourier transform and wavelet transform is converted into a logarithmic energy spectrum to enhance the signal's dynamic range. For each signal channel, spectral statistical features are extracted, including 20 features such as energy per frequency band, dominant frequency, spectral centroid, spectral bandwidth, and spectral entropy, forming a 140-dimensional (7 channels × 20 features) frequency-domain feature vector.

[0027] Step 23: Extract and fuse the time series features and frequency domain features of the time series data and the time-frequency spectrum through a pre-trained time-frequency domain joint feature extraction network to generate a comprehensive time-frequency domain feature representation of the relay.

[0028] Specifically, in this embodiment, the time-frequency domain joint feature extraction network adopts a multimodal fusion architecture with an attention mechanism. The time-frequency domain joint feature extraction network includes three parts: a time domain branch, a frequency domain branch, and a fusion module.

[0029] The time-domain branch consists of the multi-layer bidirectional long short-term memory network from step 21, outputting a 256-dimensional time-domain feature vector. The frequency-domain branch uses a two-dimensional convolutional neural network (2D-CNN) to process the spectrogram. This 2D-CNN consists of four convolutional blocks, each consisting of a 3×3 convolutional layer, a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer. The first convolutional block has 32 output channels, and this number doubles with each subsequent block. The 2D-CNN output is converted to a 128-dimensional frequency-domain feature vector through global average pooling and a fully connected layer.

[0030] The fusion module uses a cross-attention mechanism to mutually enhance the multi-channel time-series features and frequency-domain feature vectors. Specifically, the attention mechanism first calculates the attention weights of the multi-channel time-series features and frequency-domain feature vectors to generate frequency-domain-enhanced time-domain features. Then, the attention weights of the frequency-domain features on the time-domain features are calculated to generate time-domain-enhanced frequency-domain features. Finally, the two enhanced time-domain and frequency-domain features are concatenated with the original multi-channel time-series features and frequency-domain feature vectors, and dimensionality reduction is performed using a multi-layer perceptron to generate a 384-dimensional joint time-frequency feature vector.

[0031] The final generated time-frequency domain features are output in the form of structured data, which include the following three parts: a 256-dimensional time domain feature part, which encodes the timing dynamic characteristics of the relay operating state, including time domain characteristics such as current rise / fall time, voltage stability, and temperature change rate; a 128-dimensional frequency domain feature part, which characterizes the spectral characteristics of the relay during operation, including characteristic frequency, harmonic structure, spectral energy distribution, etc.; feature confidence information, including reliability assessment of feature extraction and uncertainty estimation based on signal quality and network output.

[0032] Furthermore, the step of inputting the time-frequency domain features into a pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state includes the following steps: Step 31: Construct a state judgment model combining a multi-layer perceptron and an attention mechanism, process the time-frequency domain features through the state judgment model and generate the electromagnetic interference state; Specifically, in this embodiment, the state judgment model adopts a hybrid architecture combining a multi-layer perceptron and a self-attention mechanism. The state judgment model comprises three main components: a feature encoding layer, a self-attention layer, and a state classification layer. The feature encoding layer consists of three fully connected layers, each containing 256 neurons. LeakyReLU is used as the activation function, and the dropout rate is set to 0.3. The feature encoding layer receives 384-dimensional time-frequency domain features and maps them to a 128-dimensional latent feature space through a nonlinear transformation. The self-attention layer uses a multi-head self-attention mechanism with 8 heads, each with a dimension of 16. This mechanism enhances the ability to perceive electromagnetic interference patterns by learning the correlations between the time and frequency domains within the time-frequency domain features. The state classification layer consists of two fully connected layers. The first layer contains 64 neurons, and the second layer has an output dimension of 2, corresponding to two electromagnetic interference states: weak interference and strong interference. Finally, the output is converted into a probability distribution using the Softmax function. The category with the highest probability is the predicted electromagnetic interference state.

[0033] Step 32: Constructing a sample data set of electromagnetic interference status, wherein the sample data set includes relay operation data under weak electromagnetic interference and strong electromagnetic interference environments; Specifically, the sample dataset for electromagnetic interference states consists of two types of data: weak interference environment data and strong interference environment data. The weak interference environment data was collected in a simulated vehicle environment with an electromagnetic field strength of 5-15V / m; the strong interference environment data was collected near a high-voltage substation with an electromagnetic field strength greater than 30V / m. For each type of environment, operational data for the relay was collected in different operating states (closed, open, and transition). 100 samples were collected for each operating state, with each sample lasting 5 seconds. The sample data were processed using a time-frequency domain feature extraction network to generate a 384-dimensional joint time-frequency domain feature vector, annotated with the corresponding electromagnetic interference state label. The resulting dataset contains 400 sample sets (two environments × two operating states × 100 sets), divided into training, validation, and test sets in a 7:2:1 ratio.

[0034] Step 33: Based on the sample data set, a batch gradient descent algorithm is used to train the state judgment model to obtain a trained state judgment model; Specifically, the state judgment model is trained using a batch gradient descent algorithm with a batch size of 32 and an initial learning rate of 0.001. The cross-entropy loss is used as the loss function to measure the difference between the predicted electromagnetic interference state and the true label. The Adam algorithm is used as the optimizer, with momentum parameters β1 set to 0.9, β2 set to 0.999, and a weight decay coefficient of 1e. -5 . A learning rate decay strategy is used during training, and the learning rate is reduced to 0.8 times the original value every 10 epochs. To prevent overfitting, an early stopping strategy is adopted, and training is stopped when the loss on the validation set has not decreased for 5 consecutive epochs. In this embodiment, the training of the state judgment model is performed on a GPU, using CUDA accelerated computing, and a single training session takes about 2 hours. After training, the accuracy of the state judgment model on the test set reached 95.7%, and the F1 score was 0.943, indicating that the model has a high ability to recognize electromagnetic interference states.

[0035] Step 34: Input the time-frequency domain features extracted in real time into the trained state judgment model to generate a probability distribution of the electromagnetic interference state; Specifically, during the real-time operation phase, the 384-dimensional time-frequency domain features generated in step 23 are input into the trained state judgment model. The time-frequency domain features are first transformed nonlinearly through the feature encoding layer, then captured through the self-attention layer to capture the internal correlation of the time-frequency domain features, and finally generated through the state classification layer to generate the probability distribution of two electromagnetic interference states (weak interference and strong interference). The probability distribution is output in the form of a two-dimensional vector, where each element in the vector corresponds to the probability value of an electromagnetic interference state, and the sum of the two probability values ​​is 1. The model takes the state with the highest probability as the current electromagnetic interference state. Furthermore, the method of judging whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment according to the electromagnetic interference state includes: Calculating an electromagnetic interference intensity index based on a probability distribution of the electromagnetic interference state, and determining whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment based on the electromagnetic interference intensity index; Specifically, in this embodiment, the electromagnetic interference intensity index is calculated by weighted summation: interference intensity = 0.5×P(weak interference) + 1×P(strong interference), where P(weak interference) and P(strong interference) respectively represent the probability values ​​of weak interference and strong interference in the probability distribution of the electromagnetic interference state.

[0036] It's understood that electromagnetic interference intensity is a continuous value ranging from 0 to 1, with larger values ​​indicating more severe electromagnetic interference. The preset interference intensity threshold is 0.4. When the interference intensity is greater than or equal to the threshold, the current operating state is determined to be in an electromagnetic interference environment. When the interference intensity is less than the threshold, the current operating state is determined to be not in an electromagnetic interference environment.

[0037] It is understandable that different levels of electromagnetic interference will have different degrees of impact on the working state of the relay. In an interference-free environment, the relay contacts operate stably and reliably. In a weak interference environment, it may cause increased contact jitter, but it will not affect basic functions, so future state prediction is necessary, but excessive intervention is not required. In a strong interference environment, it may cause contact malfunction, adhesion, or abnormal wear. In response to these varying degrees of impact, corresponding technical treatment measures need to be taken to ensure the reliable operation of the relay in various environments. Furthermore, the dynamically adjusting the timing parameters of the relay drive signal through the reinforcement learning model includes the following steps: Step 41: Construct a temporal parameter space model of the relay drive signal for the reinforcement learning model, and define the state space, action space, and reward function. Specifically, in an embodiment of the present invention, the time series parameter space model of the relay drive signal adopts a Markov decision process framework. Specifically, the state space consists of the electromagnetic interference state and the current operating state of the battery relay, and includes at least the following key dimensions: an electromagnetic interference intensity indicator, expressed as a continuous value between 0 and 1; the current contact state of the battery relay, including closed, open, and transition states; and a coil current stability indicator (e.g., represented by the current variance), reflecting the degree of coil current fluctuation.

[0038] The action space is defined as the set of timing parameters for the relay drive signal, including at least the following timing parameters: coil current rise time, with a control range of 0.5ms to 5ms; coil current hold time, with a control range of 10ms to 50ms; coil current fall time, with a control range of 1ms to 10ms; coil current peak value, with a control range of 80% to 120% of the rated value; and drive waveform types, including square, trapezoidal, and S-shaped. The action space uses a hybrid discrete-continuous representation, with the discrete portion representing the waveform type selection and the continuous portion representing the time and current parameters, for a total of five dimensions.

[0039] In this embodiment, the reward function design comprehensively considers the performance and reliability of the relay, and can be composed of the following components: a contact closure reliability reward, for example, a reward of +10 is given when the contact is reliably closed, and a penalty of -20 is given otherwise; a contact bounce penalty, for example, proportional to the number and duration of contact bounces, with a maximum penalty value of -15; an energy efficiency reward, for example, inversely proportional to the energy consumption of the drive signal, with a maximum reward value of +5; an electromagnetic interference suppression reward, for example, proportional to the system's resistance to electromagnetic interference, with a maximum reward value of +8; an operation delay penalty, proportional to the contact action response time, with a maximum penalty value of -5.

[0040] Step 42: Build the policy network and value network of the reinforcement learning model based on the deep deterministic policy gradient algorithm; Specifically, the reinforcement learning model constructed by the deep deterministic policy gradient algorithm includes four neural networks: a main policy network, a target policy network, a main value network, and a target value network.

[0041] In this embodiment, the main policy network adopts a fully connected neural network structure, consisting of three hidden layers, with 128, 256, and 128 neurons in each layer, respectively. The input layer receives a 12-dimensional state vector, and the output layer generates a 5-dimensional action vector, corresponding to the timing parameters of the relay drive signal. The hidden layers use the LeakyReLU activation function, while the output layer uses the Tanh activation function for continuous parameters and maps them to the corresponding range. For discrete waveform types, the Softmax activation function is selected. To enhance the network's expressiveness, a batch normalization layer can be introduced after the first hidden layer to improve training stability.

[0042] The main value network also uses a fully connected architecture, consisting of three hidden layers, with 128, 256, and 128 neurons in each layer, respectively. The input layer receives the concatenation of the state vector and the action vector, and the output layer consists of a single neuron, representing the Q-value estimate for the current state-action pair. The hidden layers use the LeakyReLU activation function, while the output layer does not use an activation function to allow the Q-value to vary within an arbitrary range. To prevent overfitting, a dropout layer is added after the second hidden layer, with a dropout rate of 0.2.

[0043] The structures of the target policy network and target value network are the same as the corresponding main policy network and main value network, but the parameters are slowly updated from the main network through a soft update mechanism, and the update coefficient τ is set to 0.001 to ensure smooth changes in the target network parameters and improve learning stability.

[0044] Step 43: Build an experience replay buffer for the reinforcement learning model, and store the interaction data during the reinforcement learning model training and application process through a priority sampling mechanism; Specifically, the experience replay buffer uses a circular queue structure with a maximum capacity of 100,000 experience tuples. Each experience tuple contains five elements: the current state, the action performed, the reward obtained, the next state, and a termination flag. The current and next states contain the electromagnetic interference characteristics and the relay operating status, the action performed contains the timing parameters of the relay drive signal, and the reward value is calculated according to the reward function defined in step 41.

[0045] Step 44: Implementing a gradient descent-based network parameter update algorithm for the reinforcement learning model, optimizing the policy network and value network of the reinforcement learning model, and completing the construction of the reinforcement learning model; Specifically, the network parameters of each neural network were updated using mini-batch gradient descent, sampling 256 experience tuples from the experience replay buffer at a time to form a batch. For the main value network and the target value network, the loss function was defined as the mean squared error, and the Adam optimizer was used with a learning rate of 0.001, β1=0.9, and β2=0.999.

[0046] The loss functions of the main policy network and the target policy network are defined as negative expected returns. The policy is optimized by maximizing the expected Q value under the current policy. Both the main policy network and the target policy network use the Adam optimizer, with a learning rate of 0.0005, β1 of the main policy network set to 0.9, and β2 of the target policy network set to 0.999. To prevent gradient explosion, a gradient clipping mechanism is implemented to constrain the gradient norm to the interval [-1, 1].

[0047] Step 45: Build a laboratory test platform to simulate the dynamic response characteristics of the relay under electromagnetic interference conditions to train the reinforcement learning model; Specifically, the laboratory test platform consists of an electromagnetic interference generator, a relay drive circuit, a sensor network, and a data acquisition system. The electromagnetic interference generator is capable of generating various types of electromagnetic interference, including pulse interference, continuous wave interference, and random noise interference. The pulse interference parameters are adjustable in the following ranges: amplitude, 0.5-3 times the rated voltage, duration, 0.1-5ms, and repetition frequency, 10-1000Hz; the continuous wave interference parameters are adjustable in the following ranges: frequency, 1-100kHz, and amplitude, 0.1-1 times the rated voltage; and the random noise interference parameters are adjustable in the following ranges: power spectrum density 10^-6-10^-3 V. 2 / Hz and bandwidth 1-500kHz.

[0048] The relay drive circuit is implemented using a programmable power supply and a digital signal processor to control the waveform, amplitude, and timing parameters of the drive signal. The sensor network includes current sensors, voltage sensors, temperature sensors, and contact status sensors, sampling at a frequency of 10kHz. The data acquisition system, based on a high-speed data acquisition card and real-time processing software, records and processes sensor data in real time, thereby constructing a precisely controlled sample data set. Step 46: Execute the training process of the reinforcement learning model, optimize the timing parameter strategy of the relay drive signal by interacting with the laboratory test platform, and obtain the trained reinforcement learning model; Specifically, in the specific implementation process of the present invention, the reinforcement learning training process adopts a staged training strategy, which is divided into three stages in total: a pre-training stage, a main training stage, and a fine-tuning stage.

[0049] During the pre-training phase, a rule-based control strategy was first used to collect 10,000 initial experience data points to populate the experience replay buffer. This rule-based control strategy linearly adjusts the drive signal parameters based on the electromagnetic interference intensity, providing foundational knowledge for subsequent learning. The pre-training phase performed 50,000 interaction steps, using a large exploration noise of σ = 0.3 and a learning rate of lr = 0.002 to rapidly explore the parameter space and establish an initial policy.

[0050] During the main training phase, 500,000 steps of environment interaction are performed, gradually reducing the exploration noise, for example, linearly decaying from σ = 0.2 to σ = 0.05, and the learning rate, for example, from lr = 0.001 to lr = 0.0002. Each training round contains 200 steps of interaction, corresponding to 200 operations of the relay.

[0051] In the fine-tuning phase, more complex and diverse electromagnetic interference patterns are used, 100,000 steps of environment interaction are performed, and smaller exploration noise, such as σ = 0.05 and learning rate, such as lr = 0.0001, are used to fine-tune the policy to adapt to extreme interference conditions.

[0052] Step 47: Based on the current electromagnetic interference state and the current working state of the relay, the timing parameters of the relay drive signal adapted to the current electromagnetic interference state are generated in real time by the trained reinforcement learning model; Specifically, the real-time timing parameter generation process is as follows: first, the latest current, voltage, and temperature signals are obtained from the sensor network; then, the time-frequency domain features are extracted through a joint time-frequency domain feature extraction network; then, the time-frequency domain features are input into the state judgment model to generate the electromagnetic interference state; finally, the electromagnetic interference state and the current working state of the relay are combined into a state vector, which is input into the policy network of the trained reinforcement learning model to generate the timing parameters of the relay drive signal.

[0053] The resulting timing parameters include: coil current rise time t_rise, which controls the time it takes for the relay coil current to increase from zero to the set value, with a typical range of 0.5-5ms; coil current hold time t_hold, which controls the time it takes for the relay coil current to remain at the set value, with a typical range of 10-50ms; coil current fall time t_fall, which controls the time it takes for the relay coil current to drop from the set value to zero, with a typical range of 1-10ms; coil current peak value I_peak, which controls the maximum value of the relay coil current, with a typical range of 80%-120% of the rated value; and drive waveform type type, which selects the waveform type that best suits the current electromagnetic interference environment, including square wave, trapezoidal wave, or S-shaped wave.

[0054] In a further preferred embodiment of the present invention, in order to prevent frequent fluctuations in parameters, a parameter smoothing mechanism may be further provided, that is, weighting the newly generated timing parameters and the currently applied timing parameters is fused to avoid short-term excessive fluctuations in the timing parameters; Furthermore, the method of predicting the multi-dimensional contact wear state of the relay by using a deep residual diagnosis network based on the time-frequency domain characteristics includes the following steps: Step 51: constructing a network structure of a deep residual diagnosis network, wherein the deep residual diagnosis network includes a feature input layer, a residual block stacking layer, and a contact wear state output layer; Specifically, the deep residual diagnostic network adopts a multi-layer residual network structure, which includes 1 feature input layer, a residual block stacking layer consisting of 16 residual blocks, and 1 contact wear status output layer. The feature input layer receives 384-dimensional time-frequency domain features and maps them to a 512-dimensional feature space through a fully connected layer. The residual block adopts a "bottleneck" design, and each residual block contains a three-layer structure: the first layer is a 1×1 convolution layer for dimensionality reduction; the second layer is a 3×3 convolution layer for feature extraction; the third layer is a 1×1 convolution layer for dimensionality increase and restoration of the number of feature channels. Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The residual connection bypasses these three layers and directly links the feature input layer to the contact wear status output layer to form a residual learning mechanism.

[0055] The 16 residual blocks are divided into four stages in a 4-4-4-4 pattern. The first residual block in each stage uses a convolution with a stride of 2 for downsampling, while a 1×1 convolution adjusts the dimensionality of the residual connections. The first stage outputs 64 feature channels, the second stage outputs 128 feature channels, the third stage outputs 256 feature channels, and the fourth stage outputs 512 feature channels.

[0056] The contact wear status output layer of the deep residual diagnosis network consists of a global average pooling layer and a fully connected layer. The global average pooling layer compresses the feature map output by the residual block stacking layer into a fixed-dimensional vector. The fully connected layer maps this fixed-dimensional vector into a multidimensional representation of the contact wear status, which includes four key indicators: the remaining contact life percentage, the contact surface roughness index, the contact resistance value, and the contact temperature rise coefficient.

[0057] Step 52: Collect sample data sets of relay contacts at different wear levels; Specifically, the sample dataset was constructed through a combination of laboratory accelerated life testing and field data collection. For example, the laboratory accelerated life testing used a current ramp-up method, conducting cyclic operation tests at 1.2 to 2 times the rated current, recording data on the relay's entire lifecycle, from new to complete failure. Field data collection involved collecting real-world relay usage data from electric vehicles with varying operating times, covering various operating conditions and environmental conditions.

[0058] The samples in the sample dataset are divided into five categories based on the degree of contact wear: new state (wear rate 0-5%), light wear (wear rate 5-25%), moderate wear (wear rate 25-50%), heavy wear (wear rate 50-75%), and near failure (wear rate 75-100%). Each type of contact wear sample contains 1,000-5,000 data records, each of which contains the time-frequency domain characteristics of the relay and the corresponding contact wear state label. The physical parameters corresponding to the contact wear state label are the contact remaining life percentage, contact surface roughness index, contact resistance value, contact temperature rise coefficient, and contact wear degree output by the contact wear state output layer of the deep residual diagnostic network. The above-mentioned contact wear state label can be determined through methods such as three-dimensional contact surface profile scanning, contact resistance measurement, temperature rise testing, and metallographic analysis.

[0059] Step 53: Combining regression loss and classification loss, designing a loss function of a deep residual diagnosis network including the regression loss function and the classification loss function to optimize the prediction accuracy of the contact wear state; Specifically, the loss function of the deep residual diagnosis network employs a multi-task learning framework to simultaneously optimize regression and classification tasks. The regression task predicts four continuous indicators: the remaining contact life percentage, contact surface roughness index, contact resistance, and contact temperature rise coefficient; while the classification task determines the five discrete degrees of contact wear.

[0060] The regression loss function, also known as the regression loss function, uses a weighted mean squared error (MSE) function to calculate the mean squared error between the predicted and true values ​​for each of the four continuous indicators. Each indicator is weighted based on its importance. The weight for the remaining contact life percentage is 0.4, the weight for the contact resistance value is 0.3, the weight for the contact surface roughness index is 0.2, and the weight for the contact temperature rise coefficient is 0.1.

[0061] The classification loss function is the cross-entropy loss function, which calculates the difference between the predicted wear level and the actual wear level. To address the class imbalance problem, class weights are introduced, giving higher weight to classes with fewer samples nearing failure.

[0062] The total loss function is the weighted sum of the regression loss function and the classification loss function.

[0063] Step 54: Based on the sample training set, a phased training strategy and a learning rate scheduling mechanism are adopted to train the deep residual diagnosis network to obtain a trained deep residual diagnosis network; In a specific embodiment of the present invention, the training of the deep residual diagnostic network adopts a three-stage strategy: a pre-training stage, a main training stage, and a fine-tuning stage. In the pre-training stage of the deep residual diagnostic network, a larger learning rate is used, such as an initial value of 0.01 and fewer residual blocks, such as 8, to learn the basic feature mapping relationship and train for 50 cycles. In the main training stage, the model parameters of the deep residual diagnostic network after pre-training are loaded, the complete 16 residual blocks are used, the learning rate is set to 0.001, and training is carried out for 100 cycles. In the fine-tuning stage, the parameters of the first 12 residual blocks of the deep residual diagnostic network are frozen, and only the last 4 residual blocks and the output layer are fine-tuned, the learning rate is reduced to 0.0001, and training is carried out for 50 cycles.

[0064] The optimizer for the Deep Residual Diagnostic Network uses a stochastic gradient descent algorithm with momentum, with a momentum coefficient of 0.9 and a weight decay coefficient of 0.0001. The learning rate schedule uses a cosine annealing strategy, where the learning rate decays as a cosine function over the training cycle to avoid oscillations during training.

[0065] The batch size is set to 64, and the training set is randomly shuffled at each training epoch to ensure that the model does not memorize the sample order. To prevent overfitting, in addition to the batch normalization layer in the residual block, a dropout layer with a dropout rate of 0.3 is added after the fully connected layer. An early stopping strategy is used during training, terminating training early when the validation set loss stops decreasing for 10 consecutive epochs.

[0066] Step 56: Based on the trained deep residual diagnosis network, according to the time-frequency domain features, predict the contact wear state of the relay in real time and generate a wear feature representation; Specifically, the real-time prediction process is as follows: first, the latest current signal, voltage signal and temperature signal are obtained from the sensor network; then, 384-dimensional time-frequency domain features are extracted through the time-frequency domain joint feature extraction network; then, the time-frequency domain features are input into the trained deep residual diagnosis network; finally, the network outputs a multi-dimensional representation of the contact wear status.

[0067] The contact wear status includes the following key information: the percentage of remaining contact life, which indicates the remaining service life of the relay contact relative to the new state, ranging from 0-100%; the contact surface roughness index, which characterizes the microscopic morphological characteristics of the contact surface, ranging from 0-10, with larger values ​​indicating rougher surfaces; the contact resistance value, which reflects the contact quality, typically ranging from 0.5-10mΩ, with larger values ​​indicating worse contact quality; the contact temperature rise coefficient, which indicates the temperature rise generated by the contact under unit current, ranging from 0.05-0.5℃ / A, with larger values ​​indicating more severe heat loss; the contact wear level, which discretizes the degree of wear into five levels, including new product, light wear, moderate wear, heavy wear, and near failure.

[0068] The contact wear status is output as structured data, including numerical indicators and confidence information. The confidence information is calculated based on the probability distribution of the model output and reflects the reliability of the prediction results, providing a decision basis for subsequent drive signal optimization.

[0069] Furthermore, generating the optimized parameters of the relay drive signal according to the contact wear state includes the following steps: For each dimension of the contact wear state, based on the physical characteristics of the dimension, a preset optimization strategy is used to generate the corresponding optimized parameters of the driving signal; In the specific implementation process of the present invention, with respect to the contact surface roughness index, when an increase in surface roughness is detected, the rising edge time of the drive signal is extended, for example, for every unit increase in roughness, the rise time is extended by 0.5 milliseconds, so as to reduce the impact force and bounce phenomenon when the contact is closed; with respect to the contact resistance value, when it exceeds 1.5 times the standard value, the drive current amplitude is increased according to a preset ratio, up to a maximum of 120% of the rated value, to ensure that the contact can overcome the increased contact resistance and achieve reliable closure; with respect to the contact temperature rise coefficient, when the temperature exceeds the preset temperature threshold, the drive signal adopts a segmented waveform, first achieving rapid attraction with a high current, for example, lasting for 2-3 milliseconds, and then decreasing to a lower holding current, for example, about 60% of the peak value, to reduce heat accumulation; with respect to the remaining contact life percentage, when the remaining contact life is less than 30%, the drive signal is decreased in a step-by-step manner, with the decrease in each step not exceeding 20% ​​of the total current.

[0070] Example 2 like Figure 2 As shown, the driving system of the electric vehicle battery relay based on deep learning includes a real-time data collection module, a time-frequency domain feature extraction module, an interference environment judgment module, a weak interference parameter optimization module, and a strong interference parameter optimization module; wherein each module is electrically connected; A real-time data collection module acquires the real-time operating data of the electric vehicle battery relay and sends the real-time operating data to the time-frequency domain feature extraction module; A time-frequency domain feature extraction module extracts the time-frequency domain features of the battery relay using the real-time operation data of the pre-trained time-frequency domain joint feature extraction network, and sends the time-frequency domain features to the interference environment judgment module and the strong interference parameter optimization module; The interference environment judgment module inputs the time-frequency domain features into the pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state. Based on the electromagnetic interference state, it determines whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment. If it is in an electromagnetic interference environment, it switches to the weak interference parameter optimization module; if it is in a strong electromagnetic interference environment, it switches to the strong interference parameter optimization module; In some embodiments, the system further includes: a weak interference parameter optimization module for dynamically adjusting the timing parameters of the relay drive signal through a reinforcement learning model when the current operating state is in a weak electromagnetic interference environment; The strong interference parameter optimization module predicts the multi-dimensional contact wear status of the relay based on the time-frequency domain characteristics through a deep residual diagnostic network, and generates optimized parameters of the relay drive signal based on the contact wear status to drive the electric vehicle battery relay.

[0071] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A driving method for an electric vehicle battery relay based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain real-time operating data of the electric vehicle battery relay; Step 2: extracting the time-frequency domain features of the battery relay from the real-time operation data through a pre-trained time-frequency domain joint feature extraction network; Step 3: Input the time-frequency domain features into a pre-trained state judgment model to generate an electromagnetic interference state, and determine whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment based on the electromagnetic interference state; Step 4: If the current operating state is in the strong electromagnetic interference environment, predict the multi-dimensional contact wear state of the relay through a deep residual diagnosis network based on the time-frequency domain characteristics; Step 5: Generate optimized parameters of the relay drive signal according to the contact wear state to drive the electric vehicle battery relay.

2. The driving method of the electric vehicle battery relay based on deep learning according to claim 1, characterized in that: After determining whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment, the method further includes: If the current operating state is in the weak electromagnetic interference environment, the timing parameters of the relay drive signal are dynamically adjusted through the reinforcement learning model.

3. The driving method of the electric vehicle battery relay based on deep learning according to claim 2, characterized in that: The real-time operation data includes current signals, voltage signals, and temperature signals, wherein the current signals include main circuit current and coil current, the voltage signals include main contact voltage and coil voltage, and the temperature signals include contact temperature, coil temperature, and ambient temperature. Extracting the time-frequency domain features of the battery relay from the real-time operation data using a pre-trained time-frequency domain joint feature extraction network includes the following steps: Step 21: organizing the main circuit current, coil current, main contact voltage, coil voltage, contact temperature, coil temperature, and ambient temperature in the real-time operation data into time series data of 7 signal channels; Step 22: using a method combining short-time Fourier transform and wavelet transform to perform frequency domain transform on the time series data of each signal channel to extract spectrum statistical features and form a time-frequency spectrum graph; Step 23: Extract and fuse the time series features and frequency domain features of the time series data and the time-frequency spectrum through a pre-trained time-frequency domain joint feature extraction network to generate a time-frequency domain feature representation of the relay.

4. The driving method of the electric vehicle battery relay based on deep learning according to claim 3 is characterized in that: The time-frequency domain joint feature extraction network adopts a multimodal fusion architecture with an attention mechanism. The time-frequency domain joint feature extraction network includes three parts: a time domain branch, a frequency domain branch, and a fusion module. The pre-trained time-frequency domain joint feature extraction network extracts and fuses the time series features and frequency domain features of the time series data and the time-frequency spectrum to generate the time-frequency domain feature representation of the relay, including: Through the time domain branch, a multi-layer bidirectional long short-term memory network is used to extract time domain features from the time series data to obtain multi-channel time series features; The frequency domain branch is used to process the time-frequency spectrum using a two-dimensional convolutional neural network, and the output of the two-dimensional convolutional neural network is converted into a multi-dimensional frequency domain feature vector through global average pooling and a fully connected layer; Through the fusion module, the multi-channel time series features and the frequency domain feature vectors are enhanced and spliced ​​using a cross-attention mechanism, and the output results after the enhancement and splicing are subjected to multi-layer perceptron dimensionality reduction to generate a multi-dimensional joint time-frequency domain feature vector as the time-frequency domain feature representation of the relay.

5. The driving method of the electric vehicle battery relay based on deep learning according to claim 4 is characterized in that, Inputting the time-frequency domain features into a pre-trained state judgment model to generate an electromagnetic interference state includes the following steps: Step 31: constructing a state judgment model combining a multi-layer perceptron and an attention mechanism, processing the time-frequency domain features through the state judgment model and generating an electromagnetic interference state; Step 32: constructing a sample data set of the electromagnetic interference state, wherein the sample data set includes relay operation data under weak electromagnetic interference and strong electromagnetic interference environments; Step 33: Based on the sample data set, the state judgment model is trained using a batch gradient descent algorithm to obtain a trained state judgment model; Step 34: Input the time-frequency domain features extracted in real time into the trained state judgment model to generate a probability distribution of the electromagnetic interference state.

6. The driving method of electric vehicle battery relay based on deep learning according to claim 5, characterized in that: The determining, based on the electromagnetic interference state, whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment includes: Calculating an electromagnetic interference intensity index according to a probability distribution of the electromagnetic interference state; It is determined whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment according to the electromagnetic interference intensity index.

7. The driving method of the electric vehicle battery relay based on deep learning according to claim 6, characterized in that: The method of dynamically adjusting the timing parameters of the relay drive signal by using a reinforcement learning model includes the following steps: Step 41: Construct a time series parameter space model of the relay drive signal for the reinforcement learning model, and define a state space, an action space, and a reward function, wherein the state space is used to indicate the electromagnetic interference state and the current working state of the battery relay, the action space is used to indicate the set of time series parameters of the relay drive signal, and the reward function is used to reward or punish based on the performance of the relay; Step 42: Constructing a policy network and a value network of the reinforcement learning model based on a deep deterministic policy gradient algorithm; Step 43: Constructing an experience replay buffer for the reinforcement learning model, and storing the interaction data during the training and application of the reinforcement learning model through a priority sampling mechanism; Step 44: Optimizing the policy network and value network of the reinforcement learning model based on a network parameter update algorithm using gradient descent to complete the construction of the reinforcement learning model; Step 45: Construct a laboratory test platform to simulate the dynamic response characteristics of the relay under electromagnetic interference conditions to train the reinforcement learning model; Step 46: adopting a phased training strategy to execute the training process of the reinforcement learning model, optimizing the timing parameter strategy of the relay drive signal by interacting with the laboratory test platform, and obtaining a trained reinforcement learning model; Step 47: Based on the current electromagnetic interference state and the current working state of the relay, the timing parameters of the relay drive signal adapted to the current electromagnetic interference state are generated in real time by the trained reinforcement learning model.

8. The driving method of the electric vehicle battery relay based on deep learning according to claim 7, characterized in that: The method of predicting the multi-dimensional contact wear state of a relay by using a deep residual diagnosis network based on time-frequency domain characteristics includes the following steps: Step 51: constructing a network structure of a deep residual diagnosis network, wherein the deep residual diagnosis network includes a feature input layer, a residual block stacking layer, and a contact wear state output layer; Step 52: Collecting sample data sets of relay contacts at different wear levels, wherein each data in the sample data set includes the time-frequency domain features of the relay and the corresponding contact wear state label; Step 53: combining regression loss and classification loss to design a loss function of the deep residual diagnosis network including the regression loss function and the classification loss function to optimize the prediction accuracy of the contact wear state; Step 54: Based on the sample training set, a phased training strategy and a learning rate scheduling mechanism are adopted to train the deep residual diagnosis network to obtain a trained deep residual diagnosis network; Step 56: Using the trained deep residual diagnosis network, predict the contact wear state of the current relay in real time according to the time-frequency domain features, and generate a wear feature representation.

9. The driving method of electric vehicle battery relay based on deep learning according to claim 8, characterized in that: Generating optimized parameters of the relay drive signal according to the contact wear state comprises the following steps: For each dimension of the contact wear state, a corresponding optimized parameter of the driving signal is generated using a preset optimization strategy according to the physical characteristics of the dimension.

10. A driving system for an electric vehicle battery relay based on deep learning, which is used to implement the driving method for an electric vehicle battery relay based on deep learning according to any one of claims 1 to 9, characterized in that: The driving system includes a real-time data collection module, a time-frequency domain feature extraction module, an interference environment judgment module, a weak interference parameter optimization module, and a strong interference parameter optimization module; wherein each module is electrically connected; A real-time data collection module acquires the real-time operating data of the electric vehicle battery relay and sends the real-time operating data to the time-frequency domain feature extraction module; A time-frequency domain feature extraction module extracts the time-frequency domain features of the battery relay using the real-time operation data of the pre-trained time-frequency domain joint feature extraction network, and sends the time-frequency domain features to the interference environment judgment module and the strong interference parameter optimization module; The interference environment judgment module inputs the time-frequency domain features into the pre-trained state judgment model to generate a probability distribution of the electromagnetic interference state. Based on the electromagnetic interference state, it determines whether the current operating state is in a weak electromagnetic interference environment or a strong electromagnetic interference environment. If it is in a strong electromagnetic interference environment, it switches to the strong interference parameter optimization module; The strong interference parameter optimization module predicts the multi-dimensional contact wear status of the relay based on the time-frequency domain characteristics through a deep residual diagnosis network, and generates the optimized parameters of the relay drive signal based on the contact wear status.

Images