Permanent magnet synchronous motor state intelligent prediction method and system based on time-frequency-space analysis

Abstract

The application provides a permanent magnet synchronous motor state intelligent prediction method and system based on time-frequency-space analysis, and belongs to the field of permanent magnet synchronous motor state prediction.The method firstly extracts global time domain information of motor physical quantities, then extracts local time domain and space information and frequency domain information, fuses the obtained information, and corresponds with output physical quantities to construct a permanent magnet synchronous motor state intelligent prediction model; after training and testing the model based on historical data, a mature permanent magnet synchronous motor state intelligent prediction model is obtained; then input physical quantities to be tested are input into the mature permanent magnet synchronous motor state intelligent prediction model, output predicted output physical quantities are output, and motor state prediction is performed.The application captures high-frequency characteristics and mixed periodicity of motor physical quantities themselves, dynamically captures nonlinear coupling relationships among motor physical quantities, improves state prediction precision and accuracy, and improves real-time performance of permanent magnet synchronous motor state prediction.

Description

Technical Field

[0001] This invention belongs to the field of permanent magnet synchronous motor state prediction, specifically relating to a method and system for intelligent state prediction of permanent magnet synchronous motors based on time-frequency-space analysis. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) use permanent magnets for excitation, resulting in a simpler motor structure. This reduces processing and assembly costs and eliminates the need for potentially problematic slip rings and brushes, improving motor reliability. Furthermore, the absence of excitation current and losses increases efficiency and power density. Compared to other motors, PMSMs offer long service life and high safety, leading to numerous applications in aerospace, automotive manufacturing, home appliances, and intelligent robotics. Therefore, the condition of PMSMs is crucial for ensuring normal operation within these industries.

[0003] With the development of related intelligent industries, higher requirements have been placed on the state prediction of permanent magnet synchronous motors. Currently, the main method for predicting the state of permanent magnet synchronous motors is to use traditional machine learning models based on existing relationships between motor physical quantities. Traditional machine learning models extract a large number of features of motor physical quantities for modeling, but these models still cannot solve the following problems: (1) The relationships between physical quantities of permanent magnet synchronous motors are nonlinear. Although there are definite formulas to describe the relationships between motor physical quantities, in actual operating conditions, parameters such as temperature and resistance will change continuously, while these parameters in the formula are constant values. Therefore, the described relationships between motor physical quantities have a large deviation from the actual relationships; (2) The physical quantity data generated by permanent magnet synchronous motors are discrete time series. Traditional machine learning models rely too much on short-term historical information and ignore the influence of long-term historical information. Some deep learning models use methods such as LSTM to learn long-term historical information in the sequence and ignore the influence of short-term historical information. Inappropriate data processing will also lead to varying degrees of time lag in the predicted motor physical quantities; (3) The physical quantity data generated by permanent magnet synchronous motors are high-frequency and have mixed periodicity. Traditional machine learning models can only capture low-frequency or single-periodic information in the data.

[0004] Deep learning, due to its powerful modeling and function fitting capabilities, is commonly used to handle a wide variety of time series modeling problems. Currently, a few methods also use deep learning models for state prediction of permanent magnet synchronous motors, but these methods typically only solve one of the aforementioned challenges; there is still no method that can simultaneously solve all three challenges. Summary of the Invention

[0005] In view of the above problems, embodiments of the present invention provide a method and system for intelligent prediction of the state of a permanent magnet synchronous motor based on time-frequency-space analysis.

[0006] The technical solution adopted in the embodiments of the present invention is as follows:

[0007] In a first aspect, embodiments of the present invention provide a method for intelligent prediction of the state of a permanent magnet synchronous motor based on time-frequency-space analysis, comprising the following steps:

[0008] Step S1: Set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor;

[0009] Step S2: Based on the input physical quantity, a global time-domain information representation model is constructed using a long short-term memory neural network to obtain the global time-domain information of the input physical quantity in the time domain;

[0010] Step S3: Based on global temporal information, a spatiotemporal convolutional neural network STConV is used to construct a local temporal and spatial information representation model to obtain the local temporal and spatial information of the input physical quantity;

[0011] Step S4: Based on global time-domain information, a frequency domain information representation model is constructed using a spectrum analysis network (SFD) to obtain the frequency domain information of the input physical quantity;

[0012] Step S5: The local time-domain and spatial information and frequency-domain information are fused, and the fused information is matched with the output physical quantity to obtain the intelligent prediction model of the permanent magnet synchronous motor state.

[0013] Step S6: Collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor, divide the historical data into training data and test data, and perform data preprocessing to obtain the training set and test set respectively.

[0014] Step S7: The intelligent prediction model of permanent magnet synchronous motor state is trained using the training set. After training, the parameters corresponding to each model are obtained, thus obtaining a mature intelligent prediction model of permanent magnet synchronous motor state. Then, the intelligent prediction model of permanent magnet synchronous motor state is tested using the test set.

[0015] Step S8: Collect the M-element input physical quantity of the permanent magnet synchronous motor at the current moment, and input the input physical quantity into the mature permanent magnet synchronous motor state intelligent prediction model to obtain the predicted N-element output physical quantity, and determine the state of the permanent magnet synchronous motor.

[0016] In a preferred embodiment of the present invention, the input physical quantities include four-axis voltages of d, q, a, and b, and the output physical quantities include rotational speed, torque, d-axis current, and q-axis current, with M=4 and N=4.

[0017] In a preferred embodiment of the present invention, in step S2, the global time-domain information representation model is as shown in equation (1):

[0018] h t =G(h t-1 x t (1)

[0019] In equation (1), x t ∈R M*1 h represents the input physical quantity at time t. t-1 h t ∈R M*1 These represent the global time-domain information at times t-1 and t, respectively. The global time-domain information at all times constitutes a time series, and G is a built-in function of the LSTM network.

[0020] In a preferred embodiment of the present invention, in step S3, the STConV network can simultaneously perform convolution on the temporal and spatial domains, and the constructed local temporal and spatial information representation model is shown in equation (2):

[0021]

[0022] In equation (2), h represents the local temporal and spatial information acquired at time t. t-w+1：t This represents the global temporal information from time t-w+1 to time t, where w is the window length of the global temporal information time series for information extraction; k is the number of convolutional kernels in the STConV network, and α... si ∈R M*1 α fi ∈R w*1 This represents two newly introduced convolution kernels along W. i Attention is calculated in the spatial and temporal dimensions.

[0023] In a preferred embodiment of the present invention, in step S3, the SFD network includes a Fourier transform (FFT) layer, a spatial linear layer, and an inverse Fourier transform (iFFT) layer.

[0024] As a preferred embodiment of the present invention, the frequency domain information representation model is as shown in equation (3):

[0025]

[0026] In equation (3), Let A ∈ R be the Fourier transform function and the inverse Fourier transform function, respectively. F*M*M , B∈R F*M These are parameters in a spatially linear layer. This means converting global time-domain information into learnable frequency-domain information using a Fourier transform. This means that the learned frequency domain information is converted into a time domain representation after inverse Fourier transform;

[0027] Pick In This serves as the learned frequency domain information at time t.

[0028] In a preferred embodiment of the present invention, the fusion in step S5 includes the following steps:

[0029] Step S51: The local time-domain and spatial information obtained through STConV and the frequency-domain information obtained through SFD are concatenated, using the following formula:

[0030]

[0031] In equation (4), These are the local time-domain and spatial information representations and the frequency-domain information representations obtained at time t, respectively. t It is the fused information obtained at time t.

[0032] Step S52: Input the fused information into the transformation linear layer and correlate it with the output physical quantity at the current time to obtain the final model architecture; the correspondence model between the fused information and the output physical quantity is shown in formula (5):

[0033]

[0034] In equation (5), This represents the output physical quantity at time t, and W′ and bias are the weight and bias parameters of the linear layer.

[0035] In a preferred embodiment of the present invention, step S6 includes data preprocessing, which includes standardization and sliding windowing; wherein...

[0036] The formula for standardizing the training data is as follows:

[0037] p' = p min +(pp min ) / (p max -p min (6)

[0038] In formula (6), p min p max Let p and p' represent the minimum and maximum values ​​in sequence p, respectively, where p represents the original sequence and p' represents the standardized sequence.

[0039] When standardizing the test data, formula (6) is still used, where p min p max From the training set;

[0040] The sliding window is used to extract input data from standardized data by using a window size w as the basic structural unit length, thereby reshaping the input data into a sample of a specified length.

[0041] In a preferred embodiment of the present invention, step S7, the test specifically includes: inputting the input physical quantities in the test set after data preprocessing into the mature permanent magnet synchronous motor state prediction intelligent model, outputting the predicted output physical quantities, and calculating the mean square error (MSE) between the predicted data and the actual data of the output physical quantities to evaluate the model.

[0042] The formula for evaluating the model using MSE is as follows:

[0043]

[0044] In equation (7), y i This represents the i-th element in the output physical quantity. Let represent the actual value and predicted value of the i-th output physical quantity at time t, respectively.

[0045] Secondly, embodiments of the present invention also provide an intelligent prediction system for the state of a permanent magnet synchronous motor based on time-frequency-space analysis. The system includes: a data acquisition module, a data preprocessing module, a global time-domain information representation module, a local time-domain and spatial information representation module, a frequency-domain information representation module, an intelligent prediction model construction module, and a result output module; wherein...

[0046] The data acquisition module is used to set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor, and to collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor, as well as to collect the M-element input physical quantity to be predicted at the current moment.

[0047] The data preprocessing module is used to divide historical data into training data and test data, and to obtain training sets and test sets after performing data preprocessing respectively.

[0048] The global time-domain information representation module is used to construct a global time-domain information representation model based on the input physical quantity using a long short-term memory neural network, and send the obtained global time-domain information of the input physical quantity in the time domain to the local time-domain and spatial information representation module and the frequency-domain information representation module.

[0049] The local temporal and spatial information representation module is used to construct a local temporal and spatial information representation model based on global temporal information using a spatiotemporal convolutional neural network, and output local temporal and spatial information;

[0050] The frequency domain information representation module is used to construct a frequency domain information representation model based on global time domain information using a spectrum analysis network (SFD), and output the frequency domain information of the input physical quantities.

[0051] The intelligent prediction model construction module is used to fuse the local time-domain, spatial, and frequency-domain information, and map the fused information to the output physical quantity to construct an intelligent prediction model for the permanent magnet synchronous motor (PMSM) state. It is also used to train the PMSM state intelligent prediction model using a training set, obtaining the parameters corresponding to each model after training, thus obtaining a mature PMSM state intelligent prediction model. The model is then tested using a test set. Furthermore, it is used to input the collected M-element input physical quantity at the current moment into the mature PMSM state intelligent prediction model to obtain the predicted N-element output physical quantity, and send the prediction result to the result output module.

[0052] The result output module is used to output the predicted N-element output physical quantity to determine the state of the permanent magnet synchronous motor.

[0053] The present invention has the following beneficial effects:

[0054] The present invention provides a method and system for intelligent prediction of permanent magnet synchronous motor state based on time-frequency-space analysis. This method addresses the time lag problem in motor state prediction by learning global information of motor physical quantities in the time domain. Furthermore, it simultaneously learns frequency domain information of motor physical quantities, as well as spatial and local time domain information between these quantities. Learning the frequency domain information captures the high-frequency characteristics and mixed periodicity of the motor physical quantities themselves, improving the accuracy of motor state prediction. Learning the spatial information between motor physical quantities dynamically captures the nonlinear coupling relationships between them, making the motor prediction more consistent with reality and improving the accuracy of the prediction results. Learning the local time domain information of motor physical quantities allows for better learning of short-term state changes, further improving the accuracy of motor state prediction. By analyzing motor physical quantities from the time, frequency, and space directions, various features in the data can be fully captured, improving the accuracy, precision, and real-time performance of motor state prediction. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1This is a flowchart of the intelligent state prediction method for permanent magnet synchronous motors based on time-frequency-space analysis proposed in an embodiment of the present invention.

[0057] Figure 2 This is a network structure diagram of the intelligent prediction model for the state of a permanent magnet synchronous motor based on time-frequency-space analysis proposed in an embodiment of the present invention. Detailed Implementation

[0058] The technical problems, technical solutions, and advantages of the present invention will be explained in detail below with reference to exemplary embodiments and accompanying drawings. The exemplary embodiments described below are for illustrative purposes only and should not be construed as limiting the invention. Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless defined herein.

[0059] The inventors of this application, through detailed research, discovered that there is currently no deep learning model specifically designed for permanent magnet synchronous motor (PMSM) state prediction from the three perspectives of time, frequency, and space. Traditional machine learning models and general deep learning models cannot learn the information of the motor's physical quantities themselves or the relationships between electrical physical quantities. Based on this, this application proposes an intelligent prediction method and system for PMSM state based on time-frequency-space analysis. This method can learn the global, local time-domain, and frequency-domain information of the motor's physical quantities, solving the time lag problem in PMSM state prediction. It can also extract the high-frequency and mixed periodic characteristics of the data itself, greatly improving the accuracy of motor state prediction. While the relationships between motor physical quantities have clear formula definitions, parameters such as temperature and resistance change constantly in reality. Therefore, fixed formulas cannot accurately describe the relationships between motor physical quantities. Currently, some works use convolutional neural networks (CNNs) to dynamically learn the relationships between motor physical quantities, but these CNNs only have one convolutional kernel, resulting in a limited receptive field and insufficient ability to fully learn the relationships between motor physical quantities. This invention uses a CNN with multiple convolutional kernels, expanding the network's receptive field and significantly improving the real-time performance and accuracy of motor state prediction.

[0060] See Figure 1 and Figure 2 The present invention provides an intelligent prediction method for the state of a permanent magnet synchronous motor based on time-frequency-space analysis. The intelligent prediction method includes the following steps:

[0061] Step S1: Set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor.

[0062] In this step, the input and output physical quantities are set according to the actual working state of the permanent magnet synchronous motor. For example, the input physical quantities include the voltages of the four axes d, q, a, and b, and the output physical quantities include the speed, torque, d-axis current, and q-axis current, with M=4 and N=4.

[0063] Step S2: Analyze the global time-domain information of the input physical quantity. Based on the input physical quantity, a global time-domain information representation model is constructed using a Long and Short Time Memory (LSTM) neural network to obtain the global time-domain information of the input physical quantity in the time domain.

[0064] In this step, the global time-domain information representation model constructed is shown in equation (1):

[0065] h t =G(h t-1 x t (1)

[0066] In equation (1), x t ∈R M*1 h represents the input physical quantity at time t. t-1 h t ∈R M*1 These represent the global time-domain information at time t-1 and t, respectively. The global time-domain information at all times constitutes a time series, and G is a built-in function of the LSTM network.

[0067] The set input physical quantity is input into the global time domain information representation model to obtain the global time domain information of each input physical quantity at each time step in the time domain.

[0068] Step S3: Analyze the local temporal and spatial information of the input physical quantity. Based on the global temporal information, a spatiotemporal convolutional neural network (STConV) is used to construct a representation model of the local temporal and spatial information to obtain the local temporal and spatial information of the input physical quantity.

[0069] In this step, the STConV network can simultaneously perform convolutions in the temporal and spatial domains, and the constructed local temporal and spatial information representation model is shown in Equation (2):

[0070]

[0071] In equation (2), h represents the local temporal and spatial information acquired at time t. t-w+1：t This represents the global temporal information from time t-w+1 to time t, where w is the window length of the global temporal information time series for information extraction; k is the number of convolutional kernels in the STConV network, and α... si ∈RM*1 α fi ∈R w*1 This represents two newly introduced convolution kernels along W. i Attention is calculated in both spatial and temporal dimensions, where i represents any convolutional kernel.

[0072] The global time-domain information from time t-w+1 to time t is input into the local time-domain and spatial information representation model to obtain the local time-domain and spatial information at time t.

[0073] By performing multi-kernel convolution on the time series composed of global time-domain information to learn local time-domain information, the problem of lag in the prediction of motor physical quantities is solved. Performing multi-kernel convolution in the spatial dimension expands the receptive field of the network, enabling the network to learn more information. By combining the information learned by different convolution kernels through attention, the accuracy of dynamically learning the relationship between motor physical quantities is greatly improved, as well as the accuracy and precision of motor physical quantity prediction.

[0074] Step S4: Analyze the frequency domain information of the input physical quantity. Based on the global time domain information, a frequency domain information representation model is constructed using a spectrum analysis network (SFD) to obtain the frequency domain information of the input physical quantity.

[0075] In this step, the SFD network includes a Fourier Transform (FFT) layer, a spatial linear layer, and an Inverse Fourier Transform (iFFT) layer. First, the global time-domain information time series of length w is transformed into a series of length w using a Fourier Transform. The frequency domain sequence is then input into a spatial linear layer to learn frequency domain information. Finally, the output of the spatial linear layer is input into an inverse Fourier transform layer to convert the frequency domain sequence into a time series in the time domain. Then, the frequency domain information at time t is extracted from the time series in the time domain representation. The model formula is shown in equation (3):

[0076]

[0077] In equation (3), Let A ∈ R be the Fourier transform function and the inverse Fourier transform function, respectively. F*M*M , B∈R F*M These are parameters in a spatially linear layer. This means converting global time-domain information into learnable frequency-domain information using a Fourier transform. This means that after inverse Fourier transform, the learned frequency domain information is converted into a time domain representation. Therefore, the learned frequency domain information at time t is:

[0078] The global time-domain information from time t-w+1 to time t is input into the frequency-domain information representation model to obtain the frequency-domain information from time t-w+1 to time t. By performing spectral analysis on the time series composed of the global time-domain information, the high-frequency, mixed periodicity characteristics of the motor's physical quantities can be captured from the frequency-domain information.

[0079] Step S5: The local time-domain and spatial information and frequency-domain information are fused, and the fused information is mapped to the output physical quantity to obtain the intelligent prediction model of the permanent magnet synchronous motor state. The model uses M-element input physical quantity and N-element output physical quantity as input and output, including a global time-domain information representation model, a local time-domain and spatial information representation model, a frequency-domain information representation model, a fusion model, and a model for mapping fused information to output physical quantity.

[0080] In this step, the feature representations of STConV and SFD outputs for permanent magnet synchronous motor state prediction are fused, and then the fused information is input into a transformation linear layer. The output of the transformation linear layer is mapped to the output physical quantity corresponding to the current feature representation to obtain a smart prediction model for permanent magnet synchronous motor state based on time-frequency-space analysis.

[0081] In this step, the fusion includes the following steps:

[0082] Step S51: The local time-domain and spatial information obtained through STConV and the frequency-domain information obtained through SFD are concatenated, and the formula is as follows:

[0083]

[0084] In equation (4), These are the local time-domain and spatial information representations and the frequency-domain information representations obtained at time t, respectively. t It is the fused information obtained at time t.

[0085] Step S52: Input the fusion information obtained in step S51 into the transformation linear layer and match it with the output physical quantity at the current time to obtain the final model architecture.

[0086] In this step, the corresponding model between the fused information and the output physical quantity is shown in formula (5):

[0087]

[0088] In equation (5), This represents the output physical quantity at time t, and W′ and bias are the weight and bias parameters of the linear layer.

[0089] Step S6: Collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor. Divide the historical data into training data and test data, and perform data preprocessing on each to obtain the training set and test set.

[0090] In this step, historical data is preprocessed, including standardization and sliding windowing.

[0091] The formula for standardizing the training data is as follows:

[0092] p' = p min +(pp min ) / (p max -p min (6)

[0093] In formula (6), p min p max Let p and p' represent the minimum and maximum values ​​in sequence p, respectively, where p represents the original sequence and p' represents the standardized sequence. When standardizing the test data, formula (6) is still used, except that p' is... min p max From the training set.

[0094] The sliding window is used to extract input data from standardized data by using a window size w as the basic structural unit length, thereby reshaping the input data into samples of a specified length for model training and testing.

[0095] Preferably, historical data is divided into training data and test data in a 5:5 ratio. Other data allocation ratios can also be used as needed.

[0096] Step S7: The intelligent prediction model of permanent magnet synchronous motor state is trained using the training set. After training, the parameters corresponding to each model are obtained, thus obtaining a mature intelligent prediction model of permanent magnet synchronous motor state. Then, the intelligent prediction model of permanent magnet synchronous motor state is tested using the test set.

[0097] In this step, the specific steps of the testing part include: inputting the input physical quantities from the test set after data preprocessing into the mature permanent magnet synchronous motor state prediction intelligent model, outputting the predicted physical quantities, calculating the mean square error (MSE) between the predicted data and the actual data of the output physical quantities, and evaluating the model.

[0098] The model was evaluated using MSE, and the calculation formula is as follows:

[0099]

[0100] In equation (7), y iThis represents the i-th element in the output physical quantity. Let represent the actual value and predicted value of the i-th output physical quantity at time t, respectively.

[0101] Step S8: Collect the M-element input physical quantity of the permanent magnet synchronous motor at the current moment, input the input physical quantity into the mature permanent magnet synchronous motor state intelligent prediction model, obtain the predicted N-element output physical quantity, and determine the state of the permanent magnet synchronous motor to complete the state prediction of the permanent magnet synchronous motor.

[0102] As can be seen from the above technical solutions, the intelligent prediction method for permanent magnet synchronous motor state based on time-frequency-space analysis described in this invention solves the problem of time lag in motor state prediction by learning global information of motor physical quantities in the time domain. Furthermore, it simultaneously learns the frequency domain information of motor physical quantities, as well as the spatial and local time domain information between them. Learning the frequency domain information captures the high-frequency characteristics and mixed periodicity of the motor physical quantities themselves, improving the accuracy of motor state prediction; learning the spatial information between motor physical quantities dynamically captures the nonlinear coupling relationships between them, making the motor prediction more consistent with reality; learning the local time domain information allows for better learning of short-term state changes in the motor physical quantities, further improving the accuracy of motor state prediction. In summary, by analyzing motor physical quantities from the time, frequency, and space directions, various features in the data can be fully captured, greatly improving the accuracy of motor state prediction.

[0103] Based on the same idea, this invention also provides an intelligent prediction system for the state of a permanent magnet synchronous motor based on time-frequency-space analysis. The system includes: a data acquisition module, a data preprocessing module, a global time-domain information representation module, a local time-domain and spatial information representation module, a frequency-domain information representation module, an intelligent prediction model construction module, and a result output module.

[0104] The data acquisition module is used to set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor, and to collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor, as well as to collect the M-element input physical quantity to be predicted at the current moment.

[0105] The data preprocessing module is used to divide historical data into training data and test data, and to obtain training sets and test sets after performing data preprocessing respectively.

[0106] The global time-domain information representation module is used to construct a global time-domain information representation model based on the input physical quantity using a long short-term memory neural network, and send the obtained global time-domain information of the input physical quantity in the time domain to the local time-domain and spatial information representation module and the frequency-domain information representation module.

[0107] The local temporal and spatial information representation module is used to construct a local temporal and spatial information representation model based on global temporal information using a spatiotemporal convolutional neural network, and output local temporal and spatial information;

[0108] The frequency domain information representation module is used to construct a frequency domain information representation model based on global time domain information using a spectrum analysis network (SFD), and output the frequency domain information of the input physical quantities.

[0109] The intelligent prediction model construction module is used to fuse the local time-domain, spatial, and frequency-domain information, and map the fused information to the output physical quantity to construct an intelligent prediction model for the permanent magnet synchronous motor (PMSM) state. It is also used to train the PMSM state intelligent prediction model using a training set, obtaining the parameters corresponding to each model after training, thus obtaining a mature PMSM state intelligent prediction model. The model is then tested using a test set. Furthermore, it is used to input the collected M-element input physical quantity at the current moment into the mature PMSM state intelligent prediction model to obtain the predicted N-element output physical quantity, and send the prediction result to the result output module.

[0110] The result output module is used to output the predicted N-element output physical quantity to determine the state of the permanent magnet synchronous motor.

[0111] In this embodiment, each module is implemented using a processor, with additional memory added as needed for storage. The processor can be, but is not limited to, a microprocessor (MPU), a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), other programmable logic devices, discrete gates, transistor logic devices, discrete hardware components, etc. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0112] It should also be noted that the intelligent prediction system for the state of a permanent magnet synchronous motor based on time-frequency-space analysis described in this embodiment corresponds to the intelligent prediction method for the state of a permanent magnet synchronous motor based on time-frequency-space analysis. The description and limitations of the method also apply to the system, and will not be repeated here.

[0113] The above description represents preferred embodiments of the present invention. It should be noted that the present invention is not limited to the exemplary embodiments disclosed above. The specification is merely intended to help those skilled in the art to comprehensively understand the specific details of the present invention. For those skilled in the art, any improvements, modifications, easily conceivable variations, or substitutions made within the scope of the technology disclosed in the present invention without departing from the principles described herein should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of the state of a permanent magnet synchronous motor based on time-frequency-space analysis, characterized in that, Includes the following steps: Step S1: Set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor; Step S2: Based on the input physical quantity, a global time-domain information representation model is constructed using a long short-term memory neural network to obtain the global time-domain information of the input physical quantity in the time domain; Step S3: Based on global temporal information, a spatiotemporal convolutional neural network STConV is used to construct a local temporal and spatial information representation model to obtain the local temporal and spatial information of the input physical quantity; Step S4: Based on global time-domain information, a frequency domain information representation model is constructed using a spectrum analysis network (SFD) to obtain the frequency domain information of the input physical quantity; the SFD network includes a Fourier transform (FFT) layer, a spatial linear layer, and an inverse Fourier transform (iFFT) layer. Step S5: The local time-domain and spatial information and frequency-domain information are fused, and the fused information is matched with the output physical quantity to obtain the intelligent prediction model of the permanent magnet synchronous motor state. Step S6: Collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor, divide the historical data into training data and test data, and perform data preprocessing to obtain the training set and test set respectively. Step S7: The intelligent prediction model of permanent magnet synchronous motor state is trained using the training set. After training, the parameters corresponding to each model are obtained, thus obtaining a mature intelligent prediction model of permanent magnet synchronous motor state. Then, the intelligent prediction model of permanent magnet synchronous motor state is tested using the test set. Step S8: Collect the M-element input physical quantity of the permanent magnet synchronous motor at the current moment, and input the input physical quantity into the mature permanent magnet synchronous motor state intelligent prediction model to obtain the predicted N-element output physical quantity, and determine the state of the permanent magnet synchronous motor.

2. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 1, characterized in that, The input physical quantities include four-axis voltages (d, q, a, b), and the output physical quantities include rotational speed, torque, d-axis current, and q-axis current. M=4, N=4.

3. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 1, characterized in that, In step S2, the global time-domain information representation model is as shown in equation (1): (1) In equation (1), It means The input physical quantity at any given time. , They represent respectively , The global time-domain information at each moment, and the global time-domain information at all moments constitute a time series. It is a built-in function of LSTM networks.

4. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 3, characterized in that, In step S3, the STConV network can simultaneously perform convolutions in the temporal and spatial domains, and the constructed local temporal and spatial information representation model is shown in equation (2): (2) In equation (2), express Local temporal and spatial information acquired at any given time. Indicates from Time's up Global time-domain information at any given moment. The window length for the global temporal information time series used for information extraction; k is the number of convolutional kernels in the STConV network. This indicates two newly introduced convolutional kernels. Attention is calculated in the spatial and temporal dimensions.

5. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 4, characterized in that, The frequency domain information representation model is shown in equation (3): (3) In equation (3), Let these represent the Fourier transform function and the inverse Fourier transform function, respectively. These are parameters in a spatially linear layer. This means converting global time-domain information into learnable frequency-domain information using a Fourier transform. This means that the learned frequency domain information is converted into a time domain representation after inverse Fourier transform; Pick In This serves as the learned frequency domain information at time t.

6. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 5, characterized in that, The fusion described in step S5 includes the following steps: Step S51: The local time-domain and spatial information obtained through STConV and the frequency-domain information obtained through SFD are concatenated, using the following formula: (4) In equation (4), , These are the local time-domain information and spatial information representations, and the frequency-domain information representations obtained at time t, respectively. It is the fused information obtained at time t; Step S52: Input the fused information into the transformation linear layer and correlate it with the output physical quantity at the current time to obtain the final model architecture; the correspondence model between the fused information and the output physical quantity is shown in formula (5): (5) In equation (5), This represents the output physical quantity at time t. These are the weight and bias parameters of the linear layer.

7. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 1, characterized in that, In step S6, data preprocessing includes standardization and sliding windowing; among which, The formula for standardizing the training data is as follows: (6) In formula (6), , Representing sequences respectively The minimum and maximum values ​​in Represents the original sequence. This represents the standardized sequence; When standardizing the test data, formula (6) is still used, where... , From the training set; The sliding window is used to adjust the standardized data by window size. The input data is truncated using the length of the basic structural unit, thereby reshaping the input data into a sample of the specified length.

8. The intelligent prediction method for the state of a permanent magnet synchronous motor according to claim 1, characterized in that, In step S7, the test specifically includes: inputting the input physical quantities from the test set after data preprocessing into the mature permanent magnet synchronous motor state prediction intelligent model, outputting the predicted output physical quantities, and calculating the mean square error (MSE) between the predicted data and the actual data of the output physical quantities to evaluate the model. The formula for evaluating the model using MSE is as follows: (7) In equation (7), Indicates the first physical quantity in the output Yuan, , They represent Time of the first The meta-output provides the actual and predicted values ​​of physical quantities.

9. A state intelligent prediction system for permanent magnet synchronous motors based on time-frequency-space analysis, characterized in that, The system includes: a data acquisition module, a data preprocessing module, a global time-domain information representation module, a local time-domain and spatial information representation module, a frequency-domain information representation module, an intelligent prediction model construction module, and a result output module; wherein... The data acquisition module is used to set the M-element input physical quantity and N-element output physical quantity of the permanent magnet synchronous motor, and to collect historical data of the M-element input physical quantity and N-element output physical quantity generated during the current operation of the motor, as well as to collect the M-element input physical quantity to be predicted at the current moment. The data preprocessing module is used to divide historical data into training data and test data, and to obtain training sets and test sets after performing data preprocessing respectively. The global time-domain information representation module is used to construct a global time-domain information representation model based on the input physical quantity using a long short-term memory neural network, and send the obtained global time-domain information of the input physical quantity in the time domain to the local time-domain and spatial information representation module and the frequency-domain information representation module. The local temporal and spatial information representation module is used to construct a local temporal and spatial information representation model based on global temporal information using a spatiotemporal convolutional neural network, and output local temporal and spatial information; The frequency domain information representation module is used to construct a frequency domain information representation model based on global time domain information using a spectrum analysis network (SFD), and output the frequency domain information of the input physical quantity; the SFD network includes a Fourier transform (FFT) layer, a spatial linear layer, and an inverse Fourier transform (iFFT) layer; The intelligent prediction model construction module is used to fuse the local time-domain, spatial, and frequency-domain information, and map the fused information to the output physical quantity to construct an intelligent prediction model for the permanent magnet synchronous motor (PMSM) state. It is also used to train the PMSM state intelligent prediction model using a training set, obtaining the parameters corresponding to each model after training, thus obtaining a mature PMSM state intelligent prediction model. The model is then tested using a test set. Furthermore, it is used to input the collected M-element input physical quantity at the current moment into the mature PMSM state intelligent prediction model to obtain the predicted N-element output physical quantity, and send the prediction result to the result output module. The result output module is used to output the predicted N-element output physical quantity to determine the state of the permanent magnet synchronous motor.

Images