Method and system for adaptively setting parameters of power controller of electric vehicle

Abstract

The invention provides an electric vehicle power controller parameter adaptive setting method and system, and relates to the technical field of electric vehicle control, and the method comprises the steps: obtaining a motor torque feedback signal, carrying out the wavelet decomposition, extracting a working condition identifier, determining a phase margin and an amplitude margin offset based on Fourier transform, and obtaining a motor torque feedback signal; parameters are adjusted in combination with a gradient projection algorithm and a primal-dual interior point algorithm, and an optimal parameter vector is obtained through a multi-target particle swarm optimization algorithm. According to the invention, on-line adaptive adjustment of the control parameters is realized, and the power performance and the energy efficiency performance of the electric vehicle are improved.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to an adaptive tuning method and system for the power controller parameters of an electric vehicle. Background Technology

[0002] With increasing environmental awareness and a worsening energy crisis, electric vehicles, as a green and efficient mode of transportation, are gradually becoming an important direction for the development of the automotive industry. The power control system is the core component of an electric vehicle, and its performance directly affects the vehicle's power, comfort, and economy. The tuning of the electric vehicle's power controller parameters has a decisive impact on the vehicle's driving performance; traditional parameter tuning methods mainly rely on expert experience and practical testing.

[0003] With the development of electric vehicle technology, the operating conditions faced by power control systems are becoming increasingly complex. Traditional fixed-parameter control methods can no longer meet the optimal control requirements under varying operating conditions. In recent years, adaptive control methods have received widespread attention in the field of electric vehicle power control. By adjusting control parameters in real time, they can adapt to the control needs under different operating conditions, thereby improving the system's adaptability and robustness. However, existing technologies still suffer from problems such as a lack of effective operating condition identification mechanisms, an inability to accurately identify the vehicle's current operating condition, resulting in inaccurate parameter adjustments, a focus on a single indicator of power performance or energy efficiency while neglecting the need for multi-objective optimization, an inability to achieve a good balance between power performance and energy efficiency, high computational complexity, difficulty in achieving real-time adjustments, and an inability to meet the control requirements of electric vehicles in highly dynamic environments. Summary of the Invention

[0004] This invention provides an adaptive tuning method and system for electric vehicle power controller parameters, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of the present invention provides an adaptive tuning method for the parameters of an electric vehicle power controller, comprising: The motor torque feedback signal of the electric vehicle is acquired and multi-scale wavelet decomposition is performed to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features. The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slow-varying component is extracted. Based on the slow-varying component, a steady-state trend feature matrix is ​​constructed and Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and the amplitude margin offset is extracted. According to the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. The gain correction vector is calculated by combining the gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector. The damping correction vector is calculated by the original dual interior point algorithm and applied to the intermediate parameter vector to obtain a candidate parameter set. Based on the motor torque feedback signal and the battery current feedback signal, power performance index and energy efficiency performance index are constructed. The Pareto front search of the candidate parameter set is performed by combining the multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector and load it into the controller. The motor output torque is controlled by combining the pre-acquired accelerator pedal command signal.

[0006] In one alternative implementation, The motor torque feedback signal of the electric vehicle is acquired and subjected to multi-scale wavelet decomposition to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components, and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features, including: The torque measurement value of the motor output shaft is collected by the motor torque sensor and converted from analog to digital to obtain the motor torque feedback signal. The torque sequence is then obtained by performing a sliding window interception. Based on the torque sequence, a wavelet mother function is selected and a multi-level decomposition scale is set to perform discrete wavelet transform to obtain detail coefficients at multiple scales. The energy values ​​of the detail coefficients are calculated and a scale energy distribution vector is constructed. Based on the scale energy distribution vector, key scales are identified and the detail coefficients corresponding to the key scales are extracted to perform wavelet reconstruction to obtain fast change components. The instantaneous amplitude envelope and instantaneous phase corresponding to the fast-changing component are calculated to obtain the time-domain transient features. The Hilbert transform is performed on the fast-changing component to extract the instantaneous frequency sequence and the frequency change rate is calculated to obtain the frequency-domain transient features. The time-domain transient features and the frequency-domain transient features are time-frequency synchronized to obtain the aligned feature sequence and tensor reconstruction is performed to obtain the three-dimensional time-frequency feature tensor. The three-dimensional time-frequency feature tensor is subjected to Tucker decomposition to obtain the core tensor and mode factor matrix and tensor product reconstruction is performed to obtain the transient response feature matrix. Singular value decomposition is performed on the transient response feature matrix to construct dominant mode features. Based on the dominant mode features, a working condition feature descriptor is constructed and matched with a preset working condition pattern library using distance metric matching to determine the working condition identifier.

[0007] In one alternative implementation, The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform, and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slowly varying component is extracted. Based on the slowly varying component, a steady-state trend feature matrix is ​​constructed and a Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset, including: Windowing and fast Fourier transform are performed on the transient response feature matrix to obtain a frequency domain complex matrix. The amplitude spectrum and phase spectrum are calculated based on the frequency domain complex matrix. Phase unwrapping is performed on the phase spectrum to obtain a continuous phase curve and the phase crossing frequency is identified. The phase margin offset is calculated based on the amplitude spectrum at the phase crossing frequency. The current measurement value is collected and converted from analog to digital to obtain the battery current feedback signal. The battery current feedback signal is subjected to sliding window truncation to obtain the current sequence and low-pass filtering to obtain the low-frequency component. The low-frequency component is subjected to moving average smoothing to obtain the trend baseline. Based on the trend baseline, the segment is divided and the local slope and local mean in each segment are calculated to construct the segmented trend feature vector as the slow variable component. Based on the slowly varying components, a state vector is constructed. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of state posterior mean values. This sequence is then expanded over time. Finally, the enhanced slowly varying feature matrix is ​​obtained by combining the segmented trend feature vector with the solution. The enhanced slowly varying feature matrix is ​​aligned with a time window and zero-padding is applied to obtain a steady-state trend feature matrix. The steady-state trend feature matrix is ​​then subjected to a fast Fourier transform to obtain a steady-state frequency domain feature spectrum and a steady-state phase spectrum is calculated. The phase values ​​corresponding to the steady-state phase spectrum are extracted and the magnitude margin offset is calculated.

[0008] In one alternative implementation, A state vector is constructed based on the slowly varying components. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of posterior mean values. This sequence is then expanded over time. The enhanced slowly varying feature matrix is ​​obtained by combining this with the segmented trend feature vector. The slowly varying components are sorted according to time order, and the local slope and local mean corresponding to each time point are extracted. The state vector is constructed based on the combination of the local slope and the local mean. The prior mean and prior covariance are set for the state vector to obtain the prior probability distribution. Based on the prior probability distribution, define the state transition matrix and observation matrix and establish the state space equation. Based on the prior probability distribution and the pre-acquired observation data, derive the posterior probability distribution of the state vector. Perform iterative optimization of maximizing the lower bound of evidence on the posterior probability distribution using the variational Bayesian inference algorithm. Update the state prediction based on the state transition matrix and update the observation prediction in combination with the observation matrix. Repeat the iteration until the posterior mean and posterior covariance of the state converge. The time dimension is expanded to obtain expanded state features by performing time dimension expansion on the posterior mean sequence of the state. The difference between adjacent time points is calculated based on the expanded state features to obtain the time derivative sequence. The time derivative sequence is normalized to obtain the dynamic rate of change vector. The dynamic rate of change vector and the slow-varying component are matched with timestamps to obtain aligned feature pairs. Based on the aligned features, a fused feature vector is calculated, and the fused feature vector is matrix reorganized according to the time dimension and the feature dimension to obtain the enhanced slow-varying feature matrix.

[0009] In one alternative implementation, Based on the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. A gain correction vector is calculated using a gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector, including: Based on the operating condition identifier, an index query is performed in the preset benchmark parameter library to extract the benchmark gain parameter and benchmark time constant parameter corresponding to the operating condition identifier from the benchmark parameter library and combine them to construct the initial parameter vector; The phase margin offset is converted into a phase constraint boundary, the magnitude margin offset is converted into a magnitude constraint boundary, the constraint boundary range is determined based on the phase constraint boundary and the magnitude constraint boundary, a linear constraint matrix is ​​constructed based on the constraint boundary range, and the parameter adjustment constraint space is defined based on the linear constraint matrix. The gradient vector is obtained by calculating the partial derivatives of each component of the initial parameter vector. The gradient vector is then projected onto the parameter adjustment constraint space to obtain the constraint gradient vector. Based on the constraint gradient vector and the linear constraint matrix, the gain correction vector is obtained by performing iterative update calculation through the gradient projection algorithm. The gain correction vector is added element-wise to the initial parameter vector to obtain the corrected parameter vector. The corrected parameter vector is then subjected to constraint boundary detection and truncation to obtain the intermediate parameter vector.

[0010] In one alternative implementation, The damping correction vector is calculated using the original dual interior point algorithm and applied to the intermediate parameter vector to obtain the candidate parameter set, which includes: Extract the linear constraint matrix and constraint boundary range in the parameter adjustment constraint space to construct the feasible region. Perform feasible region projection mapping on the intermediate parameter vector and calculate the shortest distance from the intermediate parameter vector to the feasible region boundary to obtain the relaxation variable vector. The original variable vector is constructed based on the intermediate parameter vector and the relaxation variable vector, and the dual variable vector is constructed based on the transpose of the linear constraint matrix. The original variable vector and the dual variable vector are combined to construct the extended variable vector, and strict interior point initialization detection is performed to obtain the initial extended variable vector. The initial extended variable vector is iteratively updated using the primal-dual interior point algorithm. The complementary gap is calculated based on the current extended variable vector. The first and second partial derivatives of the current extended variable vector are calculated to construct the gradient vector and Hessian matrix. The Newton direction vector is solved based on the gradient vector and the Hessian matrix, and the optimal step size factor is obtained by performing a step size optimization search. The current extended variable vector is updated along the Newton direction vector based on the optimal step size factor until the complementary gap converges, thus obtaining the optimized parameter vector. The damping coefficient adjustment and time constant adjustment are extracted from the optimized parameter vector to construct a damping correction vector. Based on the damping correction vector and the intermediate parameter vector, the updated parameter vector is obtained and a candidate parameter set is constructed.

[0011] In one alternative implementation, Based on the motor torque feedback signal and battery current feedback signal, power performance indicators and energy efficiency performance indicators are constructed. A Pareto front search is performed on the candidate parameter set using a multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector, which is then loaded into the controller. The motor output torque is controlled in conjunction with the pre-acquired accelerator pedal command signal, including: Perform time window statistical analysis on the motor torque feedback signal, calculate the torque response time and torque overshoot, and construct the power performance index. Perform integral operation on the battery current feedback signal, calculate the energy consumption ratio in combination with the driving distance, and construct the energy efficiency performance index. Extract each candidate parameter vector from the candidate parameter set and calculate the bi-objective fitness vector corresponding to each candidate parameter vector by combining the power performance index and the energy efficiency performance index. Perform a dominance relationship determination on any two candidate parameter vectors in the candidate parameter set to obtain the initial solution set of the Pareto front. Initialize the particle swarm and assign the parameter vector from the initial Pareto front solution set to each particle. Iteratively update the particle velocity and position based on the multi-objective particle swarm optimization algorithm. Combine the bi-objective fitness vector to calculate the individual optimal position and global optimal position of each particle and update the particle velocity and position again. Based on the updated particle velocity and position, perform dominance relationship determination and update the initial Pareto front solution set to obtain the Pareto front solution set and extract the optimal solution to obtain the optimal parameter vector. The optimal parameter vector is loaded into the controller, the pre-acquired accelerator pedal command signal is obtained and the target torque requirement is calculated, and the motor control command is generated based on the optimal parameter vector and the target torque requirement to control the motor output torque.

[0012] A second aspect of the present invention provides an adaptive tuning system for the parameters of an electric vehicle power controller, comprising: The feature extraction unit is used to acquire the motor torque feedback signal of the electric vehicle and perform multi-scale wavelet decomposition to obtain fast-changing components, construct a transient response feature matrix based on the fast-changing components and extract the corresponding dominant mode features, and determine the operating condition identifier based on the dominant mode features. The margin analysis unit is used to determine the transient frequency domain feature spectrum corresponding to the transient response feature matrix through Fourier transform and extract the phase margin offset, obtain the battery current feedback signal of the electric vehicle and extract the slow-varying component, construct the steady-state trend feature matrix based on the slow-varying component and perform Fourier transform to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset. The parameter optimization unit is used to extract an initial parameter vector from a preset reference parameter library according to the operating condition identifier, construct a parameter adjustment constraint space based on the phase margin offset and the gain margin offset, calculate a gain correction vector by combining the gradient projection algorithm and apply it to the initial parameter vector to obtain an intermediate parameter vector, and calculate a damping correction vector by using the original dual interior point algorithm and apply it to the intermediate parameter vector to obtain a candidate parameter set. The strategy execution unit is used to construct power performance indicators and energy efficiency performance indicators based on motor torque feedback signals and battery current feedback signals. It combines a multi-objective particle swarm optimization algorithm to perform Pareto front search on the candidate parameter set, obtain the optimal parameter vector, and load it into the controller. It then controls the motor output torque in conjunction with the pre-acquired accelerator pedal command signal.

[0013] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this invention, fast-changing components are obtained by multi-scale wavelet decomposition of the motor torque feedback signal, and operating condition identifiers are determined based on the dominant mode characteristics, thus achieving accurate identification of different operating conditions and improving the pertinence and adaptability of parameter tuning. By combining the feedback signals of motor torque and battery current, phase margin offset and amplitude margin offset are extracted respectively, constructing a complete parameter adjustment constraint space, ensuring the stability and robustness of the control system. By combining the gradient projection algorithm with the original dual interior point algorithm, the synergistic optimization of gain correction and damping correction is achieved, avoiding the problem of local optima. Based on the power performance and energy efficiency performance indicators, the Pareto front search is performed using the particle swarm optimization algorithm to achieve a balanced optimization of power performance and energy efficiency performance, meeting the optimal control effect under different driving needs. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the adaptive tuning method for electric vehicle power controller parameters according to an embodiment of the present invention. Figure 2 This is a flowchart of the original-dual interior-point method for parameter optimization in the adaptive tuning method of electric vehicle power controller parameters according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the adaptive tuning method for electric vehicle power controller parameters according to an embodiment of the present invention. Figure 1 As shown, the method includes: The motor torque feedback signal of the electric vehicle is acquired and multi-scale wavelet decomposition is performed to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features. The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slow-varying component is extracted. Based on the slow-varying component, a steady-state trend feature matrix is ​​constructed and Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and the amplitude margin offset is extracted. According to the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. The gain correction vector is calculated by combining the gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector. The damping correction vector is calculated by the original dual interior point algorithm and applied to the intermediate parameter vector to obtain a candidate parameter set. Based on the motor torque feedback signal and the battery current feedback signal, power performance index and energy efficiency performance index are constructed. The Pareto front search of the candidate parameter set is performed by combining the multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector and load it into the controller. The motor output torque is controlled by combining the pre-acquired accelerator pedal command signal.

[0020] In one alternative implementation, The motor torque feedback signal of the electric vehicle is acquired and subjected to multi-scale wavelet decomposition to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components, and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features, including: The torque measurement value of the motor output shaft is collected by the motor torque sensor and converted from analog to digital to obtain the motor torque feedback signal. The torque sequence is then obtained by performing a sliding window interception. Based on the torque sequence, a wavelet mother function is selected and a multi-level decomposition scale is set to perform discrete wavelet transform to obtain detail coefficients at multiple scales. The energy values ​​of the detail coefficients are calculated and a scale energy distribution vector is constructed. Based on the scale energy distribution vector, key scales are identified and the detail coefficients corresponding to the key scales are extracted to perform wavelet reconstruction to obtain fast change components. The instantaneous amplitude envelope and instantaneous phase corresponding to the fast-changing component are calculated to obtain the time-domain transient features. The Hilbert transform is performed on the fast-changing component to extract the instantaneous frequency sequence and the frequency change rate is calculated to obtain the frequency-domain transient features. The time-domain transient features and the frequency-domain transient features are time-frequency synchronized to obtain the aligned feature sequence and tensor reconstruction is performed to obtain the three-dimensional time-frequency feature tensor. The three-dimensional time-frequency feature tensor is subjected to Tucker decomposition to obtain the core tensor and mode factor matrix and tensor product reconstruction is performed to obtain the transient response feature matrix. Singular value decomposition is performed on the transient response feature matrix to construct dominant mode features. Based on the dominant mode features, a working condition feature descriptor is constructed and matched with a preset working condition pattern library using distance metric matching to determine the working condition identifier.

[0021] Motor torque data is acquired in real time by a torque sensor installed on the motor output shaft at a sampling frequency of 10,000 Hz. The analog signal is converted into a digital signal by a 16-bit analog-to-digital converter to form a motor torque feedback signal. A sliding window operation is performed on the motor torque feedback signal, with a window length of 1024 sampling points and a sliding step size of 128 sampling points, to obtain the torque time series.

[0022] Multi-scale analysis was performed based on the acquired torque sequence. The db4 wavelet mother function was selected as the basic analysis tool, and a discrete wavelet transform was performed with six decomposition scales. The transform yielded a set of detail coefficients (d1 to d6) and an approximate coefficient set at six different scales. The energy value of the detail coefficients at each scale was calculated by summing the squares of each element. Taking a specific test as an example, the energy values ​​at the six scales were 12.75, 38.24, 24.61, 9.33, 3.45, and 1.62, forming an energy distribution vector. Key scales were identified based on an energy percentage exceeding 15%, with the second and third scales identified as key scales, corresponding to a frequency range of 625 to 2500 Hz. Subsequently, the detail coefficients corresponding to these two key scales were extracted, and wavelet reconstruction was performed to obtain the fast-changing components in the motor torque signal.

[0023] Time-frequency features are extracted from the extracted fast-changing components, and the instantaneous amplitude envelope of the fast-changing components is calculated. The amplitude envelope curve is obtained by calculating the sum of squares at each time point of the fast-changing component, adding a weighted average with the sum of squares of adjacent points, and then taking the square root. Simultaneously, the instantaneous phase change is calculated, yielding the phase-time curve as a time-domain transient feature. In the frequency domain analysis, a Hilbert transform is performed on the fast-changing components, and the instantaneous frequency sequence is calculated by dividing the phase difference between adjacent sampling points by the sampling time interval. This instantaneous frequency sequence reflects the frequency variation over time. Furthermore, the rate of change of the instantaneous frequency, i.e., the first-order difference of frequency with respect to time, is calculated to obtain the frequency rate of change sequence, which serves as the frequency-domain transient feature.

[0024] To achieve collaborative analysis of time-frequency information, time-domain transient features and frequency-domain transient features are synchronized and aligned. The four sequences—amplitude envelope, phase change, instantaneous frequency, and rate of change of frequency—are combined according to their corresponding sampling time points to form an aligned feature sequence. The aligned feature sequence contains 1024 time points, each corresponding to 4 feature values. Tensor reconstruction is performed to organize it into a three-dimensional time-frequency feature tensor with dimensions of 1024×4×1.

[0025] Tucker decomposition was performed on the constructed 3D time-frequency feature tensor with a rank of 4×3×1, yielding a core tensor and three modality factor matrices. The core tensor reflects the interaction relationships between the three dimensions. The first modality factor matrix represents the principal components of the time dimension (1024×4), the second represents the principal components of the feature dimension (4×3), and the third is the identity matrix. Tensor product reconstruction was performed using the core tensor and modality factor matrices to obtain a compressed transient response feature matrix with a dimension of 1024×3, preserving the main information of the original data.

[0026] Singular value decomposition is performed on the reconstructed transient response feature matrix to obtain the left singular matrix, singular value vector, and right singular matrix. The left singular vector corresponding to the two largest singular values ​​is selected as the dominant mode feature. Based on the dominant mode feature, a working condition feature descriptor is constructed. The mean, standard deviation, kurtosis, and skewness of the dominant mode feature vector are calculated and concatenated into a single feature vector.

[0027] A pre-established driving condition pattern library contains feature descriptors for typical driving conditions such as acceleration, deceleration, hill climbing, flat-road cruising, and steering. The similarity between the current feature descriptor and the feature descriptors of each driving condition in the pattern library is calculated using Euclidean distance, and the driving condition with the smallest similarity is selected as the recognition result. In one test, the current driving condition had a distance of 0.187 with the acceleration driving condition, 0.563 with the deceleration driving condition, and 0.329 with the hill climbing driving condition; therefore, the current driving condition was ultimately identified as the acceleration driving condition.

[0028] In this embodiment, by introducing a key scale adaptive selection mechanism based on wavelet energy distribution into the torque signal, the intrinsic scale information corresponding to sudden changes in operating conditions and rapid load changes can be highlighted, effectively suppressing non-critical information and background disturbances, and improving the discriminability of transient features. By simultaneously characterizing the instantaneous amplitude, instantaneous phase, and instantaneous frequency change rate of the fast-changing components and synchronizing them in the time-frequency dimension, a unified three-dimensional time-frequency feature representation is constructed, enabling different types of transient features to achieve collaborative expression in the same feature space. This avoids the information mismatch problem caused by asynchronous multi-source features and enhances the overall characterization capability of complex dynamic operating conditions. By extracting the dominant correlation structure across time, frequency, and feature modes through tensor decomposition and reconstruction, noise and redundant dimensions are effectively compressed, highlighting the core response mode that is highly correlated with changes in operating conditions, so that the transient response features have stronger discriminative power while maintaining physical meaning.

[0029] In one alternative implementation, The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform, and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slowly varying component is extracted. Based on the slowly varying component, a steady-state trend feature matrix is ​​constructed and a Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset, including: Windowing and fast Fourier transform are performed on the transient response feature matrix to obtain a frequency domain complex matrix. The amplitude spectrum and phase spectrum are calculated based on the frequency domain complex matrix. Phase unwrapping is performed on the phase spectrum to obtain a continuous phase curve and the phase crossing frequency is identified. The phase margin offset is calculated based on the amplitude spectrum at the phase crossing frequency. The current measurement value is collected and converted from analog to digital to obtain the battery current feedback signal. The battery current feedback signal is subjected to sliding window truncation to obtain the current sequence and low-pass filtering to obtain the low-frequency component. The low-frequency component is subjected to moving average smoothing to obtain the trend baseline. Based on the trend baseline, the segment is divided and the local slope and local mean in each segment are calculated to construct the segmented trend feature vector as the slow variable component. Based on the slowly varying components, a state vector is constructed. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of state posterior mean values. This sequence is then expanded over time. Finally, the enhanced slowly varying feature matrix is ​​obtained by combining the segmented trend feature vector with the solution. The enhanced slowly varying feature matrix is ​​aligned with a time window and zero-padding is applied to obtain a steady-state trend feature matrix. The steady-state trend feature matrix is ​​then subjected to a fast Fourier transform to obtain a steady-state frequency domain feature spectrum and a steady-state phase spectrum is calculated. The phase values ​​corresponding to the steady-state phase spectrum are extracted and the magnitude margin offset is calculated.

[0030] Windowing is applied to the transient response feature matrix using a Hanning window function to reduce spectral leakage. The window length is 1024 sampling points, and the window overlap rate is 50%. The windowed data undergoes a Fast Fourier Transform (FFT) with 2048 transform points to obtain a frequency domain complex matrix. The amplitude and phase spectra are calculated based on this matrix. The amplitude spectrum is obtained by calculating the modulus of the complex elements, and the phase spectrum is obtained by calculating the argument of the complex elements. Phase unwinding is performed on the phase spectrum. The unwinding algorithm employs a phase continuity constraint method: when the absolute value of the phase difference between adjacent frequency points is greater than π, an appropriate multiple of 2π is added to make the phase change continuous, resulting in a continuous phase curve. The phase crossing frequency is identified on the continuous phase curve, i.e., the frequency corresponding to the intersection of the phase curve and the -180-degree line. For example, under a certain test condition, the phase crossing frequency is 157 Hz. Based on the amplitude spectrum value of 0.67 at the phase crossing frequency, the phase margin offset is calculated to be 3.47 dB.

[0031] The current measurements of the electric vehicle battery pack were collected at a sampling frequency of 1000 Hz and converted into digital signals using a 16-bit analog-to-digital converter to obtain the battery current feedback signal. A sliding window truncation operation was performed on the battery current feedback signal, with a window length of 2000 sampling points and a sliding step size of 200 sampling points, to obtain the current sequence. A Butterworth low-pass filter was applied to the current sequence with a filter order of 4 and a cutoff frequency of 10 Hz to obtain the low-frequency component of the current signal. A 51-point moving average smoothing process was then applied to the low-frequency component to eliminate residual high-frequency noise, obtaining the current trend baseline. Adaptive segmentation was performed based on the trend baseline's change characteristics. When the trend slope change exceeded a set threshold of 0.05 amperes per second, a segmentation point was marked. In actual testing, the 2000-point data segment was divided into 5 sub-segments with lengths of 356, 427, 512, 389, and 316 points, respectively. The local slope and local mean of the current value within each segment were calculated to construct a segmented trend feature vector, which serves as the slowly varying component of the current signal.

[0032] A state vector is constructed based on the slowly varying components of the current signal, containing three state variables: mean current, current slope, and current fluctuation. A variational Bayesian inference algorithm is used to iteratively optimize the posterior probability distribution of the state vector. The algorithm sets the prior distribution to Gaussian, the prior mean vector to 0, and the prior covariance matrix to be the identity matrix multiplied by 0.01. During iteration, the convergence threshold is set to 0.0001, and the maximum number of iterations is 100. After iterative optimization, a posterior mean sequence of states is obtained. Under the acceleration condition of an electric vehicle, the mean current in the posterior state mean sequence increases from 12.6 amps to 47.8 amps, the current slope decreases from 0.43 amps per second to 0.12 amps per second, and the current fluctuation decreases from 0.68 to 0.23. The posterior mean sequence of states is expanded along the time dimension and combined with the piecewise trend feature vector calculated in the previous example. A weighted fusion method is used to obtain the enhanced slowly varying feature matrix, with a matrix dimension of 2000×3.

[0033] A time window alignment is performed on the enhanced slowly varying feature matrix, with the window length being the same as that of the transient response feature matrix, both being 1024 points. Zero-padding is applied to any insufficient length to obtain a steady-state trend feature matrix with uniform dimensions. A Fast Fourier Transform (FFT) is then performed on the steady-state trend feature matrix, with 2048 points transformed, yielding the steady-state frequency domain feature spectrum. The steady-state phase spectrum is calculated based on the frequency domain feature spectrum, and phase values ​​in the low-frequency range of 0.1 to 10 Hz are extracted. For example, at the point where the amplitude reaches 0.707 times the maximum value, the corresponding frequency is 1.53 Hz, and the phase delay at this point is -43.2 degrees. Based on this, the amplitude margin offset is calculated to be 6.83 dB.

[0034] In this embodiment, by windowing and performing frequency domain complex modeling on the transient response feature matrix, and introducing a phase crossover frequency identification mechanism based on phase unwrapping, the key phase change positions of the system under dynamic excitation can be accurately captured, improving the reliability of transient stability assessment under actual working conditions. By extracting low-frequency components of the battery current signal and performing piecewise trend modeling, the slow-evolving steady-state characteristic changes during long-term operation of the system can be effectively characterized, avoiding the interference of instantaneous fluctuations on trend judgment. Combining variational Bayesian inference to perform probabilistic modeling of slow-changing states, the steady-state characteristics maintain continuity and consistency even in noisy environments, significantly improving the robustness and reliability of slow-changing characteristics.

[0035] In one alternative implementation, A state vector is constructed based on the slowly varying components. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of posterior mean values. This sequence is then expanded over time. The enhanced slowly varying feature matrix is ​​obtained by combining this with the segmented trend feature vector. The slowly varying components are sorted according to time order, and the local slope and local mean corresponding to each time point are extracted. The state vector is constructed based on the combination of the local slope and the local mean. The prior mean and prior covariance are set for the state vector to obtain the prior probability distribution. Based on the prior probability distribution, define the state transition matrix and observation matrix and establish the state space equation. Based on the prior probability distribution and the pre-acquired observation data, derive the posterior probability distribution of the state vector. Perform iterative optimization of maximizing the lower bound of evidence on the posterior probability distribution using the variational Bayesian inference algorithm. Update the state prediction based on the state transition matrix and update the observation prediction in combination with the observation matrix. Repeat the iteration until the posterior mean and posterior covariance of the state converge. The time dimension is expanded to obtain expanded state features by performing time dimension expansion on the posterior mean sequence of the state. The difference between adjacent time points is calculated based on the expanded state features to obtain the time derivative sequence. The time derivative sequence is normalized to obtain the dynamic rate of change vector. The dynamic rate of change vector and the slow-varying component are matched with timestamps to obtain aligned feature pairs. Based on the aligned features, a fused feature vector is calculated, and the fused feature vector is matrix reorganized according to the time dimension and the feature dimension to obtain the enhanced slow-varying feature matrix.

[0036] The slowly varying components are sorted according to time sequence, based on the sampling timestamp, with a sampling frequency of 1000 Hz. Local slopes and local means are calculated for the sorted slowly varying components. The local slope is calculated by dividing the change in current value at 50 adjacent time points by the corresponding time interval, in amperes per second (A / s). The local mean is obtained by calculating the average current value at 50 adjacent time points, in amperes. Under the starting condition of a certain electric vehicle, the local slope gradually increases from 0.02 A / s to 5.73 A / s, and the local mean gradually increases from 2.45 A / s to 38.62 A / s. A state vector is constructed based on the combination of local slope and local mean. The state vector has two dimensions, where the first dimension is the local mean and the second dimension is the local slope. Prior means and prior covariance are defined for the state vector. The prior mean is set as a zero vector, and the prior covariance matrix is ​​set as a diagonal matrix with diagonal elements of 0.01, resulting in the prior probability distribution.

[0037] Based on prior probability distributions, a state transition matrix and an observation matrix are defined, and a state-space equation is established. The state-space equation consists of two parts: the state equation and the observation equation. The state equation is expressed as the current state vector equals the state transition matrix multiplied by the previous state vector, plus process noise. The observation equation is expressed as the current observation equals the observation matrix multiplied by the current state vector, plus observation noise. The state vector contains two components: local mean and local slope. The state transition matrix is ​​a 2×2 matrix, with diagonal elements set to 0.98, indicating temporal continuity of the states; off-diagonal elements are set to 0.05, indicating weak correlation between states. The process noise follows a normal distribution with a mean of 0 and a covariance matrix of 0.0025. The observation matrix is ​​a 1×2 matrix with element values ​​of 1 and 0.1, indicating that the observation is mainly determined by the local mean, with the local slope playing a secondary role; the observation noise follows a normal distribution with a mean of 0 and a variance of 0.01.

[0038] Based on the prior probability distribution and pre-acquired observation data, the posterior probability distribution of the state vector is derived using Bayes' theorem. A variational Bayesian inference algorithm is then employed to perform iterative optimization by maximizing the lower bound of evidence on the posterior probability distribution. During the iteration process, the predicted mean and covariance of the state are first calculated based on the state transition matrix, and then updated by combining the observation matrix and newly acquired observations. The state prediction process uses the state transition matrix to perform forward propagation of the previous state, calculating the prior estimate of the current state. The observation update process combines the observation matrix and actual observations to correct the state prediction results, obtaining the posterior estimate of the current state. An iteration convergence threshold of 0.0001 is set; convergence is considered achieved when the change in the posterior mean of the state between two iterations is less than this threshold. For example, in the acceleration test of an electric vehicle, the algorithm converged after 27 iterations, and the resulting sequence of posterior means accurately reflects the current change trend.

[0039] A time-dimensional expansion is performed on the posterior mean sequence of states, expanding the posterior mean of the states at each time point into feature values ​​for the corresponding time, resulting in expanded state features. The dimension of the expanded state features is the same as that of the original observation data, which is 2000 data points in the test. The difference between adjacent time points is calculated based on the expanded state features, by subtracting the state value of the previous time point from the state value of the subsequent time point, resulting in a time derivative sequence. Normalization is then performed on the time derivative sequence using a max-min scaling method, mapping the data to a range of 0 to 1, resulting in a dynamic rate of change vector. In practical applications, the dynamic rate of change vector reflects the relative magnitude of the current change rate; the dynamic rate of change reaches 0.95 during the initial startup phase of an electric vehicle, decreasing to approximately 0.15 during stable operation.

[0040] Timestamp matching is performed on the dynamic rate of change vector and the slow-varying component. The matching process finds the closest data points by comparing the timestamp differences, resulting in aligned feature pairs. The timestamp matching accuracy is controlled within 1 millisecond to ensure the temporal consistency of the feature pairs. A fused feature vector is calculated based on the aligned features using a weighted summation method, with a weight allocation of 0.7 for the slow-varying component and 0.3 for the dynamic rate of change. The fused feature vector is then matrix-reorganized according to the time and feature dimensions to obtain an enhanced slow-varying feature matrix. The enhanced slow-varying feature matrix has 2000 rows (equal to the original number of data points) and 3 columns, containing the original slow-varying component, dynamic rate of change, and fused features.

[0041] In this embodiment, by organizing the slowly varying components in chronological order and introducing local slopes and local means to jointly construct the state vector, the state description not only reflects the absolute level of the slowly varying signal but also simultaneously characterizes its direction and intensity of change. This avoids the problem that a single mean feature is insufficient to distinguish different evolution paths, improving the ability to distinguish the slowly varying characteristics of different operating stages. Variational Bayesian inference is used to iteratively optimize the posterior distribution of the state, enabling the state estimation process to maintain continuity and consistency under noise and incomplete observation conditions. It can adaptively balance the influence of prior information and observation information, significantly reducing the interference of random fluctuations on the results of slowly varying feature extraction, and improving the robustness and reliability of slowly varying state estimation. By performing time expansion on the converged posterior mean of the state and introducing a dynamic rate of change representation, the slowly varying features simultaneously contain steady-state level information and evolution rate information, thereby enhancing the sensitivity to long-term changes such as slow drift and gradual degradation, and reflecting the potential trend of system state changes earlier and more accurately.

[0042] In one alternative implementation, Based on the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. A gain correction vector is calculated using a gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector, including: Based on the operating condition identifier, an index query is performed in the preset benchmark parameter library to extract the benchmark gain parameter and benchmark time constant parameter corresponding to the operating condition identifier from the benchmark parameter library and combine them to construct the initial parameter vector; The phase margin offset is converted into a phase constraint boundary, the magnitude margin offset is converted into a magnitude constraint boundary, the constraint boundary range is determined based on the phase constraint boundary and the magnitude constraint boundary, a linear constraint matrix is ​​constructed based on the constraint boundary range, and the parameter adjustment constraint space is defined based on the linear constraint matrix. The gradient vector is obtained by calculating the partial derivatives of each component of the initial parameter vector. The gradient vector is then projected onto the parameter adjustment constraint space to obtain the constraint gradient vector. Based on the constraint gradient vector and the linear constraint matrix, the gain correction vector is obtained by performing iterative update calculation through the gradient projection algorithm. The gain correction vector is added element-wise to the initial parameter vector to obtain the corrected parameter vector. The corrected parameter vector is then subjected to constraint boundary detection and truncation to obtain the intermediate parameter vector.

[0043] Based on the identified operating condition identifier, an index query is performed in a pre-defined baseline parameter library. The baseline parameter library is stored in the form of a hash table, with the operating condition identifier as the key and the corresponding parameter set as the value. The baseline parameter library contains optimal control parameters for common operating conditions, such as acceleration, deceleration, hill climbing, flat-road cruising, and steering. Baseline gain parameters and baseline time constant parameters corresponding to the current operating condition identifier are extracted from the baseline parameter library. Baseline gain parameters include the speed loop proportional gain and torque loop proportional gain; baseline time constant parameters include the speed loop integral time constant and torque loop integral time constant. In a practical application, an electric vehicle identifies the current operating condition as a rapid deceleration condition, with the corresponding baseline gain parameters being: speed loop proportional gain 5.80, torque loop proportional gain 6.70; and the baseline time constant parameters being: speed loop integral time constant 0.025 seconds, torque loop integral time constant 0.028 seconds. These parameters are combined to construct an initial parameter vector with a dimension of 4.

[0044] The phase margin offset calculated in the previous example is converted into phase constraint boundaries. The conversion method is to multiply the phase margin offset by a sensitivity factor of 0.35 to obtain the phase boundary offset value. This phase boundary offset value is then added to the reference phase margin value of 45 degrees to obtain the upper boundary of the phase constraint. The phase boundary offset value is then subtracted from the reference phase margin value to obtain the lower boundary of the phase constraint. For a measured phase margin offset of 4.86 dB, the calculated phase boundary offset value is 1.70 degrees, the upper boundary of the phase constraint is 46.70 degrees, and the lower boundary of the phase constraint is 43.30 degrees. The gain margin offset is then converted into gain constraint boundaries. The conversion method is to multiply the gain margin offset by a sensitivity factor of 0.40 to obtain the gain boundary offset value. This gain boundary offset value is then added to the reference gain margin value of 6 dB to obtain the upper boundary of the gain constraint. The gain boundary offset value is then subtracted from the reference gain margin value to obtain the lower boundary of the gain constraint. For a measured amplitude margin offset of 7.95 dB, the calculated amplitude boundary offset is 3.18 dB, the upper boundary of the amplitude constraint is 9.18 dB, and the lower boundary of the amplitude constraint is 2.82 dB. It should be noted that the sensitivity factors corresponding to the phase margin offset and amplitude margin offset need to be determined jointly based on actual requirements and the electric vehicle, and are not fixed values.

[0045] The range of constraint boundaries for parameter adjustment is determined based on the calculated phase and magnitude constraint boundaries. According to the stability criterion in control theory, the phase margin and magnitude margin together constitute the stability constraints. These stability constraints are transformed into linear inequality constraints in the parameter space, constructing a linear constraint matrix. Each row of the linear constraint matrix corresponds to a constraint condition, and each column corresponds to a parameter in the initial parameter vector. In this embodiment, the linear constraint matrix has a 4×4 dimension. The first row represents the upper limit constraint of the speed loop proportional gain, the second row represents the lower limit constraint of the speed loop proportional gain, the third row represents the upper limit constraint of the torque loop proportional gain, and the fourth row represents the lower limit constraint of the torque loop proportional gain. A parameter adjustment constraint space is defined based on the linear constraint matrix; this constraint space is a subspace within the four-dimensional parameter space that satisfies all constraint conditions.

[0046] The partial derivatives of each component of the initial parameter vector are calculated to obtain the gradient vector. The partial derivative calculation uses the central difference approximation method. For each parameter, a small perturbation value of 0.05 is added and subtracted, and the change in system response is calculated. This change is then divided by twice the perturbation value to obtain the approximate partial derivative. For example, the partial derivative of the speed loop proportional gain is calculated by increasing the initial value from 5.80 to 5.85 and decreasing it to 5.75, calculating the system response, and dividing the difference by 0.10 to obtain the partial derivative. The partial derivative values ​​calculated for the four parameters are: partial derivative of speed loop proportional gain 1.25, partial derivative of speed loop integral time constant -3.10, partial derivative of torque loop proportional gain 0.92, and partial derivative of torque loop integral time constant -2.35, which together form the gradient vector.

[0047] The gradient vector is projected onto the parameter adjustment constraint space to obtain the constraint gradient vector. The projection method projects the original gradient vector onto the feasible region defined by the linear constraint matrix, ensuring that the parameter adjustment direction does not violate the constraint conditions. Specifically, a gradient projection algorithm is used. First, the orthogonal projection of the gradient vector onto the constraint plane is calculated, and then the projection result is adjusted according to the constraint conditions. In this embodiment, the calculated values ​​of the constraint gradient vector are: speed loop proportional gain gradient 0.98, speed loop integral time constant gradient -2.45, torque loop proportional gain gradient 0.75, and torque loop integral time constant gradient -1.90.

[0048] Based on the constraint gradient vector and the linear constraint matrix, an iterative update calculation is performed using a gradient projection algorithm to obtain the gain correction vector. During the iteration, the step size factor is set to 0.05, the maximum number of iterations is 50, and the convergence threshold is 0.001. In each iteration, the current parameter vector is moved along the constraint gradient direction, and it is checked whether all constraints are satisfied. If not, projection correction is performed. Convergence is achieved after 32 iterations, and the obtained gain correction vector is as follows: speed loop proportional gain correction value 0.45, speed loop integral time constant correction value -0.012, torque loop proportional gain correction value 0.38, and torque loop integral time constant correction value -0.015.

[0049] The gain correction vector is element-wise added to the initial parameter vector to obtain the corrected parameter vector. The calculation results are: corrected speed loop proportional gain 6.25, corrected speed loop integral time constant 0.013, corrected torque loop proportional gain 7.08, and corrected torque loop integral time constant 0.013. Constraint boundary checks are performed on the corrected parameter vector to verify that each parameter is within the preset hard constraint range. The hard constraint ranges are: speed loop proportional gain 3.0 to 6.0, speed loop integral time constant 0.02 to 0.10, torque loop proportional gain 4.0 to 7.0, and torque loop integral time constant 0.02 to 0.08.

[0050] Constraint boundary checks revealed several parameters exceeding the constraint range: the corrected speed loop proportional gain of 6.25 exceeded the upper limit of 6.0; the corrected speed loop integral time constant of 0.013 was below the lower limit of 0.02; the corrected torque loop proportional gain of 7.08 exceeded the upper limit of 7.0; and the corrected torque loop integral time constant of 0.013 was below the lower limit of 0.02. Therefore, truncation was required to cut off the out-of-range parameter values ​​to their corresponding boundary values. The truncation resulted in the following intermediate parameter vector: speed loop proportional gain 6.0, speed loop integral time constant 0.02, torque loop proportional gain 7.0, and torque loop integral time constant 0.02.

[0051] In this embodiment, by retrieving the corresponding benchmark gain parameter and time constant parameter from the benchmark parameter library based on the operating condition identifier, the parameter adjustment has a clear starting point, avoiding the problem of arbitrary selection of initial parameter values ​​or deviation from the actual operating condition. This shortens the parameter convergence process and reduces the probability of unreasonable parameter combinations, improving the reliability of the parameter adjustment process. By converting the phase margin offset and amplitude margin offset into explicit phase and amplitude constraint boundaries, and constructing a linear constraint space for parameter adjustment accordingly, the parameter optimization process can be carried out under the premise of meeting the stability margin requirements. This directly limits the parameter search range, effectively avoiding the introduction of potential instability risks during parameter adjustment, and significantly improving the safety and interpretability of the parameter self-tuning process. By introducing a gradient projection algorithm to constrain the parameter gradient, the parameter update direction is always within the allowed constraint space, thereby ensuring the continuity and controllability of the parameter iterative update process. While ensuring that the stability constraints are not violated, the convergence efficiency and numerical stability of parameter adjustment are improved.

[0052] In one alternative implementation, The damping correction vector is calculated using the original dual interior point algorithm and applied to the intermediate parameter vector to obtain the candidate parameter set, which includes: Extract the linear constraint matrix and constraint boundary range in the parameter adjustment constraint space to construct the feasible region. Perform feasible region projection mapping on the intermediate parameter vector and calculate the shortest distance from the intermediate parameter vector to the feasible region boundary to obtain the relaxation variable vector. The original variable vector is constructed based on the intermediate parameter vector and the relaxation variable vector, and the dual variable vector is constructed based on the transpose of the linear constraint matrix. The original variable vector and the dual variable vector are combined to construct the extended variable vector, and strict interior point initialization detection is performed to obtain the initial extended variable vector. The initial extended variable vector is iteratively updated using the primal-dual interior point algorithm. The complementary gap is calculated based on the current extended variable vector. The first and second partial derivatives of the current extended variable vector are calculated to construct the gradient vector and Hessian matrix. The Newton direction vector is solved based on the gradient vector and the Hessian matrix, and the optimal step size factor is obtained by performing a step size optimization search. The current extended variable vector is updated along the Newton direction vector based on the optimal step size factor until the complementary gap converges, thus obtaining the optimized parameter vector. The damping coefficient adjustment and time constant adjustment are extracted from the optimized parameter vector to construct a damping correction vector. Based on the damping correction vector and the intermediate parameter vector, the updated parameter vector is obtained and a candidate parameter set is constructed.

[0053] The feasible region is constructed by extracting and adjusting the linear constraint matrix and constraint boundary range in the constraint space. The linear constraint matrix has a dimension of 4×4 and includes upper and lower limit constraints on four parameters: speed loop proportional gain, speed loop integral time constant, torque loop proportional gain, and torque loop integral time constant. The constraint boundary ranges are: speed loop proportional gain 3.0 to 6.0, speed loop integral time constant 0.02 to 0.10, torque loop proportional gain 4.0 to 7.0, and torque loop integral time constant 0.02 to 0.08. A feasible region projection mapping is performed on the intermediate parameter vector, and the shortest distance from the intermediate parameter vector to the feasible region boundary is calculated to obtain the relaxation variable vector. For example, in the case of a sharp turn of an electric vehicle, the intermediate parameter vector is: speed loop proportional gain 5.85, speed loop integral time constant 0.018, torque loop proportional gain 6.92, and torque loop integral time constant 0.022. Calculate the distance from this vector to each constraint boundary to obtain the relaxation variable vector: speed loop proportional gain relaxation value 0.15, speed loop integral time constant relaxation value -0.002, torque loop proportional gain relaxation value 0.08, torque loop integral time constant relaxation value 0.002.

[0054] The original variable vector is constructed based on the intermediate parameter vector and the relaxation variable vector. The original variable vector has a dimension of 8, with the first 4 dimensions representing the values ​​of the intermediate parameter vector and the last 4 dimensions representing the values ​​of the relaxation variable vector. The dual variable vector is constructed based on the transpose of the linear constraint matrix. The dual variable vector has a dimension of 4, corresponding to the Lagrange multipliers of the four linear constraints. The original and dual variable vectors are combined to construct the extended variable vector. The extended variable vector has a dimension of 12 and contains complete information about both the original and dual variables. A strict interior-point initialization check is performed to ensure that the initial extended variable vector strictly satisfies all constraints and that all variables are positive. The check method is to calculate the difference between the original variables and the constraint boundaries; if values ​​are close to or exceed the boundaries, appropriate adjustments are made. The adjusted initial extended variable vector is: original part (5.80, 0.025, 6.85, 0.025, 0.20, 0.005, 0.15, 0.005), dual part (0.10, 0.15, 0.10, 0.15).

[0055] The initial extended variable vector is iteratively updated using the primal-dual interior-point algorithm. This algorithm simultaneously optimizes both the original and dual variables, effectively handling constrained optimization problems. During iteration, the complementary gap is calculated based on the current extended variable vector. The complementary gap represents the degree of constraint relaxation and is calculated as the inner product of the relaxed portion of the original variable and its corresponding dual variable. The initial complementary gap value is 0.085. The first-order partial derivatives of the current extended variable vector are calculated to construct the gradient vector, which has a dimension of 12 and represents the sensitivity of the objective function to each variable. The second-order partial derivatives are calculated to construct the Hessian matrix, which has a dimension of 12×12 and represents the curvature information of the objective function. The objective function is defined as a controller performance index, including a weighted sum of overshoot, settling time, and steady-state error.

[0056] The Newton direction vector is solved using the gradient vector and Hessian matrix. The solution method involves solving a system of linear equations, aiming to find the direction that causes the objective function to decrease the fastest. The resulting Newton direction vector also has a dimension of 12, representing the adjustment direction and magnitude of each variable. A step-size optimization search is performed to determine the optimal step-size factor. The search method uses a backtracking algorithm, starting with an initial step size of 1.0 and decreasing by 0.8 at a rate until the optimal step size satisfying the Amhorst condition is found. In the first iteration, the optimal step-size factor is 0.64. Based on the optimal step-size factor, the current extended variable vector is updated along the Newton direction vector, and the complementary gap is recalculated. The iteration process continues until the complementary gap is less than a preset threshold of 0.0001 or the maximum number of iterations (50) is reached. After 14 iterations, the complementary gap decreases to 0.000096, reaching the convergence condition, and the optimized parameter vector is obtained: speed loop proportional gain 5.65, speed loop integral time constant 0.028, torque loop proportional gain 6.80, and torque loop integral time constant 0.026.

[0057] The damping coefficient adjustment and time constant adjustment are extracted from the optimized parameter vector and combined to construct a damping correction vector. The damping coefficient adjustment is calculated by comparing the optimized proportional gain with the proportional gain in the intermediate parameter vector, and the time constant adjustment is calculated by comparing the optimized integral time constant with the integral time constant in the intermediate parameter vector. The damping correction vector is as follows: speed loop proportional gain correction -0.20, speed loop integral time constant correction 0.010, torque loop proportional gain correction -0.12, torque loop integral time constant correction 0.004. Based on the damping correction vector and the intermediate parameter vector, the updated parameter vector is obtained: speed loop proportional gain 5.65, speed loop integral time constant 0.028, torque loop proportional gain 6.80, torque loop integral time constant 0.026.

[0058] A candidate parameter set is constructed, containing multiple possible parameter configurations for the controller to choose from. In addition to the optimal parameter combination obtained based on the primal-dual interior-point algorithm, several other parameter configurations are constructed as candidates. The candidate parameters are obtained by making small adjustments to the optimal parameters, with the adjustment range within ±5%. The resulting candidate parameter set contains five parameter configurations: baseline group (5.65, 0.028, 6.80, 0.026); candidate group one (5.80, 0.027, 6.75, 0.027); candidate group two (5.50, 0.029, 6.85, 0.025); candidate group three (5.70, 0.026, 6.90, 0.024); and candidate group four (5.60, 0.030, 6.70, 0.028).

[0059] In this embodiment, the feasible region of parameters is explicitly constructed based on the linear constraint matrix and the range of the constraint boundary. The intermediate parameter vector is then projected into the feasible region. Simultaneously, relaxation variables are introduced to characterize the distance between the parameters and the boundary of the feasible region. This allows the parameter optimization process to accurately perceive the degree of constraint tension, improving the smoothness and numerical stability of the parameter update process. By simultaneously constructing the original and dual variables and performing strict interior point initialization, the parameter iteration always remains within the feasible region. This fundamentally avoids the problem of repeatedly touching or crossing the constraint boundary during the parameter update process, significantly reducing the risk of constraint violation and improving the reliability of the solution in high-dimensional constraint space. By combining the original-dual interior point algorithm with Newton's direction search, the parameters are updated iteratively with second-order information-driven processing. This allows the parameter adjustment process to quickly approach the optimal solution while ensuring that the constraints are satisfied, resulting in faster convergence speed and higher solution accuracy under complex constraint conditions.

[0060] Figure 2 This is a flowchart of the original-dual interior-point method for parameter optimization in the adaptive tuning method of electric vehicle power controller parameters according to an embodiment of the present invention.

[0061] In one alternative implementation, Based on the motor torque feedback signal and battery current feedback signal, power performance indicators and energy efficiency performance indicators are constructed. A Pareto front search is performed on the candidate parameter set using a multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector, which is then loaded into the controller. The motor output torque is controlled in conjunction with the pre-acquired accelerator pedal command signal, including: Perform time window statistical analysis on the motor torque feedback signal, calculate the torque response time and torque overshoot, and construct the power performance index. Perform integral operation on the battery current feedback signal, calculate the energy consumption ratio in combination with the driving distance, and construct the energy efficiency performance index. Extract each candidate parameter vector from the candidate parameter set and calculate the bi-objective fitness vector corresponding to each candidate parameter vector by combining the power performance index and the energy efficiency performance index. Perform a dominance relationship determination on any two candidate parameter vectors in the candidate parameter set to obtain the initial solution set of the Pareto front. Initialize the particle swarm and assign the parameter vector from the initial Pareto front solution set to each particle. Iteratively update the particle velocity and position based on the multi-objective particle swarm optimization algorithm. Combine the bi-objective fitness vector to calculate the individual optimal position and global optimal position of each particle and update the particle velocity and position again. Based on the updated particle velocity and position, perform dominance relationship determination and update the initial Pareto front solution set to obtain the Pareto front solution set and extract the optimal solution to obtain the optimal parameter vector. The optimal parameter vector is loaded into the controller, the pre-acquired accelerator pedal command signal is obtained and the target torque requirement is calculated, and the motor control command is generated based on the optimal parameter vector and the target torque requirement to control the motor output torque.

[0062] Statistical analysis of the motor torque feedback signal was performed using a sliding time window method with a window length of 1.0 second and a sliding step size of 0.1 seconds. The rise time and peak overshoot of the torque signal were recorded within each window. Torque response time was defined as the time required for the torque to rise from 10% to 90% of the target value, and interpolation fitting was performed based on the time-domain characteristics of the torque feedback signal. Torque overshoot was defined as the ratio of the difference between the peak torque and the steady-state torque to the steady-state torque, expressed as a percentage. In a test of a four-wheel-drive electric vehicle, the motor torque feedback signal sampling frequency was 500 Hz, and the statistical analysis yielded a torque response time of 0.26 seconds and a torque overshoot of 6.5%. When constructing the power performance index, the torque response time and torque overshoot were weighted and combined with weights of 0.6 and 0.4, respectively, resulting in a comprehensive power performance index value of 0.355. Integral calculations were performed on the battery current feedback signal with a sampling period of 0.01 seconds and an integration time of 300 seconds to obtain the total charge. Based on the vehicle's travel distance of 2.5 kilometers, the energy consumption ratio is calculated to be 0.072 kWh / km. An energy efficiency performance index is constructed, and the value of the energy efficiency performance index is equal to the normalized value of the energy consumption ratio, which is 0.625.

[0063] Candidate parameter vectors are extracted from the candidate parameter set, which contains five parameter configurations: (5.65, 0.028, 6.80, 0.026), (5.80, 0.027, 6.75, 0.027), (5.50, 0.029, 6.85, 0.025), (5.70, 0.026, 6.90, 0.024), and (5.60, 0.030, 6.70, 0.028). The bi-objective fitness vector corresponding to each candidate parameter vector is calculated based on the aforementioned power performance and energy efficiency performance indicators. Each set of parameters is applied to the controller, and a standard test condition cycle is executed, recording the power performance and energy efficiency performance indicators. The bi-objective fitness vectors corresponding to the five sets of parameters are (0.355, 0.625), (0.382, 0.598), (0.346, 0.652), (0.368, 0.612), and (0.334, 0.665). A dominance relationship is determined for any two candidate parameter vectors in the candidate parameter set. The criterion is: if the bi-objective fitness vector of parameter vector A is not inferior to parameter vector B on both objectives, and is strictly superior to B on at least one objective, then A dominates B. After determination, dominated parameter vectors are removed, and the initial solution set of the Pareto front is obtained: (5.65, 0.028, 6.80, 0.026), (5.50, 0.029, 6.85, 0.025), and (5.60, 0.030, 6.70, 0.028).

[0064] The particle swarm is initialized with 20 particles and a dimension of 4, corresponding to the four parameters of the controller. The parameter vectors from the initial Pareto front solution set are assigned to the first three particles, and the remaining 17 particles are randomly initialized within the allowed range of parameter variation. A multi-objective particle swarm optimization algorithm is used to iteratively update particle velocity and position. The particle velocity update equation considers inertia weight, individual cognitive factor, and social cognitive factor, with inertia weight set to 0.7, individual cognitive factor to 1.5, and social cognitive factor to 2.0. The maximum velocity is limited to 20% of the parameter range, specifically: maximum velocity for the proportional gain of the rotational speed loop is 0.60, maximum velocity for the integral time constant of the rotational speed loop is 0.016, maximum velocity for the proportional gain of the torque loop is 0.60, and maximum velocity for the integral time constant of the torque loop is 0.012. Particle position updates are achieved by adding velocity to the current position, while simultaneously checking if the parameters exceed the boundaries. If they do, the position is truncated to the boundary value and the velocity direction is reversed.

[0065] The individual optimal position of each particle is calculated by combining the dual-objective fitness vector, and the non-dominated solution set of each particle's history is stored. The global optimal position is selected from the non-dominated solution set of all particles, and the crowding distance sorting method is used to select solutions located in sparse regions to maintain the diversity of the solution set. The crowding distance calculation method measures the sum of the Manhattan distances between a solution and its neighboring solutions in the target space. Based on the updated individual optimal position and global optimal position, the particle velocity and particle position are updated again. A random factor is introduced into the update formula, which follows a uniform distribution from 0 to 1 to increase the randomness and diversity of the search. The three terms of the velocity update formula correspond to inertia, the tendency to move towards the individual optimal solution, and the tendency to move towards the global optimal solution, respectively. During the iteration process, the inertia weight decreases linearly from 0.7 to 0.4 to balance the global search and local search capabilities.

[0066] Domination relationships are determined based on the updated particle velocities and positions, and the initial Pareto front solution set is updated. The determination process compares the current positions of all particles with solutions in the initial Pareto front solution set, removing old solutions dominated by new solutions and adding new solutions not dominated by any old solutions. After 100 iterations, a Pareto front solution set containing 15 non-dominated solutions is obtained. The optimal solution is extracted from these 15 non-dominated solutions using a normalized weighted sum method, with a weight of 0.6 for dynamic performance and 0.4 for energy efficiency. The calculated optimal parameter vector is (5.58, 0.031, 6.75, 0.027), and the corresponding bi-objective fitness vector is (0.328, 0.638).

[0067] The optimal parameter vector is loaded into the controller, which employs a dual-loop PI control structure: an outer loop for speed control and an inner loop for torque control. The first two components of the parameter vector are used in the speed loop PI controller, and the last two components are used in the torque loop PI controller. Pre-acquired accelerator pedal command signals are acquired at a sampling frequency of 100 Hz. The accelerator pedal travel range is 0% to 100%, which is converted to the target speed via a calibration curve. The target torque requirement is then calculated using the speed-torque mapping relationship. For example, in a certain test scenario, the accelerator pedal travel is 65%, corresponding to a target speed of 3600 rpm, and the calculated target torque requirement is 120 Nm. Motor control commands are generated based on the optimal parameter vector and the target torque requirement. These commands include torque commands output from the speed loop and current commands output from the torque loop. The torque command output from the speed loop is calculated using proportional and integral terms, as is the current command output from the torque loop. The motor control commands are transmitted to the motor driver in the form of pulse-width modulated signals to control the motor's output torque.

[0068] In this embodiment, by performing time window statistical analysis on the motor torque feedback signal and extracting the torque response time and overshoot, the power performance evaluation can directly reflect the influence of control parameters on the actual torque response characteristics. By integrating the battery current feedback signal and combining it with the driving distance to construct an energy consumption index, the energy efficiency evaluation is closely linked to the actual operating energy consumption, improving the engineering effectiveness of the parameter optimization results. By constructing a dual-objective fitness and performing dominance relationship judgment to screen the Pareto front solution set, the parameter selection no longer depends on fixed weights or human preferences, but forms a set of optimal trade-off solutions between power performance and energy efficiency performance. This avoids the subjectivity of weight setting and improves the objectivity and comprehensiveness of the parameter optimization results. By introducing a multi-objective particle swarm optimization algorithm to iteratively search the Pareto solution set, the parameter optimization process can maintain diversity in the solution space and continuously approach the global optimal Pareto front. It has stronger global search capabilities and convergence stability in complex nonlinear parameter spaces, improving the quality of parameter solutions.

[0069] A second aspect of the present invention provides an adaptive tuning system for the parameters of an electric vehicle power controller, comprising: The feature extraction unit is used to acquire the motor torque feedback signal of the electric vehicle and perform multi-scale wavelet decomposition to obtain fast-changing components, construct a transient response feature matrix based on the fast-changing components and extract the corresponding dominant mode features, and determine the operating condition identifier based on the dominant mode features. The margin analysis unit is used to determine the transient frequency domain feature spectrum corresponding to the transient response feature matrix through Fourier transform and extract the phase margin offset, obtain the battery current feedback signal of the electric vehicle and extract the slow-varying component, construct the steady-state trend feature matrix based on the slow-varying component and perform Fourier transform to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset. The parameter optimization unit is used to extract an initial parameter vector from a preset reference parameter library according to the operating condition identifier, construct a parameter adjustment constraint space based on the phase margin offset and the gain margin offset, calculate a gain correction vector by combining the gradient projection algorithm and apply it to the initial parameter vector to obtain an intermediate parameter vector, and calculate a damping correction vector by using the original dual interior point algorithm and apply it to the intermediate parameter vector to obtain a candidate parameter set. The strategy execution unit is used to construct power performance indicators and energy efficiency performance indicators based on motor torque feedback signals and battery current feedback signals. It combines a multi-objective particle swarm optimization algorithm to perform Pareto front search on the candidate parameter set, obtain the optimal parameter vector, and load it into the controller. It then controls the motor output torque in conjunction with the pre-acquired accelerator pedal command signal.

[0070] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0071] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0072] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive tuning method for the parameters of an electric vehicle power controller, characterized in that, include: The motor torque feedback signal of the electric vehicle is acquired and multi-scale wavelet decomposition is performed to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features. The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slow-varying component is extracted. Based on the slow-varying component, a steady-state trend feature matrix is ​​constructed and Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and the amplitude margin offset is extracted. According to the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. The gain correction vector is calculated by combining the gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector. The damping correction vector is calculated by the original dual interior point algorithm and applied to the intermediate parameter vector to obtain a candidate parameter set. Based on the motor torque feedback signal and the battery current feedback signal, power performance index and energy efficiency performance index are constructed. The Pareto front search of the candidate parameter set is performed by combining the multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector and load it into the controller. The motor output torque is controlled by combining the pre-acquired accelerator pedal command signal.

2. The method according to claim 1, characterized in that, The motor torque feedback signal of the electric vehicle is acquired and subjected to multi-scale wavelet decomposition to obtain fast-changing components. A transient response feature matrix is ​​constructed based on the fast-changing components, and the corresponding dominant mode features are extracted. The operating condition identifier is determined based on the dominant mode features, including: The torque measurement value of the motor output shaft is collected by the motor torque sensor and converted from analog to digital to obtain the motor torque feedback signal. The torque sequence is then obtained by performing a sliding window interception. Based on the torque sequence, a wavelet mother function is selected and a multi-level decomposition scale is set to perform discrete wavelet transform to obtain detail coefficients at multiple scales. The energy values ​​of the detail coefficients are calculated and a scale energy distribution vector is constructed. Based on the scale energy distribution vector, key scales are identified and the detail coefficients corresponding to the key scales are extracted to perform wavelet reconstruction to obtain fast change components. The instantaneous amplitude envelope and instantaneous phase corresponding to the fast-changing component are calculated to obtain the time-domain transient features. The Hilbert transform is performed on the fast-changing component to extract the instantaneous frequency sequence and the frequency change rate is calculated to obtain the frequency-domain transient features. The time-domain transient features and the frequency-domain transient features are time-frequency synchronized to obtain the aligned feature sequence and tensor reconstruction is performed to obtain the three-dimensional time-frequency feature tensor. The three-dimensional time-frequency feature tensor is subjected to Tucker decomposition to obtain the core tensor and mode factor matrix and tensor product reconstruction is performed to obtain the transient response feature matrix. Singular value decomposition is performed on the transient response feature matrix to construct dominant mode features. Based on the dominant mode features, a working condition feature descriptor is constructed and matched with a preset working condition pattern library using distance metric matching to determine the working condition identifier.

3. The method according to claim 1, characterized in that, The transient frequency domain feature spectrum corresponding to the transient response feature matrix is ​​determined by Fourier transform, and the phase margin offset is extracted. The battery current feedback signal of the electric vehicle is obtained and the slowly varying component is extracted. Based on the slowly varying component, a steady-state trend feature matrix is ​​constructed and a Fourier transform is performed to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset, including: Windowing and fast Fourier transform are performed on the transient response feature matrix to obtain a frequency domain complex matrix. The amplitude spectrum and phase spectrum are calculated based on the frequency domain complex matrix. Phase unwrapping is performed on the phase spectrum to obtain a continuous phase curve and the phase crossing frequency is identified. The phase margin offset is calculated based on the amplitude spectrum at the phase crossing frequency. The current measurement value is collected and converted from analog to digital to obtain the battery current feedback signal. The battery current feedback signal is subjected to sliding window truncation to obtain the current sequence and low-pass filtering to obtain the low-frequency component. The low-frequency component is subjected to moving average smoothing to obtain the trend baseline. Based on the trend baseline, the segment is divided and the local slope and local mean in each segment are calculated to construct the segmented trend feature vector as the slow variable component. Based on the slowly varying components, a state vector is constructed. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of state posterior mean values. This sequence is then expanded over time. Finally, the enhanced slowly varying feature matrix is ​​obtained by combining the segmented trend feature vector with the solution. The enhanced slowly varying feature matrix is ​​aligned with a time window and zero-padding is applied to obtain a steady-state trend feature matrix. The steady-state trend feature matrix is ​​then subjected to a fast Fourier transform to obtain a steady-state frequency domain feature spectrum and a steady-state phase spectrum is calculated. The phase values ​​corresponding to the steady-state phase spectrum are extracted and the magnitude margin offset is calculated.

4. The method according to claim 3, characterized in that, A state vector is constructed based on the slowly varying components. The posterior probability distribution of the state vector is iteratively optimized using a variational Bayesian inference algorithm to obtain a sequence of posterior mean values. This sequence is then expanded over time. The enhanced slowly varying feature matrix is ​​obtained by combining this with the segmented trend feature vector. The slowly varying components are sorted according to time order, and the local slope and local mean corresponding to each time point are extracted. The state vector is constructed based on the combination of the local slope and the local mean. The prior mean and prior covariance are set for the state vector to obtain the prior probability distribution. Based on the prior probability distribution, define the state transition matrix and observation matrix and establish the state space equation. Based on the prior probability distribution and the pre-acquired observation data, derive the posterior probability distribution of the state vector. Perform iterative optimization of maximizing the lower bound of evidence on the posterior probability distribution using the variational Bayesian inference algorithm. Update the state prediction based on the state transition matrix and update the observation prediction in combination with the observation matrix. Repeat the iteration until the posterior mean and posterior covariance of the state converge. The time dimension is expanded to obtain expanded state features by performing time dimension expansion on the posterior mean sequence of the state. The difference between adjacent time points is calculated based on the expanded state features to obtain the time derivative sequence. The time derivative sequence is normalized to obtain the dynamic rate of change vector. The dynamic rate of change vector and the slow-varying component are matched with timestamps to obtain aligned feature pairs. Based on the aligned features, a fused feature vector is calculated, and the fused feature vector is matrix reorganized according to the time dimension and the feature dimension to obtain the enhanced slow-varying feature matrix.

5. The method according to claim 1, characterized in that, Based on the operating condition identifier, an initial parameter vector is extracted from a preset reference parameter library. A parameter adjustment constraint space is constructed based on the phase margin offset and the magnitude margin offset. A gain correction vector is calculated using a gradient projection algorithm and applied to the initial parameter vector to obtain an intermediate parameter vector, including: Based on the operating condition identifier, an index query is performed in the preset benchmark parameter library to extract the benchmark gain parameter and benchmark time constant parameter corresponding to the operating condition identifier from the benchmark parameter library and combine them to construct the initial parameter vector; The phase margin offset is converted into a phase constraint boundary, the magnitude margin offset is converted into a magnitude constraint boundary, the constraint boundary range is determined based on the phase constraint boundary and the magnitude constraint boundary, a linear constraint matrix is ​​constructed based on the constraint boundary range, and the parameter adjustment constraint space is defined based on the linear constraint matrix. The gradient vector is obtained by calculating the partial derivatives of each component of the initial parameter vector. The gradient vector is then projected onto the parameter adjustment constraint space to obtain the constraint gradient vector. Based on the constraint gradient vector and the linear constraint matrix, the gain correction vector is obtained by performing iterative update calculation through the gradient projection algorithm. The gain correction vector is added element-wise to the initial parameter vector to obtain the corrected parameter vector. The corrected parameter vector is then subjected to constraint boundary detection and truncation to obtain the intermediate parameter vector.

6. The method according to claim 1, characterized in that, The damping correction vector is calculated using the original dual interior point algorithm and applied to the intermediate parameter vector to obtain the candidate parameter set, which includes: Extract the linear constraint matrix and constraint boundary range in the parameter adjustment constraint space to construct the feasible region. Perform feasible region projection mapping on the intermediate parameter vector and calculate the shortest distance from the intermediate parameter vector to the feasible region boundary to obtain the relaxation variable vector. The original variable vector is constructed based on the intermediate parameter vector and the relaxation variable vector, and the dual variable vector is constructed based on the transpose of the linear constraint matrix. The original variable vector and the dual variable vector are combined to construct the extended variable vector, and strict interior point initialization detection is performed to obtain the initial extended variable vector. The initial extended variable vector is iteratively updated using the primal-dual interior point algorithm. The complementary gap is calculated based on the current extended variable vector. The first and second partial derivatives of the current extended variable vector are calculated to construct the gradient vector and Hessian matrix. The Newton direction vector is solved based on the gradient vector and the Hessian matrix, and the optimal step size factor is obtained by performing a step size optimization search. The current extended variable vector is updated along the Newton direction vector based on the optimal step size factor until the complementary gap converges, thus obtaining the optimized parameter vector. The damping coefficient adjustment and time constant adjustment are extracted from the optimized parameter vector to construct a damping correction vector. Based on the damping correction vector and the intermediate parameter vector, the updated parameter vector is obtained and a candidate parameter set is constructed.

7. The method according to claim 1, characterized in that, Based on the motor torque feedback signal and battery current feedback signal, power performance indicators and energy efficiency performance indicators are constructed. A Pareto front search is performed on the candidate parameter set using a multi-objective particle swarm optimization algorithm to obtain the optimal parameter vector, which is then loaded into the controller. The motor output torque is controlled in conjunction with the pre-acquired accelerator pedal command signal, including: Perform time window statistical analysis on the motor torque feedback signal, calculate the torque response time and torque overshoot, and construct the power performance index. Perform integral operation on the battery current feedback signal, calculate the energy consumption ratio in combination with the driving distance, and construct the energy efficiency performance index. Extract each candidate parameter vector from the candidate parameter set and calculate the bi-objective fitness vector corresponding to each candidate parameter vector by combining the power performance index and the energy efficiency performance index. Perform a dominance relationship determination on any two candidate parameter vectors in the candidate parameter set to obtain the initial solution set of the Pareto front. Initialize the particle swarm and assign the parameter vector from the initial Pareto front solution set to each particle. Iteratively update the particle velocity and position based on the multi-objective particle swarm optimization algorithm. Combine the bi-objective fitness vector to calculate the individual optimal position and global optimal position of each particle and update the particle velocity and position again. Based on the updated particle velocity and position, perform dominance relationship determination and update the initial Pareto front solution set to obtain the Pareto front solution set and extract the optimal solution to obtain the optimal parameter vector. The optimal parameter vector is loaded into the controller, the pre-acquired accelerator pedal command signal is obtained and the target torque requirement is calculated, and the motor control command is generated based on the optimal parameter vector and the target torque requirement to control the motor output torque.

8. An adaptive tuning system for electric vehicle power controller parameters, used to implement the method of any one of claims 1-7, characterized in that, include: The feature extraction unit is used to acquire the motor torque feedback signal of the electric vehicle and perform multi-scale wavelet decomposition to obtain fast-changing components, construct a transient response feature matrix based on the fast-changing components and extract the corresponding dominant mode features, and determine the operating condition identifier based on the dominant mode features. The margin analysis unit is used to determine the transient frequency domain feature spectrum corresponding to the transient response feature matrix through Fourier transform and extract the phase margin offset, obtain the battery current feedback signal of the electric vehicle and extract the slow-varying component, construct the steady-state trend feature matrix based on the slow-varying component and perform Fourier transform to obtain the steady-state frequency domain feature spectrum and extract the amplitude margin offset. The parameter optimization unit is used to extract an initial parameter vector from a preset reference parameter library according to the operating condition identifier, construct a parameter adjustment constraint space based on the phase margin offset and the gain margin offset, calculate a gain correction vector by combining the gradient projection algorithm and apply it to the initial parameter vector to obtain an intermediate parameter vector, and calculate a damping correction vector by using the original dual interior point algorithm and apply it to the intermediate parameter vector to obtain a candidate parameter set. The strategy execution unit is used to construct power performance indicators and energy efficiency performance indicators based on motor torque feedback signals and battery current feedback signals. It combines a multi-objective particle swarm optimization algorithm to perform Pareto front search on the candidate parameter set, obtain the optimal parameter vector, and load it into the controller. It then controls the motor output torque in conjunction with the pre-acquired accelerator pedal command signal.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images