A vibration noise grading early warning and active control method for a permanent magnet synchronous motor for vehicle

Abstract

The application discloses a kind of vehicle permanent magnet synchronous motor vibration noise grading early warning and active control method, belong to motor prediction and control field.This method will motor real-time operating parameter and historical NVH time series data input pre-trained neural network prediction model, obtain future time stator vibration acceleration distribution image and sound pressure distribution image;According to the prediction result and preset threshold value determine risk level and trigger grading early warning;Based on early warning level, solve id / iq current optimal adjustment increment using improved particle swarm algorithm, complete motor dynamic adjustment through current loop, realize vibration noise active suppression.The application can realize vibration noise early prediction and grading early warning, form "prediction-decision-control" closed loop, complete high-precision prediction under small sample, give consideration to vibration noise suppression and motor operation safety, improve motor operation smoothness and reliability.

Description

Technical Field

[0001] This invention relates to a motor control method, specifically to a method for graded early warning and active control of vibration and noise of a permanent magnet synchronous motor. Background Technology

[0002] With the rapid development of the electric vehicle industry, users' demands for overall vehicle comfort are increasing. As a core power component of electric vehicles, the permanent magnet motor's vibration acceleration and radiated noise performance directly determine the vehicle's sound quality and driving experience. During operation, the motor is subject to the coupling of multiple physical fields, including electromagnetic waves, mechanical structure resonance, and cooling system airflow, which excites complex vibrations and noise. Especially under dynamic conditions with frequent changes in speed and load, the vibration acceleration and radiated noise characteristics exhibit strong nonlinearity, non-stationarity, and multi-scale coupling features, posing a severe challenge to the prediction and control of the motor's vibration acceleration and radiated noise performance. This invention focuses on the prediction of motor sound and vibration fields, anomaly identification, and control optimization based on the prediction results.

[0003] Currently, the prediction of motor vibration acceleration and radiated noise performance mainly relies on traditional numerical methods such as the finite element method (FEM). The FEM, by constructing a detailed electromagnetic-structural-acoustic coupled mesh model, can accurately analyze the vibration acceleration and radiated noise characteristics of motors, playing a crucial role in the motor design phase. However, this method has inherent limitations: on the one hand, when facing dynamic operating conditions such as sudden changes in motor speed and load fluctuations, frequent mesh reconstruction and multiphysics coupling iterative calculations are required, resulting in excessive computation time and difficulty meeting the needs of real-time prediction; on the other hand, the FEM demands extremely high computational resources, making it unable to quickly generate data covering all operating conditions, which is difficult to match the sample size required for training artificial intelligence models, thus limiting its further improvement in prediction accuracy.

[0004] In engineering practice, direct measurement with vibration sensors is the most common monitoring method. This method involves placing accelerometers at key parts of the motor to collect vibration signals in real time, triggering an alarm when the signal amplitude exceeds a preset threshold. However, this threshold-based early warning method has significant drawbacks: First, the vibration signal is a manifestation of the vibration result; by the time the sensor detects an anomaly, the vibration noise has already occurred, making it a "post-event alarm" lacking proactive warning capabilities. Second, fixed thresholds cannot adapt to drastic changes in motor operating conditions, easily leading to false alarms or missed alarms. Third, this method can only monitor the vibration amplitude at a single measuring point, failing to reveal the spatiotemporal evolution and propagation path of vibration noise, making it difficult to guide precise control.

[0005] From a control perspective, existing motor control systems generally lack deep coupling with vibration acceleration and motor radiated noise early warning systems. Traditional control strategies typically treat vibration and noise as disturbances, employing passive suppression methods, and cannot actively adjust based on future vibration acceleration and motor radiated noise conditions. Even if some studies can achieve offline evaluation or short-term prediction of vibration acceleration and motor radiated noise performance, the prediction results have not been translated into real-time early warning commands and active control responses, failing to address dynamic problems such as control parameter drift and increased torque pulsation caused by vibration and noise.

[0006] In summary, existing technologies have insufficient early warning capabilities in monitoring and warning of motor vibration acceleration and motor radiated noise, and lack a closed-loop mechanism to convert high-precision prediction results into early warning commands and control optimization in real time. Summary of the Invention

[0007] Purpose of the invention: In view of the above-mentioned prior art, a vibration and noise classification early warning and active control method for automotive permanent magnet synchronous motors is proposed. This method can predict the vibration acceleration and radiated noise of the motor and provide early warning of future anomalies. The early warning results are then used for active control optimization to improve the smoothness of motor operation.

[0008] Technical solution: A method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors, comprising the following steps:

[0009] Step 1: Input the real-time operating parameters of the motor and historical NVH time series data into the pre-trained neural network prediction model to obtain the stator vibration acceleration distribution image and sound pressure distribution image of the motor at future moments;

[0010] Step 2: Based on the peak vibration acceleration and sound pressure level in the predicted image, compare them with the preset safety threshold to determine the current risk level of the motor and trigger a graded warning.

[0011] Step 3: Based on the determined risk level, use an optimization algorithm to solve for the optimal adjustment increment of the right-angle axis current id / iq, and use the current loop to dynamically adjust the motor to achieve active suppression of vibration and noise.

[0012] Furthermore, the process of establishing the neural network prediction model includes:

[0013] S1: Generating continuous time series sample data of electromagnetic-structural-acoustic multiphysics coupling of motor based on meshless method;

[0014] S2: Standardize and preprocess the generated sample data, and define the temporal relationship between the input and the label;

[0015] S3: Construct an end-to-end prediction network, which integrates a selective bidirectional Mamba modeling module, a physically guided evolution module, and a U-Net encoding / decoding structure;

[0016] S4: Train the network using a multi-dimensional combined loss function that includes physical consistency constraints to obtain the optimal model weights.

[0017] Furthermore, in S1, the moving least squares method is used to construct a field function approximation model to obtain the spatiotemporal distribution of vibration acceleration of the stator core and stator base, as well as the spatiotemporal distribution of sound pressure of the acoustic radiation field.

[0018] Furthermore, in S3, the selective bidirectional Mamba modeling module embeds an acoustic physics prior; the acoustic physics prior is initialized by diagonalizing the state transition matrix through the transient wave equation of the electroacoustic system, and the matrix elements are determined by the air acoustic attenuation coefficient, the speed of sound, and the wave number of the sound wave.

[0019] Furthermore, in S3, the physically guided evolution module initializes the acoustic MLS shape function by embedding it into the weights of the deep convolution kernel, and uses the variational weak form of the acoustic transient wave equation as the physical consistency constraint during the model forward propagation process.

[0020] Furthermore, the risk levels are classified as follows:

[0021] Safety level: Vibration acceleration < 2.5 m / s² 2 And the noise sound pressure level is < 65dB;

[0022] Warning level: Vibration acceleration is between 2.5 and 3.5 m / s². 2 Or the noise sound pressure level is between 65dB and 75dB;

[0023] Hazard level: Vibration acceleration > 3.5 m / s² 2 Or the noise sound pressure level is > 75dB.

[0024] Furthermore, in step 3, the adjustment of the orthogonal axis current id / iq adopts a constrained dimensionless multi-objective optimization. The core sub-objective is to minimize the NVH peak value within the prediction window. The degree of NVH exceeding the standard is quantified based on the difference between the predicted peak value of future time-domain vibration acceleration, the predicted peak value of noise sound pressure level and the safety level threshold. It also includes the sub-objectives of torque stability, stator current amplitude and motor efficiency loss. An adaptive penalty function method is used to penalize the constraint violation, and the penalty coefficient increases linearly with the algorithm iteration.

[0025] Furthermore, the current regulation in step 3 follows these constraints: the stator current amplitude does not exceed 1.1 times the rated current, the instantaneous peak current does not exceed the maximum allowable current of the motor, the steady-state torque fluctuation is ≤2%, the dynamic torque fluctuation is ≤5%, the voltage modulation ratio is ≤0.95, the motor speed fluctuation is ≤±1% of the rated speed, the motor system efficiency decreases by ≤0.5% during the warning level regulation, and the efficiency decreases by ≤2% during the dangerous level regulation.

[0026] Furthermore, the weight coefficients of the multi-objective optimization are adaptively adjusted according to the risk level, and the weight switching adopts a 100ms linear ramp transition; wherein the weights for safety level NVH are 0.2, torque 0.5, current 0.15, and efficiency 0.15; for warning level NVH, the weights are 0.45, torque 0.35, current 0.1, and efficiency 0.1; and for dangerous level NVH, the weights are 0.7, torque 0.15, current 0.1, and efficiency 0.05.

[0027] Furthermore, step 3 further includes decomposing the optimal adjustment amount into step amounts for a single control cycle: linear ramp decomposition is adopted under steady-state conditions, and S-shaped smooth acceleration and deceleration decomposition is adopted under dynamic conditions to achieve shock-free adjustment.

[0028] Beneficial Effects: This invention, through a complete technical solution of vibration and noise prediction, graded early warning, and active current control, solves the technical defects of existing motor vibration and noise monitoring and early warning systems, as well as passive control. Compared with existing technologies, it has the following beneficial effects:

[0029] 1. Achieving high-precision, physically compliant advance prediction of vibration and noise with small sample sizes: A meshless method is used to generate continuous time series samples of electromagnetic-structural-acoustic multiphysics coupling of motors. Standardized preprocessing is combined to construct standardized training data. The prediction network integrates a selective bidirectional Mamba modeling module and a physically guided evolution module, embedding the motor acoustic physics priors into the network structure. The variational weak form of the acoustic wave equation is used as a physical consistency constraint. Training is completed with a multi-dimensional combined loss function containing physical consistency constraints. Under the condition of limited sample size, the model is effectively avoided from overfitting, and the prediction results of the motor stator vibration acceleration distribution and sound pressure distribution at future time moments are accurately output.

[0030] 2. Achieve proactive, tiered early warning, shifting from passive monitoring to active early warning. Based on the vibration and noise data of future moments output by the neural network prediction model, and combined with preset safety thresholds, a three-level risk level is determined, enabling proactive early warning of vibration and noise anomalies. Compared with traditional real-time threshold monitoring, this effectively avoids the shortcomings of "post-event alarms," ​​and the early warning results can be directly used to guide subsequent active control.

[0031] 3. Construct a "prediction-decision-control" closed loop to achieve active suppression of vibration and noise. The vibration and noise prediction and graded early warning results are directly converted into id / iq orthogonal axis current adjustment commands. The motor is dynamically controlled through the current loop, achieving active suppression of vibration and noise from the source. This breaks through the technical limitations of traditional passive suppression and forms a complete prediction-control closed loop.

[0032] 4. Multi-objective optimization control, balancing vibration noise suppression and motor operation safety: An improved particle swarm optimization algorithm is used to solve for the optimal adjustment increment of id / iq current. The optimization objectives are to minimize NVH prediction value, suppress torque fluctuation, and minimize motor efficiency loss. At the same time, the rigid constraints of stator current amplitude and torque fluctuation are strictly followed to ensure stable and reliable motor operation during vibration and noise suppression.

[0033] 5. Shock-free adjustment ensures smooth motor operation. The optimal current adjustment is decomposed into single-control-cycle step amounts. In steady-state operation, a linear ramp decomposition is used, and in dynamic operation, an S-shaped smooth acceleration and deceleration decomposition is used to achieve shock-free adjustment of motor current, avoid secondary vibration during adjustment, and ensure smooth motor operation. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;

[0035] Figure 2 The example image shows the spatial distribution of motor vibration noise prediction error obtained through an end-to-end network integrating spatiotemporally selective bidirectional Mamba and physically guided evolution.

[0036] Figure 3 The following is a training verification error change curve obtained by training the model using a multi-dimensional combined loss function as an example.

[0037] Figure 4 The example shows a scatter plot comparing the predicted and measured values ​​of vibration noise obtained by integrating an end-to-end network of spatiotemporally selective bidirectional Mamba and physically guided evolution.

[0038] Figure 5 The example shows the model training learning rate adjustment curve obtained by cosine annealing learning rate scheduling. Detailed Implementation

[0039] The invention will now be further explained with reference to the accompanying drawings.

[0040] like Figure 1 As shown, a method for graded early warning and active control of vibration and noise of a vehicle permanent magnet synchronous motor is described, with the following specific steps:

[0041] Step 1: Generate sample data based on the gridless method.

[0042] The Moving Least Squares (MLS) method in the meshless approach is used to solve the electromagnetic-structural-acoustic multiphysics coupling equations of the motor, generating continuous time series sample data of the vibration acceleration field of the stator core and stator frame, and the acoustic radiation field of the motor. This provides a high-quality training dataset for subsequent vibration and noise prediction models. The specific implementation method is as follows:

[0043] 1.1 Determine the core geometric and physical parameters of the motor

[0044] The core geometric parameters and physical operating parameters of the permanent magnet synchronous motor are collected and input, including stator outer diameter, rated voltage, number of motor pole pairs, air gap length, stator and rotor material properties, structural stiffness parameters and air acoustic medium parameters. These parameters are used as the basic inputs for the meshless solution to construct the boundary conditions and operating constraints of the motor physical calculation model.

[0045] 1.2 Layout of Discrete Computing Nodes

[0046] Within the entire computational domain of the motor stator, rotor, air gap, and frame housing, meshless computational nodes are deployed using a uniform discrete method. The air gap region, being the core region for electromagnetic force wave excitation, employs a high-density node deployment to ensure computational accuracy. The stator, rotor, and stator frame regions utilize a conventional density node deployment to balance computational efficiency and accuracy, thus forming a discrete computational domain that does not require a structured mesh.

[0047] 1.3 Constructing a field function approximation model and solving the coupling equations

[0048] Based on the moving least squares method, a continuous smooth field function approximation model of vibration acceleration field and acoustic radiation field is constructed. The electromagnetic-structural-acoustic multi-physics coupling equations are solved simultaneously. The spatiotemporal distribution data of vibration acceleration of motor stator core and stator frame, as well as the spatiotemporal distribution data of sound pressure of motor acoustic radiation field are obtained through numerical calculation. Finally, continuous time series vibration acceleration and radiated noise samples are generated, with each time step corresponding to a complete set of physical field spatial distribution data.

[0049] 1.4 Sample Data Output

[0050] The obtained time-series physical field data is transformed into standardized image data, which serves as the input and label sample for the subsequent prediction network model. The data is time-series continuous and the physical field distribution is complete, which can truly reflect the evolution law of vibration and noise under the dynamic working conditions of the motor.

[0051] Step 2: Dataset partitioning and preprocessing.

[0052] The temporal vibration acceleration and radiated noise sample data of permanent magnet synchronous motors generated based on the meshless method are divided and standardized preprocessed to construct a standardized dataset suitable for training the vibration and noise prediction network, ensuring the stability, convergence efficiency, and prediction generalization of the model training. The specific implementation method is as follows:

[0053] 2.1 Random partitioning of the dataset

[0054] The time-series sample data is randomly divided according to a preset ratio, with the training set accounting for 70%, the validation set accounting for 20%, and the test set accounting for 10%. This random division method ensures that the operating conditions and physical field characteristics of each subset of samples are uniform and consistent, adapting to the data characteristics under the dynamic operating conditions of the motor.

[0055] 2.2 Input and Tag Timing Definition

[0056] The segmented sample data are time-series paired and standardized. Historical time-series vibration acceleration and radiation noise image data are used as model inputs for the prediction network, and vibration acceleration and radiation noise image data corresponding to future moments are used as model prediction labels. The time-series data are sorted in chronological order to ensure that the model can learn the real spatiotemporal evolution of motor vibration and noise.

[0057] 2.3 Image size is uniform and consistent

[0058] All time-series image data are uniformly scaled to a fixed size of 256×256 to meet the fixed requirements of the input image size for the convolutional layers and U-Net encoding / decoding structure in the vibration noise prediction network, ensuring the consistency, stability and compatibility of model feature extraction, downsampling and upsampling calculations.

[0059] 2.4 Tensor Transformation and Numerical Normalization

[0060] The 2D image data after size normalization is converted into a tensor format adapted to the deep learning framework, and the tensor values ​​are normalized to linearly map the numerical range to the [0,1] interval. This eliminates the dimensional differences of physical quantities such as vibration acceleration and sound pressure level, avoids the gradient vanishing and gradient exploding problems during model training, and improves the convergence speed and prediction stability of model training in small sample scenarios.

[0061] Step 3: Construction of the overall model of the vibration acceleration and motor radiated noise prediction network.

[0062] This invention constructs an end-to-end vibration acceleration and motor radiated noise prediction network model. Employing a modular, serial architecture, it sequentially completes feature extraction, temporal mining, physical correction, image reconstruction, and standardized output through a backbone network, spatiotemporally selective bidirectional Mamba modeling, a physics-guided evolution module, multi-scale feature fusion, a U-Net encoder-decoder network, a sharpness enhancement module, and an output layer. This achieves high-precision advance prediction of motor temporal vibration noise data. The model integrates spatial feature learning, long-term dependency capture, and electromagnetic-structural-acoustic physical field constraints, adapting to the highly nonlinear, non-stationary, and multi-scale coupled dynamic characteristics of motor vibration noise. Specific implementation details are as follows:

[0063] 3.1 Backbone Network Construction

[0064] The backbone network is constructed using two consecutive 3×3 convolutional layers, without introducing a batch normalization (BatchNorm) layer, to ensure the consistency of feature distribution during model training and validation. Basic spatial features of the vibration noise image are extracted through two-dimensional convolutional operations, and a Rectified Linear Activation (ReLU) function is used to perform nonlinear mapping, outputting a basic spatial feature map with edge, gradient, and local distribution characteristics, providing underlying feature support for subsequent temporal processing and physical correction.

[0065] 3.2 Spatiotemporal Selective Bidirectional Mamba Modeling

[0066] The spatiotemporally selective bidirectional Mamba modeling module, connected after the backbone network, is used to mine temporal dependencies of basic spatial features and adaptively focus on key time steps, solving the problems of difficulty in capturing long-range dependencies in the temporal evolution of motor vibration and noise, and the interference of redundant information with prediction accuracy. This module performs bidirectional multi-scale reconstruction of the traditional selective state-space model (SSM) and embeds physical priors of motor NVH (noise, vibration, and acoustic roughness), realizing joint modeling of spatial and temporal features. The specific implementation is as follows:

[0067] 1. Spatial global feature statistical extraction: For the input feature map at each time step, global average pooling (AvgPool) and global max pooling (MaxPool) operations are performed in parallel to extract spatial global features. After concatenation, the spatial statistical features of each time step are obtained, which are used to generate dynamic adjustment parameters for subsequent time series state updates, so as to realize the constraint guidance of spatial features on time series evolution.

[0068] Mathematically represented as: ,in Let be the input feature map at time step t. Let be the spatial statistical characteristics at the t-th time step.

[0069] 2. Dimensionality Reduction and Temporal Tokenization: 1×1 convolution is used to reduce the number of feature channels to 1 / 4 of the original number of channels, eliminating channel redundancy; the two-dimensional spatial feature map of each time step is flattened into a one-dimensional temporal token, completing the dimensionality adaptation between spatial features and temporal sequences, forming a standardized temporal feature sequence that meets the input requirements of the Mamba structure.

[0070] Mathematically represented as: ,in The projected weights of a 1×1 convolution. Let t be the temporal token at time step t, Flatten is the flattened one-dimensional channel function, and Conv2d is the two-dimensional convolution operation.

[0071] 3. Initialization of Selective State-Space Model (SSM) based on acoustic physics priors: The SSM state transition matrix is ​​diagonalized and initialized according to the discrete state-space form of the transient wave equation of the motor acoustics. The diagonal elements of the matrix are determined by the air acoustic attenuation coefficient, sound velocity and sound wave number, so that the model has physical priors on sound wave propagation and attenuation that conform to the physical laws of NVH in the initial stage of training, reducing the dependence on the size of the training samples.

[0072] The discrete state-space form of the transient wave equation for the electro-acoustic system is as follows: ,in Let be the system state vector at time t; A is the state transition matrix, describing the propagation and attenuation characteristics of sound pressure fluctuations; B is the input matrix; C is the output matrix; and D is the direct feedthrough term. Input timing token; To output the features, based on the acoustic propagation characteristics obtained by the meshless method, the state transition matrix A is diagonalized and initialized, with the initial values ​​of the diagonal elements being... satisfy: ,in The acoustic attenuation coefficient of air medium. The speed of sound in air. This represents the wavenumber of the corresponding order of sound waves.

[0073] 4. Bidirectional Multi-Scale Parallel SSM Modeling: A parallel structure is constructed with a forward causal SSM branch and a backward anti-causal SSM branch. The forward causal branch updates the state in forward time sequence to capture the causal evolution of vibration noise from history to the future; the backward anti-causal branch updates the state in reverse time sequence to explore the inverse dependence of noise attenuation characteristics and future states on historical features; each branch has three time-scale sub-branches with short period, medium period, and long period to adaptively match the multi-scale vibration and noise characteristics of the motor under various operating conditions.

[0074] The forward causal SSM branch uses a selective scan mechanism, and the state update formula is:

[0075] ;

[0076] in The time step parameter is determined by the spatial statistical characteristics of the current time step. Instead of the traditional Mamba which generates tokens solely from input tokens, it enables dynamic control of spatial features over temporal state updates. These are time scale coefficients specific to the vibration and noise time series; different scale sub-branches correspond to different... Values; , For gating weights, , For bias terms, This is the adaptive state transition matrix. For adaptive input matrix, In the forward-hidden state, The forward SSM outputs features. The state update formula for the backward anti-causal SSM branch is consistent with that of the forward branch, only the temporal scan direction is reversed, and the final output is the backward branch features. .

[0077] After completing the bidirectional multi-scale branch computation, the output features of each branch are concatenated along the channel dimension, and multi-scale feature fusion is achieved through 1×1 convolution to obtain the bidirectional SSM fused features. ,in The weights are 1×1 convolution kernel weights. For convolution bias terms, Multi-scale splicing features at time step t ,in It performs the splicing operation along the feature channel dimension (the last dimension).

[0078] 5. Temporal saliency weighting and feature fusion: A lightweight temporal saliency scoring branch is added. The contribution of each time step is quantified and scored through a two-layer fully connected MLP network, and the temporal saliency weight is obtained by Softmax normalization.

[0079] The lightweight time-series saliency scoring branch extracts the contribution weight of each time step from the hidden state of the bidirectional SSM, achieving focus on key time nodes. Specifically, it first processes the bidirectional SSM hidden state of each time step. and The data is then concatenated and scored using a two-layer fully connected MLP network to obtain an initial significance score for each time step. The MLP is a two-layer fully connected network that uses the ReLU activation function to output an initial saliency score for a single channel. Then, the sharpness of the attention distribution is adjusted using vibration noise parameters, and the final temporal saliency weights are obtained after Softmax normalization. ,in This is a vibration noise adjustment parameter with a default initial value of 0.1, which can be adaptively optimized during training. Let be the significance weight at time step t, satisfying This weight can be directly output for visualization analysis, pinpointing the most critical time points for NVH prediction. The smaller the time-series significance weight value, the more concentrated the weight distribution is at the critical time points.

[0080] 6. Feature Fusion and Output: Based on temporal saliency weights, the output features of the bidirectional multi-scale SSM are weighted and fused to automatically focus on key time steps for prediction and suppress redundant temporal information. Residual connections are introduced to retain the original global temporal information and avoid information loss due to over-correction of features. The weighted fused features are combined with residual features to output global temporal enhancement features with the same dimension as the backbone network input. It seamlessly integrates with the subsequent physical guidance evolution module, realizing full-link feature transmission and connection, where C, H, and W represent the three dimensions of channel, height, and width, respectively.

[0081] 3.3 Physics-Guided Evolution Module

[0082] The physics-guided evolution module, following the spatiotemporally selective bidirectional Mamba modeling module, is used to embed the physical laws of the electromagnetic-structural-acoustic coupling of the motor (the physical evolution laws of motor vibration acceleration and motor radiated noise) into the forward propagation process of the neural network. It performs physical consistency correction on the temporal attention-enhanced features, ensuring that the feature evolution matches the spatial propagation and temporal attenuation characteristics of vibration acceleration and radiated noise. The prediction results constrained from the network's bottom layer conform to engineering physics laws. The specific implementation is as follows:

[0083] 1. Meshless MLS and PINN Deeply Coupled Architecture: The architecture adopts a structure that deeply couples the meshless moving least squares (MLS) method with the physical information neural network (PINN). MLS provides the acoustic domain spatial discrete system and smooth shape function for PINN, while PINN provides the physical consistency constraint of the acoustic wave equation, realizing the organic integration of data-driven feature learning and physical field evolution rules.

[0084] 2. MLS-shaped function-driven convolution kernel initialization: Based on the principle of meshless approximation of the acoustic radiation domain of the motor, the acoustic MLS shape function is directly embedded into the weight assignment process of the 3×3 depth convolution kernel to complete the physical prior initialization of the convolution kernel, so that the convolution operation naturally simulates the anisotropic propagation characteristics of sound waves in the air medium; 1×1 pointwise convolution is used to realize the channel fusion of different frequency noise components, and its weight is initialized as an identity matrix and adaptively optimized with network training.

[0085] For any spatial node within the acoustic radiation domain of the motor The meshless approximation of the sound pressure field is: ,in spatial nodes The sound pressure approximation value at the location corresponds one-to-one with the noise physical quantity in the module feature; N is the total number of meshless discrete nodes in the acoustic radiation domain corresponding to the feature map; For discrete nodes The sound pressure characteristic value at that location; For the first The acoustic MLS shape function corresponding to each meshless node. basis functions ,matrix Weighted basis vectors Compactly supported Gaussian weighted function , For the radius of the compactly supported region, For influence domain coefficients, and All are initialized in a fixed manner within the module.

[0086] The weights of the 3×3 convolution kernels for each channel are directly assigned by the discrete matrix of the MLS shape function corresponding to the acoustic radiation direction, which is completed during the module initialization phase to simulate the anisotropic propagation characteristics of sound waves in the air domain; the 1×1 convolution is used for feature mixing of different frequency noise channels within the module, initialized as an identity matrix, and optimized synchronously with the module training.

[0087] 3. Depthwise separable convolution to simulate physical field propagation: A depthwise separable convolution structure is used to complete feature propagation calculations. The depthwise convolution branch is used to simulate the spatial diffusion, attenuation, and directional dependence of vibration acceleration fields and radiated noise fields; the pointwise convolution branch is used to complete the linear fusion and dimensionality regularization of multi-channel features, maintaining the accuracy of physical field modeling while reducing the number of parameters and computational load.

[0088] 4. Weak-form acoustic wave equation constraint: Using the acoustic transient wave equation as the only physical constraint, the second-order partial differential control equation is reduced to the first-order differential form through variational weak-form transformation and integration by parts, avoiding numerical oscillations caused by the second derivative and Laplace operator during automatic differentiation of the network; the meshless shape function is directly used as the weight function to participate in the weak-form integration, so that the physical residual calculation depends entirely on the internal characteristics of the module, without the need for external grid, external data or external equation input.

[0089] Among them, the acoustic transient wave equation is used to describe the propagation law of motor noise, and is defined as: Where t is the time variable and c is the wave speed. It is the Laplace operator. After variational weak form transformation and order reduction through integration by parts, we obtain a weak form constraint containing only first-order differential operators: ,in For the acoustic computation domain, For the weight function, The discrete approximate solution for the sound pressure field; weighting function Directly taking the acoustic MLS shape function Therefore, all spatial derivatives only need to calculate the first derivative of the shape function, without automatically calculating the second derivative of the network features or calculating the Laplace operator, thus eliminating numerical noise caused by higher-order derivatives from the computation path.

[0090] 5. Internal closed-loop physical residual loss calculation: The physical residual loss is constructed in real time during the forward propagation process. The degree of satisfaction of the feature field with the wave equation is used as the optimization objective to form an internal self-supervised physical constraint. This loss and the model supervision loss participate in the back propagation together, guiding the network parameters to converge towards the dual optimal direction of data fitting and physical compliance.

[0091] In the forward propagation, the module directly constructs the PINN physical residual loss based on a weak form: The loss is calculated independently within the module, without relying on external data, external equations, or additional meshes, thus achieving a fully closed-loop physical constraint.

[0092] 6. Residual Connection and Feature Output: A residual connection structure is introduced to perform weighted fusion of the module input features and the physically corrected features, retaining the original effective feature information and avoiding excessive suppression of data-driven features by physical constraints; the fused output is a corrected feature that conforms to the electromagnetic-structural-acoustic physical laws and is sent to the subsequent multi-scale feature fusion module.

[0093] The module employs residual connections and 1×1 convolutions to preserve original feature information, avoiding excessive suppression of effective features by physical constraints and ensuring feature integrity. Internally, pre-initialized meshless acoustic depthwise separable convolutions perform physical consistency correction on the input features, simulating the anisotropic propagation of sound waves. Its mathematical form is as follows: ,in These are input features. Depth convolutions initialized for meshless MLS within the module. This is a pointwise convolution. The result of fusing the residuals and propagation is: ,in It is the final output characteristic tensor after the physical guidance of vibration acceleration and motor radiated noise evolution.

[0094] The evolution module takes into account the features after time-attention fusion and uses depthwise separable convolution to simulate the anisotropic wave propagation of vibration acceleration and motor radiated noise physical fields. That is, the propagation rate of vibration acceleration and motor radiated noise physical quantities in different directions is different, which fits the actual distribution characteristics of motor vibration acceleration and motor radiated noise physical fields.

[0095] 3.4 Multi-scale feature fusion module

[0096] The multi-scale feature fusion module, following the physics-guided evolution module, is used to simultaneously extract fine-grained detail features and coarse-grained global distribution features from the vibration acceleration field and radiation noise field. This addresses the limitations of single-scale convolution's receptive field and incomplete feature learning, achieving an organic fusion of local details and overall evolutionary patterns. The specific implementation method is as follows:

[0097] 1. Multi-branch parallel feature extraction: Four-way parallel convolution branches are used to perform multi-scale parallel processing on the physically corrected features to extract the vibration acceleration and motor noise features of the stator at different scales.

[0098] ;

[0099] in , , , The features extracted respectively correspond to 1×1, 3×3, 5×5 and dilated convolutions. This indicates that the sampling step size interval is 2, and dilated convolution expands the receptive field without increasing the parameters.

[0100] 2. Multi-scale feature splicing and fusion: The multi-scale features output from the four branches are spliced ​​in the channel dimension. The feature dimension regularization and information fusion are completed by 1×1 convolution, so that the output features simultaneously contain local details, medium structure and global distribution information. The fusion process retains the original feature expressive power of each scale and avoids feature loss or prediction ambiguity caused by the dominance of a single scale.

[0101] 3. Feature Output: The output features multi-scale fusion that take into account both fine-grained details and global patterns, ensuring that the subsequent U-Net encoding and decoding network can accurately restore the peak position and distribution of vibration noise during image reconstruction, and correctly grasp the overall temporal evolution trend, thus providing support for high-precision image prediction output.

[0102] 3.5U-Net codec network

[0103] The U-Net encoder network is used for deep encoding and decoding reconstruction of multi-scale fused features, achieving accurate restoration from high-level semantic features to physical field images of vibration acceleration and radiation noise, while taking into account both global evolution patterns and the integrity of local details. This module adopts a combination structure of a three-layer downsampling encoder, a bottleneck layer, and a three-layer upsampling decoder. Skip connections are used to preserve shallow-level detailed features, avoiding information loss during reconstruction. The specific implementation is as follows:

[0104] 1. Encoder Downsampling: The encoder consists of three consecutive coding blocks, each containing two 3×3 convolutional layers and one max-pooling layer. The convolutional layers are used to extract high-level semantic features layer by layer, while the max-pooling layer is used to reduce spatial resolution and expand the receptive field. During downsampling, the feature map size is gradually reduced and the number of channels is gradually increased, achieving feature depth compression and key information extraction.

[0105] The downsampling formula is: ,in The input of the i-th encoder, It is the output of the i-th encoder. These are the downsampled features after pooling.

[0106] 2. Bottleneck layer feature enhancement: The bottleneck layer is located between the encoder and the decoder. Two 3×3 convolutional layers are used to further increase the number of channels. This is used to extract the highest-level global semantic features, fully characterize the overall spatiotemporal evolution trend of vibration noise, and provide core feature support for image upsampling reconstruction.

[0107] 3. Decoder upsampling and skip connections: The decoder consists of three consecutive decoding blocks. Each decoding block uses transposed convolution to upsample the feature map, restoring the spatial resolution step by step. Skip connections are used to concatenate and fuse the shallow detail features of the corresponding level of the encoder with the upsampled features of the decoder in the channel dimension, so that the decoder has both high-level semantic understanding and low-level detail restoration capabilities, and finally outputs a predictive feature map with regular size and complete details.

[0108] The upsampling formula is as follows: ,in For transposed convolution upsampling features, For the skip connection features of the corresponding level of the encoder, These are the output features of the decoder.

[0109] 3.6 Sharpness Enhancement Module

[0110] The sharpness enhancement module is used to correct detail blurring and edge blunting issues in U-Net reconstructed images, improving the clarity and accuracy of the physical field distribution of vibration acceleration and radiated noise. This module enhances high-frequency details while avoiding over-sharpening that introduces artifacts. The specific implementation method is as follows:

[0111] 1. High-frequency detail extraction: using the Laplacian operator Output features of the U-Net decoder High-frequency information is extracted to capture key details such as peak edges and gradient jumps in the vibration noise field.

[0112] 2. Adaptive sharpening intensity adjustment: Introduce learnable sharpening enhancement intensity parameters, and adaptively optimize the enhancement amplitude through network training to avoid noise amplification or lack of detail caused by fixed sharpening coefficients.

[0113] 3. Residual Fusion Output: The original features and high-frequency enhanced features are weighted and fused using a residual structure, preserving the stability of the global structure and enhancing only local details; the output features are high-precision prediction features with clear edges and accurate distribution, providing a guarantee for the final physical field output.

[0114] The mathematical representation of the sharpness enhancement module is as follows: ,in The enhancement strength parameter is a learnable parameter with an initial value of 0.3, which can be adaptively optimized through backpropagation.

[0115] 3.7 Output Layer

[0116] The output layer is used to convert the sharpened and enhanced feature map into a standardized prediction image consistent with the preprocessing format, thereby achieving the final output of the physical quantities of vibration acceleration and radiation noise. The specific implementation method is as follows:

[0117] 1. Feature Dimensionality Reduction and Single-Channel Mapping: A two-layer convolutional structure is used to complete feature mapping. The first convolutional layer performs channel dimensionality reduction and retains effective physical features through the ReLU activation function. The second convolutional layer further compresses the features into a single-channel feature map, which corresponds one-to-one with the physical quantity distribution of the vibration acceleration field or radiation noise field.

[0118] 2. Numerical normalization: The Sigmoid activation function is used to constrain the output values ​​to the [0,1] interval, which is consistent with the normalization range of the previous data preprocessing, ensuring that the numerical system is unified, inversely solvable, and comparable.

[0119] 3. Predicted Image Output: Outputs a 256×256 standardized single-channel predicted image, which can be directly restored to the actual vibration acceleration and sound pressure level physical quantities through inverse normalization, and used for subsequent graded early warning and id / iq current closed-loop control.

[0120] The mathematical representation of the output layer is: ,in It is a predicted image of the output vibration acceleration and motor radiated noise.

[0121] Step 4: Predictive network training.

[0122] Prediction training is used to optimize parameters, achieve model convergence, and validate performance of the vibration acceleration and radiated noise prediction network. To address the issues of overly smoothed predictions, loss of detail, and overfitting caused by traditional single loss functions in small-sample scenarios, a multi-dimensional combined loss function is constructed, along with an adaptive optimization strategy and learning rate scheduling mechanism. This ensures rapid and stable model convergence and simultaneous improvement in prediction accuracy and detail representation. The specific implementation is as follows:

[0123] 4.1 Combination Loss Function

[0124] To overcome the shortcomings of traditional mean squared error (MSE) loss in predicting vibration-noise temporal images, such as detail blurring, structural distortion, and small sample overfitting, this invention adopts a multi-dimensional weighted combined loss function, which integrates multiple constraints such as pixel-level error, structural similarity, spatial gradient, detail perception, and temporal attention to achieve synergistic optimization of data-driven fitting and physical field feature preservation.

[0125] The combined loss function consists of mean squared error (MSE) loss. Structural Similarity (SSIM) Loss gradient loss Perceived loss and attention regularization loss The loss is weighted, with the following preset weights for each loss term: mean squared error loss weight 0.5, structural similarity loss weight 0.2, gradient loss weight 0.2, perceptual loss weight 0.1, and attention regularization loss weight 0.01.

[0126] Among these, mean squared error loss ensures pixel-level matching between predicted and true values; structural similarity loss constrains the overall distribution and structural consistency of the vibration and noise physical field; gradient loss preserves key gradient features such as vibration peaks and noise edges, preventing overly smooth prediction results; perceptual loss enhances the ability to reproduce high-frequency details and improves feature fitting accuracy in small sample scenarios; and attention regularization loss constrains the temporal saliency weight distribution, preventing the model from overemphasizing local time steps and causing prediction bias. These multi-dimensional constraints work together to achieve effective regularization in small sample training scenarios, guiding the model towards optimization towards high accuracy, high fidelity, and physical compliance.

[0127] The combined loss function is expressed as: ,in This represents the total loss.

[0128] 4.2 Network Training

[0129] The trainer integrates loss calculation, parameter update, and learning rate adjustment functions to construct a complete model training closed loop. The specific configuration is as follows:

[0130] 1. Loss function initialization: Load the combined loss function and preset weight configuration defined in this invention to ensure that the loss calculation logic is completely matched with the output features and sample labels of the prediction network, so as to provide a precise optimization target for model backpropagation.

[0131] 2. Optimizer Configuration: The Adam adaptive learning rate optimization algorithm is used to update network parameters. This algorithm can adaptively adjust the parameter update step size, adapting to the nonlinear and non-stationary characteristics of vibration noise data, effectively improving convergence speed and training stability. The initial learning rate of the optimizer is set to 1e-3, and the optimization object covers all trainable parameters of the prediction network, including convolution kernel weights, bias terms, sharpening enhancement intensity parameters, temporal attention parameters, etc.

[0132] 3. Learning Rate Scheduling Configuration: A cosine annealing learning rate scheduling strategy is adopted to achieve a periodic and smooth decay of the learning rate, avoiding oscillations and overfitting in the later stages of training. The scheduling period length is set to 100 epochs, and the minimum learning rate is set to 1e-5. In the early stage of training, a larger learning rate is used to quickly converge to the optimal solution region, and in the later stage, a smaller learning rate is used to fine-tune the parameters, ensuring that the model converges stably to the global optimum and improving generalization ability.

[0133] After training, the optimal model weights are selected based on the performance of the validation set and used for subsequent vibration and noise advance prediction, graded early warning and current closed-loop control.

[0134] Step 5: Implementation of closed-loop control for vibration acceleration, noise classification early warning, and dynamic adjustment of id / iq current.

[0135] Based on the trained vibration acceleration and radiated noise prediction network, a closed-loop control system for permanent magnet synchronous motors is constructed, encompassing the entire process of "prediction-decision-control-feedback-iteration". The advanced prediction results are transformed into graded early warning commands and dynamic adjustment strategies for the id / iq orthogonal shaft currents. This proactively suppresses abnormal vibration and noise growth from the electromagnetic excitation source, achieving proactive prevention and control of NVH risks and active optimization of the motor's operating status. The specific implementation method is as follows:

[0136] 5.1 Overall Framework of Closed-Loop System

[0137] The closed-loop system consists of three core units: a vibration and noise prediction module, a graded early warning and risk decision-making module, and an id / iq current vector dynamic adjustment module. These units communicate with each other and coordinate in time, fully covering the entire process of NVH (Noise, Vibration, and Harshness) prediction, risk level assessment, and active control adjustment. The specific workflow is as follows:

[0138] 5.1.1 Tiered Early Warning and Risk Assessment

[0139] The prediction module loads the optimal weights of the trained prediction network, inputs real-time motor operating parameters and historical NVH time-series data, and outputs full-scale prediction data for future time-domain vibration acceleration and radiated noise sound pressure level. The real-time motor operating parameters include real-time motor speed, speed command setpoint, output electromagnetic torque, load rate, real-time feedback values ​​of the right-angle shaft current (id, iq), real-time DC bus voltage, real-time voltage modulation ratio, motor rated current, maximum allowable peak current, rated speed, and rated voltage. The graded early warning module classifies risks based on preset safety thresholds, triggers corresponding early warning mechanisms, and outputs control commands. The threshold settings and response rules are as follows:

[0140] Safety level: Vibration acceleration < 2.5 m / s² 2 Furthermore, the noise sound pressure level is <65dB, the motor operates normally, there is no warning prompt, and only the predicted data is recorded for trend analysis;

[0141] Warning level: Vibration acceleration is between 2.5 and 3.5 m / s². 2 If the noise sound pressure level is between 65 and 75 dB, it indicates that the vibration acceleration and motor radiated noise are about to exceed the safe range, triggering a yellow audible and visual warning to alert maintenance personnel to pay attention to the operating status. At the same time, the warning information (including predicted vibration acceleration and motor radiated noise data, risk level, and expected time of exceeding the standard) is displayed on the monitoring terminal to remind relevant personnel to pay attention to the motor operating status and simultaneously output a warning-level current adjustment command to the id / iq current adjustment module.

[0142] Hazard level: Vibration acceleration > 3.5 m / s² 2If the noise sound pressure level is greater than 75dB, it indicates that the vibration acceleration and motor radiated noise are seriously exceeding the standard, and there is a risk of motor damage. This will trigger a red audible and visual warning and push a remote (SMS / APP) alarm message to the maintenance personnel. At the same time, an emergency adjustment command will be output. If the NVH prediction value does not drop within 100ms, the motor load reduction protection will be triggered.

[0143] 5.1.2 Core Constraints and Multi-Objective Optimization of Current Regulation

[0144] The id / iq current regulation module receives predicted data and risk decision instructions, with minimizing the predicted NVH value as the core optimization objective. Simultaneously, it strictly adheres to the rigid constraints of motor operation to ensure the vehicle's dynamics and safety. These constraints include:

[0145] 1. Stator current RMS value constraint: The magnitude of the stator current vector synthesized from id and iq shall not exceed 1.1 times the rated current, mathematically expressed as: ,in For direct-axis current adjustment increment, Give a value to the optimized id. This is the current steady-state initial value of the direct-axis current id; To adjust the increment of the quadrature-axis current, Let iq be the current quadrature-axis current in steady state initial value. Given a value for the optimized iq, This is the rated current.

[0146] 2. Stator current peak constraint: The instantaneous peak current shall not exceed the maximum allowable current of the motor, mathematically expressed as: ,in This represents the maximum permissible peak current of the motor.

[0147] 3. Steady-state torque ripple constraint: Steady-state torque ripple ≤ 2%, mathematically expressed as... ,in To output electromagnetic torque to the motor, This is the reference value for the motor torque requirement issued by the vehicle's VCU.

[0148] 4. Dynamic torque ripple constraint: Dynamic torque ripple ≤ 5%, mathematically expressed as... .

[0149] 5. Voltage modulation ratio constraint: Voltage modulation ratio ≤ 0.95, mathematically expressed as... ,in , These are the direct-axis and quadrature-axis voltages, respectively. This is the real-time voltage of the DC bus.

[0150] 6. Speed ​​fluctuation constraint: Motor speed fluctuation ≤ ±1% of rated speed, mathematically expressed as: ,in This refers to the real-time speed of the motor. This is the speed command value.

[0151] 7. Efficiency Decrease Constraint: During early warning-level adjustment, the efficiency decrease of the motor system is ≤0.5%; during emergency adjustment at the danger level, the efficiency decrease is ≤2%, mathematically expressed as... , To adjust the steady-state operating efficiency of the front motor, To determine the motor operating efficiency after id / iq adjustment; The efficiency threshold is 0.005 for the warning level and 0.02 for the danger level.

[0152] The above constraints are transformed into multi-objective optimization sub-terms and penalty terms, and a dimensionless multi-objective optimization function is constructed. The optimization variable is the id / iq current regulation increment. The weight coefficients are adaptively adjusted according to the risk level. The NVH suppression weight is increased in the dangerous level, and the torque stability and operating efficiency are prioritized in the safe level.

[0153] The core excitation source of vibration and noise in permanent magnet synchronous motors (PMSMs) is the radial electromagnetic force wave within the stator-rotor air gap. Its amplitude is proportional to the square of the air gap magnetic flux density, and its frequency is strongly correlated with stator current harmonics, the number of pole pairs, and the rotational speed. id and iq, as the core control variables of the PMSM's vector control, directly determine the amplitude, phase, and harmonic distribution of the stator armature reaction magnetomotive force, thereby altering the spatiotemporal distribution characteristics of the air gap magnetic flux density and ultimately achieving active control over the radial electromagnetic force wave and its derived vibration and noise. The adjustment process is based on the PMSM electromagnetic torque formula as its core constraint, mathematically expressed as: ,in To output electromagnetic torque to the motor, This represents the number of pole pairs of the motor. For rotor permanent magnet flux linkage, , These are direct-axis and quadrature-axis inductors, respectively. Furthermore, all id and iq adjustments must be performed within the constraint boundaries to ensure motor operation safety and overall vehicle driving performance.

[0154] With minimizing the predicted NVH values ​​as the core objective, the rigid constraints of motor operation are transformed into sub-objectives and penalty terms. A constrained dimensionless multi-objective optimization function is constructed to resolve the conflict between NVH suppression, torque stability, and efficiency constraints, providing a target benchmark for solving the optimal adjustment. The core independent variables for optimization are the adjustment increments of id and iq, defined with the initial values ​​of id and iq during the current steady-state operation of the motor as the benchmark: We employ a weighted summation-type multi-objective optimization architecture, unifying the core objective and constraint objective into a minimization problem, mathematically represented as: ,in The smaller the value of the objective function, the better the optimization effect. Four dimensionless sub-objective terms; The weight coefficients for each sub-objective are adaptively adjusted based on the risk level to satisfy... =1; This is a rigid constraint penalty term. The penalty coefficient is... To constrain the violation degree function, ,in For torque tracking equality constraints, And satisfy A penalty is triggered if the torque deviates from the required torque.

[0155] The specific details of each sub-objective are as follows:

[0156] 1. The core sub-objective is to minimize the NVH peak value within the prediction window, directly receiving the output data from the prediction module. Its quantification formula is: ,in For the degree of NVH exceeding the standard, To predict the peak value of future time-domain vibration acceleration based on the current adjustment increment, The predicted peak value of the noise sound pressure level after synchronous correction; , All are security level thresholds, default. =2.5m / s 2 , =65dB; The weighting coefficients for vibration and noise are set by default. =0.5, which can be adaptively adjusted according to the items exceeding the standard.

[0157] 2. The quantification formula for the torque stability sub-objective is: In the formula For torque ripple rate, This refers to the adjusted electromagnetic torque. The torque stability sub-target is used to ensure that the motor output torque remains stable before and after adjustment.

[0158] 3. The quantification formula for the stator current amplitude sub-target is: ,in For current overload ratio, This is the rated current of the motor. The stator current amplitude sub-target is used to prevent stator current overcurrent.

[0159] 4. The quantification formula for the efficiency loss sub-objective is: ,in This represents the relative increase in copper loss. The efficiency loss sub-target is used to limit the additional heat generation and energy consumption increase caused by regulation.

[0160] An adaptive penalty function method is employed, triggering a penalty only when the optimization variable violates rigid constraints. The greater the violation, the higher the penalty value, ensuring the optimal solution falls within the feasible region. The penalty coefficient λ increases linearly from 10 to 1000 with the number of algorithm iterations, expanding the search space in the early stages and strictly constraining the feasible region in the later stages. The weight coefficients for each sub-objective are as follows. Based on the dynamic switching of risk levels output by the early warning module, the core priorities under different operating conditions are balanced, and the specific allocation is as follows:

[0161]

[0162] The weight switching adopts a 100ms linear ramp transition to avoid abrupt changes in adjustment; at the same time, the weights can be finely adjusted according to the NVH exceeding the standard to achieve targeted optimization.

[0163] 5.1.3 Improved Particle Swarm Optimization (PSO) Algorithm for Real-Time Solution

[0164] The particle swarm optimization algorithm is used as the real-time solver for the optimal id / iq current regulation in this system. The algorithm inputs are NVH prediction data, current motor operating parameters, risk level and corresponding weight, and rigid constraint boundaries; the output is the optimal regulation increment that satisfies all constraints on id / iq. , To meet the real-time requirements of the vehicle controller, the standard particle swarm optimization algorithm was optimized. A small-scale population initialization was implemented based on the risk level, reducing the search space: the population size was set to 20, and the maximum number of iterations was set to 50. An adaptive inertial weight was introduced, linearly decreasing from 0.9 to 0.4 with the number of iterations, balancing global search and local optimization capabilities. A constrained adaptive penalty function was used to construct the fitness function, ensuring convergence to the optimal solution within the feasible region. These improvements ensured that the optimal solution was found within 10ms, matching the motor current loop control cycle.

[0165] The iterative solution process is as follows:

[0166] 1. Initialize algorithm parameters, particle swarm positions and velocities, calculate the initial fitness of each particle, and determine the initial individual optimum and global optimum;

[0167] 2. Update the velocity and position of particles generation by generation, perform boundary trimming on the updated positions, calculate new fitness values, and simultaneously update the individual optimal and global optimal values;

[0168] 3. When the maximum number of iterations is reached or the rate of change of the global optimal fitness is less than 1e-4 for three consecutive generations, the iteration is terminated and the global optimal solution is output.

[0169] 4. Perform a full constraint rigidity check on the output optimal solution. If the check fails, select the suboptimal feasible solution. It is strictly forbidden to output adjustment commands that violate the constraints.

[0170] 5.1.4 Shockless step adjustment execution

[0171] To avoid secondary NVH problems caused by sudden current changes due to a one-time issuance of the optimal adjustment value, the optimal adjustment value is decomposed into single-control-cycle step values, with step size constraints set in stages: the maximum single-cycle step size is ≤0.5% of the rated current for the safety level, ≤1% for the warning level, and ≤2% for the danger level; a linear ramp decomposition is used for steady-state conditions, and an S-shaped smooth decomposition is used for dynamic conditions, so as to achieve shock-free and jerk-free adjustment process; after each single-step adjustment, the real-time operating status of the motor is fed back to the prediction module for iterative correction of the adjustment strategy.

[0172] In steady-state operation, a linear ramp decomposition method is used. First, the total number of decomposition cycles is calculated based on the total adjustment and the maximum step size per cycle. Then, the total adjustment is evenly distributed to each cycle, and a fixed step size is superimposed on each control cycle, resulting in smooth execution and low computational complexity. In dynamic operation, an S-shaped smooth acceleration / deceleration decomposition method is used. The step coefficient for each cycle is calculated using an S-shaped function, achieving a "gradual start and stop" of the step size, avoiding sudden current jumps, and adapting to the closed-loop stability requirements under acceleration / deceleration and load change conditions.

[0173] The step size decomposition is not executed in a fixed manner. After each control cycle completes the single-step adjustment, the real-time operating parameters of the motor are fed back to the prediction module to update the NVH prediction data. The optimal adjustment amount after correction is re-solved, and the step size is decomposed again based on the remaining number of cycles to achieve closed-loop control of "single-step execution - status feedback - iterative optimization" to ensure that the adjustment effect always matches the latest NVH prediction trend.

[0174] 5.1.5 Hierarchical Adjustment and Control Strategy

[0175] Differentiated adjustments are implemented based on risk levels to achieve precise matching:

[0176] 1. Safety-level pre-adjustment: Only executed under steady-state conditions. id performs negative field weakening fine adjustment, with a cumulative adjustment amount ≤ ±2% of the rated value. iq provides synchronous torque compensation to ensure constant output torque. The adjustment period is 100ms. Adjustment stops after the NVH prediction value falls back to the safe range.

[0177] 2. Early warning level mild suppression: id prioritizes negative magnetic weakening, with an adjustment range of ±5% of the rated value, a single step size of 1%, and an adjustment period of 50ms; iq core completes torque compensation, with an iq fundamental frequency adjustment range of ±3% of the rated value, and synchronously superimposed with a reverse harmonic current not exceeding 3% of the fundamental frequency to specifically counteract abnormal frequency electromagnetic excitation; under dynamic operating conditions, the adjustment range is narrowed, the step size is reduced, and the adjustment period is extended to prioritize closed-loop stability.

[0178] 3. Dangerous Level Deep Coordinated Adjustment: The id adjustment range is expanded to ±10% of the rated value, with a single step size of 2% and an adjustment cycle of 20ms. The first cycle directly executes a negative field weakening of 3% of the rated value to quickly suppress NVH peaks. The iq fundamental frequency adjustment range is expanded to ±8% of the rated value, which can moderately reduce torque output and armature reaction, and simultaneously superimpose multiple reverse harmonics not exceeding 5% of the fundamental frequency. The current loop PI parameters are simultaneously optimized to accelerate the response speed. If the adjustment has no effect for 60ms, the motor load reduction protection is triggered.

[0179] 5.1.6 Feedback Control

[0180] After the controller executes the adjustment command, the real-time operating status of the motor, the actual feedback values ​​of id / iq, and the measured NVH data are synchronously transmitted back to the prediction module to complete the online iterative optimization of the prediction model and the closed-loop correction of the control strategy, so that the motor is always in a low-vibration, low-noise, and high-reliability operating state.

[0181] 5.2 Fault Tolerance and Security Protection Mechanism of Closed-Loop System

[0182] To ensure the stable and reliable operation of the closed-loop system, a multi-level fault-tolerance and safety protection mechanism is set up to prevent system failure from causing motor and vehicle malfunctions. The specific implementation method is as follows:

[0183] 1. Real-time constraint verification mechanism: Before executing the adjustment command in each control cycle, the id / iq adjustment is fully constrained for current, torque, voltage and speed. If the verification fails, the step size is reduced or the adjustment is terminated.

[0184] 2. Torque fluctuation fallback protection: Torque fluctuation is monitored in real time during the adjustment process. If the fluctuation exceeds the constraint threshold, the NVH adjustment is immediately paused and switched to torque priority mode. The adjustment is restarted after the operation is stable.

[0185] 3. Fault tolerance handling for adjustment failures: If the NVH prediction value does not improve after 5 consecutive adjustment cycles, switch to the preset fixed suppression parameter and trigger the highest level warning at the same time; if the prediction model fails or the algorithm times out, immediately lock the NVH adjustment link and restore the default vector control parameters to ensure that the system does not lose control capability.

[0186] 4. Emergency Interlock Mechanism: If a fault such as motor overcurrent, overvoltage, loss of synchronism, or insulation abnormality is detected during the adjustment process, the NVH adjustment link will be immediately cut off, all id / iq adjustment commands will be locked, and the motor main controller will execute the safety protection logic to ensure the safety of the motor and the whole vehicle.

[0187] Figure 2 This is a spatial distribution diagram of the motor vibration and noise prediction error in this embodiment. Figure 2It can be seen that the error only appears in a small amount on the surface of the casing, and the maximum error is only 0.008, indicating that the prediction network proposed in this invention highly matches the actual value in predicting the spatial distribution of vibration noise of the sample.

[0188] Figure 3 This is a graph showing the change in training and validation errors of the model in this embodiment. Figure 3 As can be seen, the training loss smoothly and monotonically decreased from the initial 0.1096 to the final 0.001067, and the validation loss synchronously decreased from the initial 0.0674 to approximately 0.0011, without any oscillations or increases throughout, clearly indicating that the loss function continuously and steadily decreased. This demonstrates that the model training process was stable, fully converged, and reached optimal performance.

[0189] Figure 4 This is a scatter plot comparing the predicted and measured vibration and noise values ​​in this embodiment. Figure 4 It can be seen that the points on the sample are basically evenly distributed on both sides of the ideal fitting curve and are very close to each other, and the coefficient of determination R is... 2 The values ​​are close to 1, and the RMSE and MAE values ​​are very small, indicating that the model's prediction results are very reliable.

[0190] Figure 5 This is a graph showing the learning rate adjustment during model training in this embodiment. Figure 5 It can be seen that the learning rate satisfies the following conditions: a high learning rate in the early stage to quickly explore the image feature space; a gradually decreasing learning rate in the middle stage to finely adjust the feature extractor; and an extremely low learning rate in the later stage to converge to the optimal classification / detection boundary.

[0191] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for graded early warning and active control of vibration and noise in automotive permanent magnet synchronous motors, characterized in that, Includes the following steps: Step 1: Input the real-time operating parameters of the motor and historical NVH time series data into the pre-trained neural network prediction model to obtain the stator vibration acceleration distribution image and sound pressure distribution image of the motor at future moments; Step 2: Based on the peak vibration acceleration and sound pressure level in the predicted image, compare them with the preset safety threshold to determine the current risk level of the motor and trigger a graded warning. Step 3: Based on the determined risk level, use an optimization algorithm to solve for the optimal adjustment increment of the orthogonal axis current id / iq, and use the current loop to dynamically adjust the motor to achieve active suppression of vibration noise; In step 3, the adjustment of the orthogonal axis current id / iq adopts a constrained dimensionless multi-objective optimization. The core sub-objective is to minimize the NVH peak value within the prediction window. The degree of NVH exceeding the standard is quantified based on the difference between the predicted peak value of future time-domain vibration acceleration, the predicted peak value of noise sound pressure level and the safety level threshold. It also includes the sub-objectives of torque stability, stator current amplitude and motor efficiency loss. An adaptive penalty function method is used to penalize the constraint violation, and the penalty coefficient increases linearly with the algorithm iteration. In step 3, the current regulation follows the following constraints: the stator current amplitude does not exceed 1.1 times the rated current, the instantaneous peak current does not exceed the maximum allowable current of the motor, the steady-state torque fluctuation is ≤2%, the dynamic torque fluctuation is ≤5%, the voltage modulation ratio is ≤0.95, the motor speed fluctuation is ≤±1% of the rated speed, the motor system efficiency decreases by ≤0.5% during the early warning level regulation, and the efficiency decreases by ≤2% during the dangerous level regulation.

2. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 1, characterized in that, The process of establishing the neural network prediction model includes: S1: Generating continuous time series sample data of electromagnetic-structural-acoustic multiphysics coupling of motor based on meshless method; S2: Standardize and preprocess the generated sample data, and define the temporal relationship between the input and the label; S3: Construct an end-to-end prediction network, which integrates a selective bidirectional Mamba modeling module, a physically guided evolution module, and a U-Net encoding / decoding structure; S4: Train the network using a multi-dimensional combined loss function that includes physical consistency constraints to obtain the optimal model weights.

3. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 2, characterized in that, In S1, the moving least squares method is used to construct a field function approximation model to obtain the spatiotemporal distribution of vibration acceleration of the stator core and stator frame, as well as the spatiotemporal distribution of sound pressure of the acoustic radiation field.

4. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 2, characterized in that, In S3, the selective bidirectional Mamba modeling module embeds an acoustic physics prior; the acoustic physics prior is initialized by diagonalizing the state transition matrix through the acoustic transient wave equation of the motor, and the matrix elements are determined by the air acoustic attenuation coefficient, sound speed and sound wave number.

5. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 2, characterized in that, In S3, the physically guided evolution module initializes the acoustic moving least squares form function by embedding it into the weights of the deep convolution kernel, and uses the variational weak form of the acoustic transient wave equation as the physical consistency constraint during the model forward propagation process.

6. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 1, characterized in that, The risk levels are classified as follows: Safety level: Vibration acceleration < 2.5 m / s² 2 And the noise sound pressure level is < 65dB; Warning level: Vibration acceleration is between 2.5 and 3.5 m / s². 2 Or the noise sound pressure level is between 65dB and 75dB; Hazard level: Vibration acceleration > 3.5 m / s² 2 Or the noise sound pressure level is > 75dB.

7. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 1, characterized in that, The weight coefficients of the multi-objective optimization are adaptively adjusted according to the risk level, and the weight switching adopts a 100ms linear ramp transition; among them, the weights for safety level NVH are 0.2, torque is 0.5, current is 0.15, and efficiency is 0.15; the weights for warning level NVH are 0.45, torque is 0.35, current is 0.1, and efficiency is 0.1; and the weights for dangerous level NVH are 0.7, torque is 0.15, current is 0.1, and efficiency is 0.

05.

8. The method for graded early warning and active control of vibration and noise of automotive permanent magnet synchronous motors according to claim 1, characterized in that, Step 3 further includes decomposing the optimal adjustment amount into step amounts for a single control cycle: linear ramp decomposition is used under steady-state conditions, and S-shaped smooth acceleration / deceleration decomposition is used under dynamic conditions to achieve shock-free adjustment.

Images