Construction method and system of motor-driven dynamic control model

Abstract

The invention relates to the technical field of motor driving control, in particular to a construction method and system of a dynamic control model driven by a motor. The method comprises the following steps of: acquiring operation data of the non-electrolytic capacitor motor driving system under different control parameters to form a full-power section operation data set; performing multi-scale analysis and high-dimensional compression feature decoupling on the full-power section operation data set to obtain a multi-scale dynamic feature vector; constructing a nonlinear regression model capable of being used for real-time control quantity prediction according to the multi-scale dynamic feature vector; deploying a nonlinear regression model in a controller of the non-electrolytic capacitor motor driving system; and executing motor driving control by utilizing the predictive control quantity output by the nonlinear regression model so as to form an electrolytic capacitor-free motor driving strategy. Through high-dimensional dynamic coupling compression representation learning, an electrolytic capacitor-free motor driving strategy of intelligent voltage stabilization regulation and control is constructed to support efficient and steady operation of the humanoid robot.

Description

Technical Field

[0001] This invention relates to the field of motor drive control technology, and in particular to a method and system for constructing a dynamic control model for a motor drive. Background Technology

[0002] The key characteristics of dynamic control technology in motor drive systems are mainly reflected in the stability and control accuracy of the bus voltage under dynamic loads, random operating conditions, and high-dimensional coupling characteristics. Specifically, electrolytic capacitor-less motor drive systems are subjected to electrical shocks such as instantaneous load changes, rapid command jumps, and high-frequency switching of power switching devices during actual operation. These factors work together to cause drastic fluctuations in the energy flow within the system. Without effective control, this will directly lead to severe oscillations or even instability of the DC bus voltage, ultimately affecting its drive performance and operational safety.

[0003] In existing technologies, machine learning requires a large amount of interactive data to build efficient and accurate models. However, in motor drive systems, frequent exploration and experimentation can damage equipment. Acquiring large-scale datasets is inefficient and costly. Furthermore, the high-dimensional feature aliasing (dimension > 30) of the system restricts the efficiency of model construction and training. It is difficult to efficiently decouple and dynamically represent these complex features, and it is impossible to quickly extract the core dynamic laws that are strongly correlated with the stability of the bus voltage from massive amounts of data. This increases the difficulty of model construction and fails to provide a strong guarantee for the stable operation of the drive system under all operating conditions. Summary of the Invention

[0004] Based on this, the present invention provides a method and system for constructing a dynamic control model for motor drive, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a method for constructing a dynamic control model for motor drive includes the following steps: Step S1: Collect the operating data of the electrolytic capacitor-free motor drive system under different control parameters, obtain the state observations including bus voltage, stator current, speed and torque, as well as the corresponding control quantities, and form a full power range operating dataset; Step S2: Construct bus voltage stability constraints, and then perform multi-scale analysis and high-dimensional compressed feature decoupling on the full power segment operation dataset to obtain multi-scale dynamic feature vectors; Step S3: Construct a nonlinear regression training dataset based on multi-scale dynamic feature vectors, and construct a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraint; Step S4: Deploy a nonlinear regression model in the controller of the capacitor-free motor drive system; acquire state observations in real time as input to the nonlinear regression model, and use the predictive control quantity output by the nonlinear regression model to execute motor drive control, thereby forming a capacitor-free motor drive strategy.

[0006] The present invention also provides a system for constructing a dynamic control model for a motor drive, which executes the method for constructing a dynamic control model for a motor drive as described above. The system for constructing a dynamic control model for a motor drive includes: The operating data acquisition module is used to collect the operating data of the electrolytic capacitor-free motor drive system under different control parameters, and obtain state observations including bus voltage, stator current, speed and torque, as well as the corresponding control quantities, to form a full power range operating dataset. The dynamic feature decoupling module is used to construct bus voltage stability constraints, and then perform multi-scale analysis and high-dimensional compressed feature decoupling on the full power segment operation dataset to obtain multi-scale dynamic feature vectors. The prediction model building module is used to construct a nonlinear regression training dataset based on multi-scale dynamic feature vectors, and to build a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraint. The drive strategy deployment module is used to deploy a nonlinear regression model in the controller of the electrolytic capacitor-free motor drive system. By acquiring state observations in real time as inputs to the nonlinear regression model, and using the predictive control quantities output by the nonlinear regression model to execute motor drive control, a drive strategy for the electrolytic capacitor-free motor is formed.

[0007] The beneficial effects of this invention are as follows: On the one hand, by collecting full-power range operating datasets and constructing a physical-data hybrid architecture for high-dimensional compressed feature decoupling, dynamic feature vectors are obtained through multi-scale analysis. This process, based on the system's inherent physical laws and a data-driven deep learning model, can efficiently and safely solve the data acquisition dilemma and high-dimensional feature aliasing problem required for intelligent model construction. Furthermore, through the physical-data hybrid architecture, hard constraints such as stator voltage equations and electromagnetic torque equations can be embedded into the data-driven decoupling process, ensuring the physical interpretability of feature extraction and avoiding the blindness of traditional pure data-driven methods. It can effectively separate highly coupled feature components such as voltage, current, load, and power. In addition, by performing multi-scale analysis on the decoupled features, the short-term fluctuations and long-term trends of the system at different time scales can be accurately captured, significantly improving the ability to characterize the dynamic characteristics of the motor drive system.

[0008] On the other hand, a nonlinear regression model is constructed using a deep neural network, and the bus voltage stability constraint based on the Lyapunov function is incorporated into the loss function for training. The final model is then deployed to form a capacitor-free motor drive strategy. The bus voltage stability constraint is used as a hard indicator for model training, ensuring that the trained model not only pursues prediction accuracy but also inherently satisfies the boundary conditions for stable system operation. This effectively avoids the risk of traditional machine learning models outputting unstable control commands under extreme conditions. Simultaneously, by deploying a complete feature extraction and model prediction process, this invention can process the collected state observations in real time and quickly output the optimal predictive control quantity, accurately reflecting and responding to the dynamic response requirements of the system under random operating conditions. Furthermore, without the support of physical electrolytic capacitors, this intelligent voltage regulation strategy enables precise and rapid adjustment of the bus voltage, obtaining a robust bus voltage. This effectively ensures the efficient and stable operation of the drive system under all operating conditions and significantly improves the system's power density, reliability, and service life. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the steps of constructing the dynamic control model for motor drive according to the present invention. Figure 2 This is a module architecture diagram of the system for constructing the dynamic control model of the motor drive of the present invention; Figure 3 This is a schematic diagram of the electrolytic capacitor-free motor drive strategy based on intelligent voltage regulation in this invention; Figure 4 This is a flowchart illustrating the construction of the high-dimensional dynamic coupling compressed representation learning model based on strong bus voltage regulation in this invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0012] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] To achieve the above objectives, please refer to Figures 1 to 4 This invention provides a method for constructing a dynamic control model for a motor drive, comprising the following steps: Step S1: Collect the operating data of the electrolytic capacitor-free motor drive system under different control parameters, obtain the state observations including bus voltage, stator current, speed and torque, as well as the corresponding control quantities, and form a full power range operating dataset; In embodiments of the present invention, including but not limited to, data acquisition is performed on a capacitor-free motor drive system on a test platform for simulating the high dynamic joints of a humanoid robot. Based on the rated operating conditions of the controlled motor (e.g., rated power 2kW, rated speed 3000rpm), control parameters covering the full power range from no-load to full-load and from low speed to high speed are set in the simulation controller. These parameters are constructed as combinations including initial speed and speed changes applied at intermediate moments, forming a total of 3321 independent control parameter configurations.

[0014] In one implementation of this invention, assuming the current execution is control parameter configuration number 101, which defines an initial speed of 1500 rpm and applies a step speed command change of +800 rpm at 0.7 seconds. During a 1.4-second sampling period, the controller synchronously acquires the DC bus voltage (e.g., instantaneous value of 48.5V), bus current, and three-phase stator current using high-precision voltage and current sensors, and obtains the actual speed and estimated actual torque using a high-resolution encoder mounted coaxially with the motor. Simultaneously, the controller's internal vector control algorithm generates corresponding direct-axis voltage control quantities and quadrature-axis voltage control quantities. These 1.4 million timestamp-aligned data samples, containing state observations and corresponding control quantities, collectively constitute the dataset under configuration number 101 and are ultimately merged into the full-power range operation dataset.

[0015] Step S2: Construct bus voltage stability constraints, and then perform multi-scale analysis and high-dimensional compressed feature decoupling on the full power segment operation dataset to obtain multi-scale dynamic feature vectors; In this embodiment of the invention, a Lyapunov function is constructed to characterize the stability of the bus voltage. This function will be used as a stability penalty term in subsequent model training to ensure that the control strategy output by the model is essentially stable.

[0016] In one implementation of this invention, high-dimensional compressed feature decoupling is achieved through a physical-data hybrid architecture. The data-driven layer of this architecture employs a deep autoencoder (DAE) model. Raw state observations (such as time-series data like voltage, three-phase current, and rotational speed) exceeding 30 dimensions, extracted from the full-power operating dataset, are fed into the DAE's encoder as a high-dimensional input vector. The encoder, through a nonlinear transformation of a multi-layer neural network, compresses this high-dimensional input vector into a low-dimensional data-driven feature vector with a dimension of no more than 10, thereby achieving feature decoupling of each physical component in the low-dimensional space and forming a decoupled feature set.

[0017] Step S3: Construct a nonlinear regression training dataset based on multi-scale dynamic feature vectors, and construct a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraint; In one implementation of this invention, a nonlinear regression deep neural network model is constructed. The number of nodes in the input layer matches the dimension of the multi-scale dynamic feature vector, and the output layer has two nodes (corresponding to the direct-axis voltage control quantity and the quadrature-axis voltage control quantity), containing multiple fully connected hidden layers. The training process employs the Adam optimizer and defines a composite loss function. During training, the backpropagation algorithm is used to simultaneously minimize the prediction error and stability penalty until the model's prediction error on the validation set is lower than a preset threshold (e.g., 0.01V). At this point, the network weight parameters are saved, ultimately generating the nonlinear regression model.

[0018] Step S4: Deploy a nonlinear regression model in the controller of the capacitor-free motor drive system; acquire state observations in real time as input to the nonlinear regression model, and use the predictive control quantity output by the nonlinear regression model to execute motor drive control, thereby forming a capacitor-free motor drive strategy.

[0019] In this embodiment of the invention, the network weight parameters of the stored nonlinear regression model are solidified into the non-volatile memory of the digital signal processor (DSP) or FPGA of the electrolytic capacitor-free motor drive system. After the system is powered on, the model is loaded into memory and enters a standby state. When the motor system starts running in real time, the controller enters a high-speed closed-loop control cycle. In each control cycle (e.g., every 10 microseconds): 1. Real-time observation: The controller acquires state observations such as DC bus voltage, three-phase stator current, and actual speed at the current moment through its ADC and encoder interfaces. 2. Real-time feature extraction: The dedicated computing unit in the controller executes the same processing flow as step S2, and constructs the real-time multi-scale dynamic feature vector of the current moment through high-dimensional compression and multi-scale analysis of the acquired state observations in real time. 3. Real-time prediction: The real-time multi-scale dynamic feature vector is used as input and fed into the deployed nonlinear regression model. The model outputs the predicted direct-axis voltage control quantity and quadrature-axis voltage control quantity in real time through one forward propagation calculation (usually taking only a few microseconds). 4. Real-time drive: The controller directly sends the predicted voltage control quantity to the space vector pulse width modulation (SVPWM) module to generate PWM signals to drive the power switching devices (such as IGBTs) in the inverter, thereby completing the drive control of the motor.

[0020] Preferably, step S1 includes the following steps: Step S11: Initialize the electrolytic capacitor-free motor drive system according to its rated power range, and set the initial speed parameter set and speed change parameter set covering the entire power range; Step S12: Based on the initial speed parameter set and the speed change parameter set, combine different initial speeds with the corresponding speed changes to form multiple sets of control parameter configurations; Step S13: Under each set of control parameter configurations, continuously sample the operation process of the electrolytic capacitor-free motor drive system at a preset sampling frequency to obtain the DC bus voltage, bus current, three-phase stator current, actual speed and actual torque as state observations. Step S14: Synchronously record the direct-axis voltage control quantity and quadrature-axis voltage control quantity at the corresponding time. Step S15: Time-align the state observations with the direct-axis voltage control quantities and quadrature-axis voltage control quantities, and bind them into a full-power segment operation dataset.

[0021] In this embodiment of the invention, the target electrolytic capacitor-free motor drive system is initialized. It is assumed that the rated power of the motor driven by the electrolytic capacitor-free motor drive system is 2 kW, and the rated speed is 3000 rpm. The initial speed parameter set covers the complete range from standstill to the rated speed, specifically set from 0 rpm to 3000 rpm, in 100 rpm increments, generating a total of 31 initial speed levels.

[0022] In one implementation of this invention, the speed change parameter group is used to simulate various acceleration, deceleration and load disturbance conditions encountered by the motor during operation. Its range is set from -1000 rpm to +1000 rpm, with a step of 100 rpm, generating a total of 21 speed change levels.

[0023] It should be noted that this parameter set ensures that the subsequent generated control parameter configuration can fully cover all typical operating conditions encountered by the motor in actual applications, such as low speed, medium speed, high speed, and different degrees of dynamic acceleration and deceleration.

[0024] In one implementation of this invention, the initial speed parameter set and the speed change parameter set are combined using a Cartesian product. For example, 1500 revolutions per minute from the initial speed parameter set is combined with +500 revolutions per minute from the speed change parameter set to form a specific control parameter configuration. This configuration means that the motor first runs stably at a speed of 1500 revolutions per minute, and then at a certain moment, the speed command will step increase by 500 revolutions per minute.

[0025] In this embodiment of the invention, data acquisition is performed in a high-fidelity electrolytic capacitor-free motor drive system model built on a computer simulation platform. It is assumed that, triggered by the configuration of control parameter number 101, the controller performs high-speed, synchronous, and continuous sampling of key physical quantities of the drive system at a preset sampling frequency of 1 MHz. The DC bus voltage and current are acquired through a high-precision shunt and voltage sensor deployed on the DC bus; the three-phase stator current is acquired through a Hall effect sensor; the actual rotational speed is obtained through a photoelectric encoder coaxially connected to the motor shaft; and the actual torque is estimated using a torque observer algorithm within the controller.

[0026] It should be noted that in this invention, the actual torque is not directly measured by physical sensors, but is calculated in real time using the electromagnetic torque equation based on the collected electrical quantities and motor parameters. Its mathematical expression is: ; in, The actual torque obtained from the calculation; This represents the number of pole pairs of the motor. and These are direct-axis and quadrature-axis inductors, respectively. and These are the direct-axis and quadrature-axis currents obtained through coordinate transformation; This refers to the flux linkage of a permanent magnet. For example, the actual torque calculated based on the current value at that moment is 5.2 Nm, and this calculated value is also recorded as part of the state observation.

[0027] In one implementation of this invention, the vector control algorithm module inside the controller calculates the voltage vector to be applied to the motor in real time based on the current speed command and the feedback of actual speed, current, and other states. The component of this voltage vector in the two-phase rotating coordinate system is the direct-axis voltage control quantity. and quadrature axis voltage control quantity It was recorded synchronously.

[0028] In one implementation of this invention, the state observation at the timestamp (a multi-dimensional vector containing DC bus voltage, current, three-phase stator current, speed, and torque) and the control quantity (a multi-dimensional vector containing...) are... and A two-dimensional vector is bound to a data sample unit. The time series data generated under all 3321 sets of control parameter configurations are collected, organized, and stored to form the final full-power segment operation dataset.

[0029] Preferably, in step S2, bus voltage stability constraints are constructed based on Lyapunov functions, and then multi-scale analysis and high-dimensional compressed feature decoupling are performed on the full-power segment operating dataset to obtain multi-scale dynamic feature vectors, including: Mathematical constraint equations for quantifying bus voltage stability constraints are constructed based on Lyapunov functions. Construct a physical-data hybrid architecture; the physical-data hybrid architecture includes a physical model layer and a data-driven layer; In the physical-data hybrid architecture, the bus voltage stability loss term constructed based on mathematical constraint equations is used as a penalty term to identify the coupling connection weights between voltage components, current components, load components, and power components. When the connection weight is lower than a preset redundancy threshold, the connection is eliminated. Dynamic decoupling feature extraction is performed on each component after removing redundant connections to obtain a decoupling feature set.

[0030] Specifically, the state variable is defined as the deviation between the DC bus voltage and its target steady-state value. ,in The target bus voltage is, for example, 310 volts. In one implementation of this invention, a quadratic Lyapunov function is used to construct the mathematical constraint equations, the specific mathematical expression of which is: ; in, It is the Lyapunov function value, which represents the energy or distance that drives the system state to deviate from the steady point; It is a state vector that includes the bus voltage deviation; It is the transpose symbol; It is a pre-selected positive definite symmetric matrix.

[0031] It should be noted that, according to Lyapunov's second stability theory, if it can be guaranteed... It is positive definite, and its time derivative is... If it is negatively definite, then the driving system is asymptotically stable.

[0032] Specifically, the physical model layer is built upon known electrical machinery theory, including stator voltage equations, electromagnetic torque equations, and power balance equations, ensuring that data processing does not violate fundamental physical laws. The data-driven layer employs a deep autoencoder model structure, consisting of an encoder and a decoder. The encoder's role is to nonlinearly compress the high-dimensional input data (dimensions exceeding 30) extracted from the full-power range operating dataset, extracting its core latent features; the decoder's role is to attempt to reconstruct the original input data using these compressed features.

[0033] In one implementation of this invention, the loss function of the deep autoencoder during training is designed as a composite loss function, the mathematical expression of which is: ; in, Total loss; The reconstruction loss is used to measure the difference between the decoder output and the original input, and is usually calculated using the mean square error. It is a preset penalty coefficient, for example, 0.05; This refers to the Lyapunov function value constructed in the previous step, which serves as the loss term for bus voltage stability. During the training process using the backpropagation algorithm, the connection weights between neurons in the network are continuously adjusted. After training, the absolute values ​​of these connection weights reflect the coupling strength between different feature components.

[0034] In one implementation of this invention, the original full-power segment operating dataset is input again into a physical-data hybrid architecture that has eliminated redundant connections. The data flow undergoes only one forward propagation calculation through the encoder portion of the deep autoencoder in this architecture. Because the encoder has undergone physical constraint training and structural simplification, its output low-dimensional vector has eliminated irrelevant couplings in the original data to the greatest extent possible. It should be noted that this process achieves a significant dimensionality reduction effect. For example, an original input data vector containing 32 dimensions (voltage, three-phase current, speed, etc.) will output a low-dimensional vector with no more than 10 dimensions after processing by this optimized encoder.

[0035] Preferably, the physical model layer of the physical-data hybrid architecture includes: In a two-phase rotating coordinate system of a capacitorless motor drive system, the stator voltage equation is established, and the direct-axis voltage and quadrature-axis voltage are expressed as functions of direct-axis current, quadrature-axis current, stator resistance, quadrature-direct-axis inductance, and electric angular velocity, respectively. An electromagnetic torque equation is established to characterize the physical relationship between electromagnetic torque, stator current, and motor parameters; Establish a power balance equation and establish an equation relating the product of DC bus voltage and bus current, and the product of electromagnetic torque and motor speed. The stator voltage equation, electromagnetic torque equation, and power balance equation are used as hard constraints and embedded in the physical model layer of the physical-data hybrid architecture.

[0036] In one implementation of this invention, during synchronous rotation... In a two-phase coordinate system, the stator voltage equations are constructed as a set of mathematical expressions containing two equations: Direct-axis voltage equation: ; Quadrature axis voltage equation: ; in, and These are the direct-axis and quadrature-axis voltages applied to the stator windings; It is the resistance of the stator winding; and These are the direct-axis and quadrature-axis currents flowing through the stator windings; It is a direct-axis inductor; It is a quadrature axis inductor; and It is the rate of change of current with respect to time; It is the rotor's electrical angular velocity, which is equal to the mechanical angular velocity multiplied by the number of pole pairs of the motor; It is the flux linkage constant of the permanent magnet, for example, 0.1 Weber.

[0037] In one implementation of this invention, the specific mathematical expression of the electromagnetic torque equation is as follows: ; in, It is the electromagnetic torque generated by the motor, measured in Newton-meters; This is the number of pole pairs of the motor; the other parameters are defined in the stator voltage equation.

[0038] In one implementation of this invention, under the ideal case where drive system losses are ignored, the power balance equation is constructed as a simple equation: Specifically, the equation can be expanded as follows: ; in, The input power that the inverter obtains from the DC bus is equal to the DC bus voltage. With bus current The product; The mechanical power output by the motor is equal to the electromagnetic torque. With the mechanical angular velocity of the motor The product; This represents the total power loss of the drive system, including copper losses, iron losses, and switching losses.

[0039] It's important to note that treating these three key equations as hard constraints means that during the training of data-driven models (such as deep autoencoders), any set of latent feature representations that does not satisfy these physical equations will suffer a significant loss penalty. This forces the neural network to strictly adhere to these known physical laws while learning the inherent patterns in the data, greatly improving the reliability and interpretability of the final generated features.

[0040] Preferably, the data-driven layer of the physical-data hybrid architecture includes: The full-power range operating dataset is normalized and used as the input data stream for the data-driven layer. A deep autoencoder model consisting of an encoder and a decoder is built as the core of the data-driven layer, where the encoder consists of multiple hidden layers for data compression. The input data stream is fed into the encoder, and through layer-by-layer forward propagation, the input data stream is nonlinearly mapped and compressed into a low-dimensional latent space of a preset dimension. Low-dimensional vectors are extracted from the low-dimensional latent space as data-driven feature vectors, and the deep autoencoder model and the entire process of generating data-driven feature vectors together constitute the data-driven layer; among them, the data-driven feature vectors are used for the identification of coupling connection weights in the physical-data hybrid architecture.

[0041] In one implementation of this invention, a min-max scaling method is used for normalization; specifically, the mathematical expression for this normalization operation is: ;in, These are the original data points. and These are the minimum and maximum values ​​of the feature in the entire dataset, respectively. It is the normalized value between 0 and 1 obtained after calculation.

[0042] Specifically, assume the original input data has a dimension of 32. The encoder is constructed as a neural network with three hidden layers, structured as follows: Input layer (32 neurons) -> Hidden layer 1 (16 neurons) -> Hidden layer 2 (10 neurons) -> Hidden layer 3 (8 neurons). It's important to note that the encoder's network layer structure decreases in size with each layer, forcing the network to continuously compress features and refine information during information propagation. The decoder has a completely opposite, symmetrical structure: Input layer (8 neurons) -> Hidden layer 1 (10 neurons) -> Hidden layer 2 (16 neurons) -> Output layer (32 neurons), its purpose being to reconstruct the compressed features back to the original dimension.

[0043] In one implementation of this invention, a normalized 32-dimensional input data vector is fed into the input layer of the encoder as an input data stream.

[0044] Specifically, during the forward propagation process, the data undergoes one linear transformation and one non-linear activation at each hidden layer. For example, the mathematical expression for the propagation from the input layer to the first hidden layer is: ; in, It is the 16-dimensional output vector of the first hidden layer; It is the input 32-dimensional data vector; It is the weight matrix (dimension 16x32) connecting the input layer and the first hidden layer. It is the bias vector of the first hidden layer (dimension 16). It is a preset nonlinear activation function, such as the modified linear unit (ReLU) function.

[0045] It's important to note that the construction of the entire data-driven layer refers not only to the deep autoencoder model itself, but also includes the complete technical process from data normalization and model building to forward propagation computation to extract data-driven feature vectors. This final output data-driven feature vector will serve as a key data representation, inputting into the loss function of the physical-data hybrid architecture to evaluate the coupling strength between different physical components, i.e., for subsequent coupling connection weight identification.

[0046] Most importantly, the data-driven layer can also be constructed as follows: Construct a neural network model consisting of one-dimensional convolutional layers and fully connected layers; wherein the fully connected layer contains at least one hidden layer located between the input and output ends; The input data stream is fed into a one-dimensional convolutional layer, and a sliding convolution operation is performed on the input data stream through a preset convolutional kernel to capture local temporal patterns and generate local temporal feature maps. The local temporal feature map is input into the fully connected layer, and linear transformation and nonlinear activation function processing are performed sequentially in each hidden layer through layer-by-layer forward propagation to generate nonlinear feature maps at each level. The nonlinear feature map output from the last hidden layer of the neural network model is extracted as a data-driven feature vector, and the neural network model and the entire process of generating the data-driven feature vector together constitute the data-driven layer.

[0047] In one implementation of this invention, the neural network model is designed as a serial structure. The front end is a one-dimensional convolutional layer used to capture local temporal dependencies; the back end is a multi-layer fully connected network used to integrate and map the extracted local features in a high dimension. Specifically, assume the input data is a data segment containing 128 consecutive time steps, each time step containing 32 feature dimensions. The one-dimensional convolutional layer is configured to contain 16 convolutional kernels, each kernel having a size of 3, used to scan these 128 time steps. The fully connected layer consists of two hidden layers, where the first hidden layer contains 64 neurons and the second hidden layer contains 32 neurons.

[0048] Specifically, 16 convolutional kernels of size 3 slide along the time dimension of the data (128 strides). At each stride, the kernel performs a weighted summation operation on the data from the three time steps it covers, thus obtaining a feature value. For example, when a kernel slides to the 5th, 6th, and 7th time steps, it combines the 32-dimensional data from these three moments to calculate an activation value that reflects a specific pattern (such as a small current overshoot) within that tiny time window.

[0049] In one implementation of this invention, the flattened one-dimensional vector is then input into the first hidden layer of the fully connected layer. In this layer, the vector is first multiplied by a weight matrix and then superimposed with a bias vector to complete a linear transformation. The transformation result is then processed by a rectified linear unit (ReLU) nonlinear activation function to increase the expressive power of the model, thereby generating a 64-dimensional output of the first hidden layer, i.e., the nonlinear feature map of the first level.

[0050] It is important to note that this 64-dimensional output will continue to serve as the input for the next layer. Similar linear transformations and nonlinear activation processes will be repeated in the second hidden layer, ultimately generating a second-level, lower-dimensional 32-dimensional nonlinear feature map.

[0051] In one implementation of this invention, the 32-dimensional nonlinear feature map output by the last hidden layer (i.e., the second hidden layer) of the fully connected network in the previous step is directly extracted.

[0052] Specifically, this 32-dimensional vector is named the data-driven feature vector because it incorporates local time-series patterns from the original time-series data and undergoes deep integration through multiple layers of nonlinear transformations. It is considered a highly condensed representation of the inherent laws of the original input data fragments.

[0053] Preferably, step S2, which involves constructing bus voltage stability constraints based on Lyapunov functions and then performing multi-scale analysis and high-dimensional compressed feature decoupling on the full-power segment operating dataset to obtain multi-scale dynamic feature vectors, further includes: The decoupled feature set is segmented into time series, and a sliding time window with a preset range is applied to perform multi-scale analysis to generate a time series data window sequence containing multiple time segments. Short-term fluctuation characteristics and long-term trend characteristics are obtained based on time-series data window sequences; Assess short-term volatility characteristic scores based on short-term volatility characteristics; Assess long-term trend characteristic scores based on long-term trend characteristics; The short-term fluctuation feature score and the long-term trend feature score are weighted and fused according to the preset weight coefficients to calculate the dynamic feature fusion score. Based on the dynamic feature fusion score, a multi-scale dynamic feature vector is constructed from short-term fluctuation features and long-term trend features.

[0054] In one implementation of this invention, a fixed-length sliding time window method is used to process the decoupled feature set. Specifically, it is assumed that the decoupled feature set is a feature sequence containing 1.4 million time points, each with 8 dimensions. The length of the sliding time window is set to 1000 sampling points (corresponding to 1 millisecond of physical duration), and the sliding step size is 100 sampling points (i.e., 90% overlap). It is important to note that a high overlap rate ensures the continuity of the analysis and avoids missing key dynamic change information. The sliding time window starts from the beginning of the sequence, extracts a 1000x8-dimensional data segment each time, then slides forward 100 sampling points, and extracts again.

[0055] In one implementation of this invention, a method combining wavelet transform and gated recurrent unit (GRU) network is used to acquire short-term fluctuation characteristics. Specifically, a time-series data window (1000x8-dimensional) is input to a wavelet transform module, which decomposes it into coefficients representing different frequency components through multi-level decomposition. Special attention is paid to the high-frequency component coefficients, as they typically correspond to short-term, drastic changes such as noise, glitches, and transient shocks in the driving system. Subsequently, these high-frequency coefficient sequences are input into a pre-trained GRU network. The GRU network effectively captures the time dependence of these high-frequency signals and ultimately outputs an 8-dimensional vector characterizing the short-term fluctuation characteristics within the time window; this vector represents the short-term fluctuation characteristics.

[0056] In one implementation of this invention, a method combining a temporal convolutional network (TCN) and an attention mechanism is used to acquire long-term trend features. Specifically, the same time-series data window is input into the TCN. Through its unique causal convolution and dilated convolution structure, the TCN can efficiently learn the long-range dependencies of the data within the entire 1-millisecond time window. At the top layer of the TCN, an attention mechanism module is cascaded, which automatically learns and assigns higher weights to the key time steps that contribute the most to the long-term trend changes. Finally, the model outputs an 8-dimensional vector that encapsulates the main trends and directions of data change within the time window; this vector represents the long-term trend feature.

[0057] In one implementation of this invention, the score is evaluated by calculating the L2 norm of the short-term volatility feature vector. Specifically, it is assumed that the obtained short-term volatility feature vector is... The mathematical expression for calculating its short-term volatility characteristic score is: ;in, This is the final short-term volatility characteristic score. For example, if a feature vector is... The score obtained by calculating its L2 norm is 1.52.

[0058] In one implementation of this invention, the score is evaluated by calculating the sum of the absolute values ​​(L1 norm) of each component of the long-term trend feature vector. Specifically, it is assumed that the obtained long-term trend feature vector is... The mathematical expression for calculating its long-term trend characteristic score is: ;in, This is the final long-term trend feature score. For example, if a feature vector is... The L1 norm score is 0.95.

[0059] Please see Figure 4 In one implementation of this invention, the preset weighting coefficients are obtained based on expert experience or through experimental optimization, aiming to balance the importance of short-term disturbances and long-term trends in the overall assessment. Specifically, a weighting coefficient for short-term fluctuation characteristics is assumed. The weighting coefficient for long-term trend characteristics is set to 0.4. It is set to 0.6. The mathematical expression for calculating the dynamic feature fusion score is: ;in, It is the final dynamic feature fusion score.

[0060] In one implementation of this invention, the original short-term fluctuation feature vector, long-term trend feature vector, and calculated dynamic feature fusion score are concatenated. This construction process uses vector concatenation operation, which directly connects and fuses the features in terms of dimensions.

[0061] Preferably, obtaining short-term fluctuation characteristics and long-term trend characteristics based on time-series data window sequences includes: Wavelet transform is performed on the time series data window sequence, and then a pre-defined gated cyclic unit network is used to capture the time series dependence of the decomposed high-frequency components in order to extract short-term fluctuation features. The time series data window sequence is input into a long-term feature extraction model consisting of a cascaded temporal convolutional network and an attention mechanism module. This model captures long-term temporal dependencies and focuses on key time steps by the attention mechanism to extract long-term trend features.

[0062] In one implementation of this invention, a 1000x8 dimensional data segment extracted from a time-series data window sequence is first processed. Specifically, "db4" (Daubechies4) is selected as the wavelet basis function, and a 3-level discrete wavelet transform decomposition is performed independently on each feature dimension (a total of 8 dimensions) of the data segment.

[0063] It is important to note that the db4 wavelet basis was chosen because of its good time-frequency localization characteristics, making it suitable for analyzing transient impact signals in motor drive systems. After three levels of decomposition, the original signal is decomposed into a low-frequency approximate component and three high-frequency detail components at different scales.

[0064] In another implementation of this invention, the focus is on extracting the highest frequency detail component sequence (usually denoted as cD1), as it best reflects the rapid and drastic fluctuations caused by power device switching, current surges, etc. These eight-dimensional high-frequency detail component sequences are then merged into a new time-series data set with the same dimensions as the original data window, but containing only high-frequency information.

[0065] It should be noted that this high-frequency component sequence is then input into a pre-trained gated recurrent unit (GRU) network. This GRU network is designed with two hidden layers, each containing 32 units, and its function is to learn and capture the temporal evolution patterns and intrinsic correlations of these high-frequency impact signals. After processing the entire sequence, the hidden layer state at the last time step is used as the final output. This output is an 8-dimensional vector that condenses the core information of all short-term fluctuations within the entire time window and is formally defined as the short-term fluctuation feature.

[0066] In one implementation of this invention, the core of the long-term feature extraction model is a Temporal Convolutional Network (TCN). Specifically, the TCN is constructed as a deep structure containing four residual blocks, with the dilation factors of its convolutional layers increasing exponentially by 2 (1, 2, 4, 8). It is noteworthy that this exponentially increasing dilated convolutional design allows the TCN to cover the entire input sequence across 1000 time steps with only a small increase in parameters, thus enabling it to capture long-range dependencies of the data throughout the entire time window.

[0067] In one implementation of this invention, the same 1000x8-dimensional temporal data window is input into the TCN model. A self-attention mechanism module is cascaded after the last convolutional layer of the TCN. This attention mechanism module weights the features extracted from all time steps by the TCN. It autonomously learns and calculates the importance score of each time step for judging the overall trend, assigning higher weights to features from key time steps that best represent trend reversals or accelerations (e.g., the response phase after a change in speed command).

[0068] Preferably, step S3 includes the following steps: Step S31: Use the multi-scale dynamic feature vector as the input variable, and extract its corresponding direct-axis voltage control quantity and quadrature-axis voltage control quantity as two-dimensional output variables to construct a deep learning training dataset; Step S32: Construct an initial nonlinear regression model based on the dimensions of the input variables and the two-dimensional output variables; Step S33: Define a composite loss function, which includes a prediction error loss term used to measure the difference between the model's predicted control quantity and the target label, and a bus voltage stability loss term, which is a penalty term for the mathematical constraint equation multiplied by a preset penalty coefficient. Step S34: Iteratively train the initial nonlinear regression model using the deep learning training dataset, minimize the composite loss function through the backpropagation algorithm, and save the network weight parameters after training when the model's prediction error is lower than the preset error threshold, finally generating a nonlinear regression model that can be used for real-time control quantity prediction.

[0069] In one implementation of this invention, it is assumed that a 17-dimensional multi-scale dynamic feature vector generated for each time-series data window is used as the input feature of the model. Specifically, for the multi-scale dynamic feature vector generated around t=0.85 seconds, the timestamp t=0.85 seconds is precisely located in the original full-power segment running dataset. It is important to note that after finding this timestamp, the direct-axis voltage control quantity and quadrature-axis voltage control quantity recorded at that time are extracted, and these two values ​​are combined into a two-dimensional vector, which is then used as the output label corresponding to this input feature. Each 17-dimensional multi-scale dynamic feature vector and its corresponding two-dimensional output label are combined to form a training data pair. This process is repeated for all time-series data windows in the entire dataset, eventually forming a set containing millions of such data pairs, which is the final deep learning training dataset.

[0070] Please see Figure 4 In one implementation of this invention, a fully connected feedforward neural network is constructed as the initial nonlinear regression model. Specifically, the network structure of this model is designed as follows: Input layer: containing 17 neurons, the number of which perfectly matches the dimension of the input features (multi-scale dynamic feature vectors). First hidden layer: containing 128 neurons. Second hidden layer: containing 64 neurons. Third hidden layer: containing 32 neurons. Output layer: containing 2 neurons, the number of which perfectly matches the dimension of the output labels (direct-axis and quadrature-axis voltage control quantities). It should be noted that all hidden layers use the Modified Linear Unit (ReLU) as the activation function to introduce nonlinear expressiveness and accelerate model convergence. The output layer does not use an activation function so that it can directly output voltage regression values ​​of any range.

[0071] In one implementation of this invention, the specific mathematical expression of the composite loss function is as follows: ; in, It is the total loss value that the model needs to minimize in a single iteration; This is the prediction error loss term, calculated using the mean square error, used to measure the voltage predicted by the model. With actual output labels The smaller the Euclidean distance between them, the more accurate the model prediction. This is the bus voltage stability loss term, which directly references the mathematical constraint equations constructed based on the Lyapunov function in the preceding steps. At each step of the training, the following is calculated based on the current state of the drive system: The larger the value, the more the drive system deviates from a stable state. This is a preset penalty coefficient, for example, set to 0.1. It's important to note that this penalty coefficient... Its function is to balance the importance of the two loss terms. By introducing When the model's predictions lead to bus voltage instability (i.e. When the value is large), total loss The number of such events will increase dramatically, forcing the optimization algorithm to adjust the network weights in order to generate a control strategy that is both accurate and does not compromise the stability of the driving system.

[0072] In one implementation of this invention, the deep learning training dataset is randomly divided into a training set and a validation set in an 8:2 ratio. Specifically, the model is iteratively trained using mini-batch stochastic gradient descent, with a batch size of 512. In each iteration, 512 data pairs are randomly selected from the training set, the average composite loss of these 512 samples is calculated, and then the gradient of the loss function with respect to each weight parameter in the network is calculated using the backpropagation algorithm. The weights are then updated slightly in the opposite direction of the gradient. After each complete training cycle (i.e., traversing all training data once), the model's prediction error loss term is evaluated on the validation set. The preset error threshold is set to 0.005. This threshold is used when the model's prediction error on the validation set... When the value first drops below 0.005, the model is considered to have converged sufficiently and reached the expected prediction accuracy. At this point, the training process is immediately stopped, and all weight and bias parameters of the current network are saved to a file. This neural network file containing the trained parameters is the final nonlinear regression model that can be directly deployed for real-time control quantity prediction.

[0073] Please see Figure 4 In another implementation of this invention, the multi-scale dynamic feature vector is... After injecting the stabilizing line, deep neural network training is performed; subsequently, a nonlinear regression model is constructed. Define the loss function .in, For deep neural networks based on input The calculated result; For data centralization and input The corresponding correct answer; This is a manually set coefficient used to balance the two loss terms; These are the learnable weights and biases in a neural network. This is the value of the Lyapunov function.

[0074] Preferably, step S4 includes the following steps: Step S41: Deploy a nonlinear regression model in the controller of the electrolytic capacitor-free motor drive system; Step S42: During the real-time operation of the electrolytic capacitor-free motor drive system, the current state observations are synchronously acquired through the hardware sensor network; Step S43: Input the state observations into the deployed nonlinear regression model, and calculate the predicted direct-axis voltage control quantity and quadrature-axis voltage control quantity in real time through forward propagation; Step S44: Convert the predicted direct-axis voltage control quantity and quadrature-axis voltage control quantity into real-time control commands for the power switching devices in the motor drive inverter to execute motor drive control, thereby forming an electrolytic capacitor-free motor drive strategy.

[0075] Please see Figure 3 In one implementation of this invention, the nonlinear regression model file containing network weight parameters is optimized and quantized using a dedicated model conversion tool. Specifically, the optimization process includes removing redundant operations and fusing layers to reduce model size and computational complexity. The quantization process converts the weight parameters, originally 32-bit floating-point numbers, into 16-bit or 8-bit fixed-point numbers to adapt to the computational characteristics of embedded processors (such as digital signal processors) and significantly improve computational speed.

[0076] In this embodiment of the invention, the controller performs data acquisition and feature construction once at the frequency of its main control cycle (e.g., every 10 microseconds). Specifically, in At this moment, the controller synchronously acquires raw state observations, such as the initial rotational speed, speed command, slip signal, and six-dimensional input variables of the three-phase stator current, through its analog-to-digital converter interface and quadrature encoder pulse interface. It's important to note that the controller does not directly feed these raw state observations into the model. Instead, a dedicated computing unit or coprocessor integrated within the controller immediately initiates a high-speed, forward computation process. This process rigorously replicates all data processing steps from data decoupling to multi-scale analysis during offline training.

[0077] Specifically, the model performs a complete forward propagation computation. The data starts from 17 input neurons, flows through three hidden layers with 128, 64, and 32 neurons, and finally yields the computation result at the two output neurons.

[0078] It should be noted that this forward propagation computation process is essentially a series of highly parallel matrix multiplications and additions, executed on modern digital signal processors using optimized math libraries, resulting in extremely short execution times. For example, in At time 1, the model calculates and outputs a pair of predicted voltage control values ​​based on the input feature vector: {predicted direct-axis voltage control value: 0.98 volts, predicted quadrature-axis voltage control value: 1.45 volts}.

[0079] In one implementation of this invention, the coordinate transformation module in the controller immediately receives the voltage control quantities [0.98, 1.45] output by the model. Specifically, the module first performs an inverse Park transformation to convert the voltage components in the dq coordinate system to the α-β stationary coordinate system; then it performs an inverse Clark transformation to convert them into the A, B, and C three-phase instantaneous voltage commands in the three-phase stationary coordinate system.

[0080] In one implementation of this invention, the three-phase voltage command is then sent to the space vector pulse width modulation module. Based on the magnitude and phase of the voltage command, this module calculates in real time the duty cycle of the pulse width modulation waveform required to drive the six power switching devices in the inverter, and generates six corresponding high-frequency gate drive signals. These signals directly control the on / off state of the power switching devices, thereby precisely chopping and modulating the high-voltage DC power on the DC bus into the desired three-phase AC power to drive the motor.

[0081] Of particular importance is that, while executing voltage modulation in the electrolytic capacitor-free motor drive strategy, the controller synchronously feeds back and adjusts real-time control commands, which also includes: Obtain the real-time DC bus voltage value; Calculate the fluctuation range of the real-time DC bus voltage during stable operation, and then take the arithmetic mean of the upper and lower thresholds of the fluctuation range as the nominal bus voltage value. The real-time voltage deviation is analyzed based on the nominal bus voltage value. When the absolute value of the real-time voltage deviation exceeds the preset deviation threshold, it is determined that the bus voltage has fluctuated, and the difference in voltage deviation between adjacent monitoring periods is calculated as the voltage deviation change rate. Using the proportional-integral-derivative controller within the controller, the voltage compensation adjustment is calculated based on the real-time voltage deviation and the rate of change of voltage deviation. The voltage compensation adjustment is superimposed on the direct and quadrature axis components of the real-time control command according to a preset ratio to generate a corrected control command, so as to achieve differentiated control and stable operation of the system.

[0082] In one implementation of this invention, the controller performs an instantaneous sampling of the bus voltage through its built-in high-precision analog-to-digital converter interface, which is directly connected to the DC bus, during each high-speed control cycle.

[0083] It is important to note that the "fluctuation range of the stable operating period" required here is prior knowledge obtained during the model building phase through offline analysis of the full-power range operating dataset, and is not dynamically calculated during the real-time operation of the drive system.

[0084] In one implementation of this invention, the controller compares the currently acquired real-time voltage value of 311.5 volts with the stored nominal bus voltage value of 310.0 volts, calculating the current real-time voltage deviation as +1.5 volts. The drive system presets a deviation threshold of 1.0 volt. Since the absolute value of the currently calculated real-time voltage deviation (1.5 volts) is greater than this threshold, the controller immediately determines that the bus voltage has fluctuated and requires intervention. It should be noted that at this time, the controller will also immediately look up and read the voltage deviation recorded in the previous control cycle, assuming its value is +1.2 volts. Subsequently, by calculating the difference between the current deviation and the deviation of the previous cycle, the voltage deviation change rate is obtained as +0.3 volts, which reflects that the voltage deviation is continuously increasing at a relatively fast rate.

[0085] In one implementation of this invention, the controller integrates a digital proportional-integral-derivative (PI-DE) controller that comprehensively considers three aspects to generate the final compensation command. Specifically, its proportional part generates a compensation component proportional to the current real-time voltage deviation of 1.5 volts, used for rapid response to the current error; its integral part generates a compensation component to eliminate the steady-state error of the drive system based on the cumulative value of historical deviations; and its derivative part generates a predictive compensation component to suppress future fluctuation overshoot based on the current voltage deviation change rate of +0.3 volts.

[0086] Specifically, assume the allocation ratio is set to superimpose 40% of the compensation adjustment onto the direct-axis voltage component and 60% onto the quadrature-axis voltage component. Simultaneously, assume the real-time control command output by the nonlinear regression model is 0.98 V for the direct-axis voltage and 1.45 V for the quadrature-axis voltage. The controller performs superposition calculations, generating a corrected direct-axis voltage control value of 1.166 V and a corrected quadrature-axis voltage control value of 1.729 V.

[0087] The present invention also provides a system 100 for constructing a dynamic control model for a motor drive, which executes the method for constructing a dynamic control model for a motor drive as described above. The system for constructing a dynamic control model for a motor drive includes: The operating data acquisition module 101 is used to acquire the operating data of the electrolytic capacitor-free motor drive system under different control parameters, obtain state observations including bus voltage, stator current, speed and torque, as well as corresponding control quantities, and form a full power range operating dataset. The dynamic feature decoupling module 102 is used to construct bus voltage stability constraints, and then perform multi-scale analysis and high-dimensional compressed feature decoupling on the full power segment operation dataset to obtain multi-scale dynamic feature vectors. The prediction model building module 103 is used to build a nonlinear regression training dataset based on multi-scale dynamic feature vectors, and to build a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraint. The drive strategy deployment module 104 is used to deploy a nonlinear regression model in the controller of the electrolytic capacitor-free motor drive system. By acquiring state observations in real time as inputs to the nonlinear regression model, and using the predictive control quantities output by the nonlinear regression model to execute motor drive control, a nonlinear capacitor-free motor drive strategy is formed.

[0088] Please see Figure 3 This diagram illustrates the electrolytic capacitor-free motor drive strategy based on intelligent voltage regulation in this invention. It shows the monitoring and control logic structure of the fuel cell stack and converter system. When the system starts up, the fuel cell stack and DC / DC converter begin operation. The observation module collects operating data from the switching circuit group, etc., while the control parameter set collects relevant control information and forms residual polarization data pairs. After processing by the ISC (Independent Voltage Suppressor), the output control quantity regulates the system. Simultaneously, the relevant data of the observation and control quantities are transmitted to the lower-level module, where they undergo dimensionality reduction, data features are visualized through a characterization process, and finally, a nonlinear regression model is input, outputting... , These results enable effective monitoring, control, and analysis of the fuel cell stack and converter system.

[0089] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0090] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method of constructing a dynamic control model of an electric machine drive, characterized by The method comprises the following steps: Step S1: Collecting the running data of the electrolytic capacitorless motor drive system under different control parameters, obtaining the state observation quantities including bus voltage, stator current, speed and torque, and corresponding control quantities, and forming a full-power segment running data set; Step S2: Constructing a bus voltage stability constraint, and then performing multi-scale analysis and high-dimensional compressed feature decoupling on the full-power segment running data set to obtain a multi-scale dynamic feature vector; Step S3: Constructing a nonlinear regression training data set according to the multi-scale dynamic feature vector, and constructing a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraint; Step S4: Deploying the nonlinear regression model in the controller of the electrolytic capacitorless motor drive system; By real-time acquisition of the state observation quantities as the input of the nonlinear regression model, and by using the predicted control quantities output by the nonlinear regression model to perform motor drive control, an electrolytic capacitorless motor drive strategy is formed.

2. The method of claim 1, wherein, Step S1 comprises the following steps: Step S11: Initialization according to the rated power range of the electrolytic capacitorless motor drive system, setting the initial speed parameter group and the speed variation parameter group covering the full power segment; Step S12: Based on the initial speed parameter group and the speed variation parameter group, different initial speeds are combined with corresponding speed variation parameters to form multiple control parameter configurations; Step S13: Under each control parameter configuration, continuously sample the running process of the electrolytic capacitorless motor drive system at a preset sampling frequency, and obtain the DC bus voltage value, bus current, three-phase stator current, actual speed and actual torque as the state observation quantities; Step S14: Synchronously record the direct-axis voltage control quantity and the quadrature-axis voltage control quantity at the corresponding time; Step S15: Time-align the state observation quantities with the direct-axis voltage control quantity and the quadrature-axis voltage control quantity, and bind them as a full-power segment running data set.

3. The method of claim 2, wherein, In step S2, the bus voltage stability constraint is constructed, and then multi-scale analysis and high-dimensional compressed feature decoupling are performed on the full-power segment running data set to obtain a multi-scale dynamic feature vector, which comprises: Based on the Lyapunov function, a mathematical constraint equation for quantifying the bus voltage stability constraint is constructed; A physical-data hybrid architecture is constructed, which includes a physical model layer and a data-driven layer; In the physical-data hybrid architecture, the bus voltage stability loss term constructed based on the mathematical constraint equation is used as a penalty term to identify the coupling connection weights between the voltage components, current components, load components and power components, and when the connection weight is lower than a preset redundancy threshold, the connection is removed; Dynamic decoupling feature extraction is performed on each component after removing the redundant connections, thereby obtaining a decoupled feature set.

4. The method of claim 3, wherein, The construction of the physical model layer of the physical-data hybrid architecture comprises: In the two-phase rotating coordinate system of the electrolytic capacitorless motor drive system, the stator voltage equation is established, and the direct-axis voltage and the quadrature-axis voltage are expressed as the functions of the direct-axis current, the quadrature-axis current, the stator resistance, the direct-quadrature axis inductance and the electrical angular velocity, respectively; An electromagnetic torque equation is established to represent the physical relationship between the electromagnetic torque and the stator current and the motor parameters; The power balance equation is established to associate the product of the DC bus voltage value and the bus current with the product of the electromagnetic torque and the motor speed by an equation; The stator voltage equation, the electromagnetic torque equation and the power balance equation are taken as hard constraints and embedded into the physical model layer of the physics-data hybrid architecture.

5. The method of claim 3, wherein the dynamic control model of the motor drive is constructed by, The data-driven layer of the physics-data hybrid architecture is constructed by: normalizing the full-power operating data set to construct an input data stream for the data-driven layer; building a deep autoencoder model composed of an encoder and a decoder as the core of the data-driven layer, wherein the encoder is composed of multiple hidden layers for data compression; inputting the input data stream into the encoder, and nonlinearly mapping and compressing the input data stream to a low-dimensional latent space of a preset dimension through layer-by-layer forward propagation; extracting a low-dimensional vector in the low-dimensional latent space as a data-driven feature vector, and the deep autoencoder model and the whole process of generating the data-driven feature vector together constitute the data-driven layer; wherein the data-driven feature vector is used for coupling connection weight identification of the physics-data hybrid architecture.

6. The method of claim 3, wherein the dynamic control model of the motor drive is constructed by, The step S2 of constructing the bus voltage stability constraint, and then performing multi-scale analysis and high-dimensional compressed feature decoupling on the full-power operating data set to obtain a multi-scale dynamic feature vector further includes: segmenting the decoupled feature set by time series and applying a preset range of sliding time windows to perform multi-scale analysis to generate a time series data window sequence containing multiple time segments; obtaining short-term fluctuation features and long-term trend features based on the time series data window sequence; evaluating a short-term fluctuation feature score according to the short-term fluctuation features; evaluating a long-term trend feature score according to the long-term trend features; weighting and fusing the short-term fluctuation feature score and the long-term trend feature score according to a preset weight coefficient to calculate a dynamic feature fusion score; constructing a multi-scale dynamic feature vector based on the dynamic feature fusion score from the short-term fluctuation features and the long-term trend features.

7. The method of claim 6, wherein the dynamic control model of the motor drive is constructed by, The step of obtaining short-term fluctuation features and long-term trend features based on the time series data window sequence includes: performing wavelet transform on the time series data window sequence, and then using a preset gated recurrent unit network to capture the time series dependence of the high-frequency components after decomposition to extract short-term fluctuation features; inputting the time series data window sequence into a long-term feature extraction model composed of a time convolution network and an attention mechanism module in cascade to capture long-time series dependence and focus on key time steps by the attention mechanism to extract long-term trend features.

8. The method of claim 1, wherein, The step S3 includes the following steps: Step S31: taking the multi-scale dynamic feature vector as an input variable, and extracting its corresponding direct-axis voltage control quantity and quadrature-axis voltage control quantity as two-dimensional output variables to construct a deep learning training data set; Step S32: constructing an initial nonlinear regression model based on the dimensions of the input variable and the two-dimensional output variable; Step S33: defining a composite loss function, which includes a prediction error loss term for measuring the difference between the model predicted control quantity and the target label, and a bus voltage stability loss term that is a penalty term of the mathematical constraint equation multiplied by a preset penalty coefficient; Step S34: iteratively train the initial nonlinear regression model using the deep learning training dataset, minimize the composite loss function through the back propagation algorithm, save the trained network weight parameters when the prediction error of the model is lower than the preset error threshold, and finally generate a nonlinear regression model that can be used for real-time control quantity prediction.

9. The method of claim 1, wherein, Step S4 includes the following steps: Step S41: deploying the nonlinear regression model in the controller of the electrolytic capacitor-free motor drive system; Step S42: synchronously acquiring the state observation at the current time through the hardware sensor network during real-time operation of the electrolytic capacitor-free motor drive system; Step S43: inputting the state observation into the deployed nonlinear regression model and calculating the predicted direct-axis voltage control quantity and quadrature-axis voltage control quantity through forward propagation; Step S44: converting the predicted direct-axis voltage control quantity and quadrature-axis voltage control quantity into real-time control instructions of the power switch devices in the motor drive inverter to perform motor drive control, thereby forming an electrolytic capacitor-free motor drive strategy.

10. A system for constructing a dynamic control model of an electric machine drive, characterized by The method for constructing a dynamic control model of a motor drive as claimed in claim 1, the system for constructing a dynamic control model of a motor drive comprises: a working condition data acquisition module for acquiring running data of the electrolytic capacitor-free motor drive system under different control parameters, obtaining state observations including bus voltage, stator current, speed and torque, and corresponding control quantities, and forming a full-power operating data set; a dynamic characteristic decoupling module for constructing bus voltage stability constraints, and then performing multi-scale analysis and high-dimensional compressed feature decoupling on the full-power operating data set to obtain a multi-scale dynamic characteristic vector; a prediction model construction module for constructing a nonlinear regression training dataset according to the multi-scale dynamic characteristic vector, and constructing a nonlinear regression model that can be used for real-time control quantity prediction under the bus voltage stability constraints; a drive strategy deployment module for deploying the nonlinear regression model in the controller of the electrolytic capacitor-free motor drive system; by acquiring the state observation as the input of the nonlinear regression model in real time, and using the predicted control quantity output by the nonlinear regression model to perform motor drive control, thereby forming an electrolytic capacitor-free motor drive strategy.

Images