CN117607680A

CN117607680A - State evaluation method of three-phase full-wave brushless motor

Info

Publication number: CN117607680A
Application number: CN202311111138.4A
Authority: CN
Inventors: 莫劲松; 韩灵
Original assignee: Tianjin Nuoxin Xinda Technology Co ltd
Current assignee: Hangzhou Quanjin Technology Co ltd
Priority date: 2023-08-30
Filing date: 2023-08-30
Publication date: 2024-02-27
Anticipated expiration: 2043-08-30

Abstract

The invention relates to the technical field of machine learning, in particular to a state evaluation method of a three-phase full-wave brushless motor. The method comprises the following steps: acquiring three-phase full-wave brushless motor operation data, performing frequency spectrum acoustic sampling and entropy driving reconstruction processing on the three-phase full-wave brushless motor operation data to generate motor remodelling acoustic wave data, acquiring real-time current and voltage data of the three-phase full-wave brushless motor, performing micro-granularity data fusion processing and nonlinear electromagnetic force field construction by combining the motor remodelling acoustic wave data to generate a motor electromagnetic force field diagram, performing micro-variation detection scanning and state causal chain analysis on the motor electromagnetic force field diagram through a deep neural network to generate motor state causal relation data, and performing state prediction evaluation processing according to the motor state causal relation data to generate a motor state prediction report. The invention can automatically learn the mode and the characteristics from the model, thereby realizing accurate evaluation of the motor state.

Description

State evaluation method of three-phase full-wave brushless motor

Technical Field

The invention relates to the technical field of machine learning, in particular to a state evaluation method of a three-phase full-wave brushless motor.

Background

The state evaluation method of applying the machine learning technology to the three-phase full-wave brushless motor is an emerging research direction, and aims to realize accurate evaluation and monitoring of the state of the motor by utilizing a machine learning algorithm. The method can provide real-time monitoring and fault diagnosis of the running state of the motor, and is beneficial to improving the reliability and efficiency of the motor. And the application of the machine learning technology can automatically learn the state characteristics and modes of the motor through a large number of data samples and training algorithms, so that the accurate assessment of the motor state is realized. Obtaining sufficient tagged training data is critical to the success of machine learning. In practice, acquiring a large and accurate amount of brushless motor status data can be challenging, particularly for different types of fault conditions.

Disclosure of Invention

The invention provides a state evaluation method of a three-phase full-wave brushless motor, which aims to solve at least one technical problem.

To achieve the above object, the present invention provides a state evaluation method of a three-phase full-wave brushless motor, the method comprising the steps of:

step S1: acquiring three-phase full-wave brushless motor operation data, and performing spectrum acoustic sampling on the three-phase full-wave brushless motor operation data to generate motor acoustic spectrum data;

Step S2: performing entropy driving reconstruction processing by utilizing motor acoustic spectrum data to generate motor remolded acoustic data;

step S3: acquiring real-time current and voltage data of the three-phase full-wave brushless motor, and performing particle data fusion processing by combining motor remodelling sound wave data to generate motor sound wave particle fusion data;

step S4: constructing a nonlinear electromagnetic force field by utilizing motor acoustic particle fusion data to generate a motor electromagnetic force field diagram;

step S5: performing micro-variation detection scanning on the motor electromagnetic force field diagram through a deep neural network to generate motor micro-variation marking data;

step S6: based on motor micro variation marking data, carrying out state causal link analysis by utilizing a complex network technology to generate motor state causal relation data;

step S7: and carrying out state prediction evaluation processing according to the motor state causal relationship data so as to generate a motor state prediction report.

The invention provides a state evaluation method of a three-phase full-wave brushless motor, which comprises the steps of acquiring three-phase full-wave brushless motor operation data, performing spectrum acoustic sampling on the three-phase full-wave brushless motor operation data to generate motor acoustic spectrum data, and acquiring the three-phase full-wave brushless motor operation data, so that the overall state of motor operation can be captured, and various changes such as vibration and noise in the motor can be captured. This full wave dynamic capture approach helps to gain insight into the operating characteristics of the motor and potential anomalies. And performing entropy driving reconstruction processing by using the motor acoustic spectrum data to generate motor remodelling acoustic spectrum data, wherein the entropy driving reconstruction method is beneficial to capturing hidden features in the acoustic spectrum data. These features can be difficult to capture in conventional frequency domain analysis, but are of great significance for state assessment and fault diagnosis of the motor. By considering the information entropy and distribution characteristics of the signals, potential rules and variation trends in the acoustic data can be revealed. Real-time current and voltage data of the three-phase full-wave brushless motor are obtained, and microparticle data fusion processing is carried out by combining motor remodelling sonic data so as to generate motor sonic microparticle fusion data, and in the data fusion process, a fusion mode of micro granularity is focused, namely, the current, the voltage and the sonic data are fused on a smaller time scale. The micro-granularity fusion strategy enables different types of data to be more densely interwoven together, thereby enhancing the visibility of fault characteristics and providing a more reliable information basis for fault detection and diagnosis. The method comprises the steps of constructing a nonlinear electromagnetic force field by utilizing motor acoustic particle fusion data to generate a motor electromagnetic force field diagram, wherein the distribution of the electromagnetic force field is usually complex nonlinear characteristics in the running process of a motor. The step introduces an innovative nonlinear electromagnetic force field modeling method, considers the diversity characteristics of sound waves and particle data, so as to more accurately describe the electromagnetic force distribution condition inside the motor, carries out micro variation detection scanning on the electromagnetic force field diagram of the motor through a deep neural network so as to generate motor micro variation mark data, and can capture micro variation in the electromagnetic force field diagram of the motor by means of the deep neural network, wherein the variation can be in a range which cannot be perceived by naked eyes, but has obvious influence on the performance and state of the motor. Through micro variation detection, the weak signals can be accurately identified, and a reliable basis is provided for subsequent analysis. Based on motor micro-variation marker data, state causal link analysis is performed by using a complex network technology to generate motor state causal relationship data, wherein the motor micro-variation marker data generally contains information of micro-variation inside a motor, and the micro-variation can play an important role in the state causal relationship. Through complex network technology, dynamic changes between states can be captured, how small changes affect the evolution of the states is revealed, and therefore richer information is provided for state prediction. And carrying out state prediction evaluation processing according to the state causal relationship data of the motor to generate a motor state prediction report, and identifying state characteristics which have important influence on the performance and stability of the motor based on analysis of the state causal relationship data. These features may be tiny, imperceptible, but may lead to failure in future developments. Through the application of the prediction model, the possibility of the key fault characteristics can be better obtained, and beneficial prediction guidance is provided for preventive maintenance and fault elimination.

Drawings

FIG. 1 is a schematic flow chart of the state evaluation method of a three-phase full-wave brushless motor according to the present invention;

FIG. 2 is a detailed implementation step flow diagram of step S1;

fig. 3 is a detailed implementation step flow diagram of step S2.

Detailed Description

It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

The embodiment of the application provides a state evaluation method of a three-phase full-wave brushless motor. The execution subjects of the state evaluation method of the three-phase full-wave brushless motor include, but are not limited to, those on which the system is mounted: mechanical devices, data processing platforms, cloud server nodes, network transmission devices, etc. may be considered general purpose computing nodes of the present application. The data processing platform includes, but is not limited to: at least one of an audio management system, an image management system and an information management system.

Referring to fig. 1 to 3, the present invention provides a state evaluation method of a three-phase full-wave brushless motor, the method comprising the steps of:

In the embodiment of the present invention, as described with reference to fig. 1, a flow chart of steps of a method for evaluating a state of a three-phase full-wave brushless motor according to the present invention is shown, and in this example, the method for evaluating a state of a three-phase full-wave brushless motor includes the following steps:

in the embodiment of the invention, three-phase full-wave brushless motor operation data are acquired, dynamic sensing positioning processing is carried out on the three-phase full-wave brushless motor operation data by utilizing a KNN algorithm to generate a sensor optimal coordinate set, sensor positioning deployment is carried out on the basis of the sensor optimal coordinate set, a sensor is started, edge frequency sampling and filtering are carried out to generate a motor edge frequency data set, time sequence synchronization processing is carried out on the motor edge frequency data set to generate a motor synchronization sound wave data set, and multidimensional frequency spectrum reconstruction is carried out on the basis of the motor synchronization sound wave data set to generate motor sound wave frequency spectrum data.

In the embodiment of the invention, a motor acoustic wave entropy calculation formula is utilized to calculate motor acoustic wave frequency spectrum data so as to generate motor frequency spectrum entropy score data, entropy threshold calibration processing is carried out through the motor frequency spectrum entropy score data so as to generate an entropy threshold data set, entropy driving frequency spectrum screening is carried out on the motor acoustic wave frequency spectrum data through an edge calculation technology so as to generate motor high-entropy frequency spectrum data, and distributed acoustic wave remodeling is carried out on the motor high-entropy frequency spectrum data so as to generate motor remolded acoustic wave data.

in the embodiment of the invention, the motor remodelling sound wave data is subjected to data granularity division to generate a motor sound wave particle data set, real-time current and voltage data of the three-phase full-wave brushless motor are obtained, motor dynamic characteristic extraction is performed on the real-time current and voltage data of the three-phase full-wave brushless motor to generate motor dynamic characteristic data, homomorphic encryption encoding is performed on the motor sound wave particle data set and the motor dynamic characteristic data by utilizing homomorphic encryption technology to generate a post-secret particle data set and post-secret characteristic data, and heterogeneous data fusion is performed on the post-secret particle data set and the post-secret characteristic data to generate motor sound wave particle fusion data.

Step S4: constructing a nonlinear electromagnetic force field by utilizing motor acoustic particle fusion data to generate a motor electromagnetic force field diagram; the method comprises the steps of carrying out a first treatment on the surface of the

In the embodiment of the invention, particle space nonlinear remodeling is carried out on motor acoustic particle fusion data to generate remodeled particle space structure data, an electromagnetic force field source point data is generated by calculating remodeled particle space structure data through an electromagnetic force source point intensity positioning formula, electromagnetic force field intensity modeling is carried out on the basis of the electromagnetic force field source point data to generate a motor electromagnetic intensity model, dynamic force field path prediction is carried out on the basis of a time sequence prediction technology and the motor electromagnetic intensity model to generate motor dynamic electromagnetic path data, and electromagnetic stand space rendering processing is carried out on the basis of the motor dynamic electromagnetic path data to generate a motor electromagnetic force field diagram.

in the embodiment of the invention, electromagnetic complex deformation state interpretation is carried out on a motor electromagnetic force field diagram to generate motor electromagnetic complex characteristic data, fault microcosmic association structure identification is carried out on the motor electromagnetic complex characteristic data based on a generation countermeasure network to generate a motor fault early warning association diagram, mutation behavior deep association analysis is carried out through the motor fault early warning association diagram to generate a microcosmic mutation behavior identification strategy set, and microcosmic mutation self-adaptive marking processing is carried out based on the microcosmic mutation behavior identification strategy set to generate motor microcosmic mutation marking data.

Step S6: and performing deep intelligent optimization based on elastic analysis on the LED grain elastic analysis data to generate an LED grain elastic driving optimization scheme.

In the embodiment of the invention, the elasticity spectrum characteristic analysis is carried out by utilizing the elasticity analysis data of the LED crystal grains to generate the elasticity spectrum characteristic data of the LED crystal grains, the deep learning characteristic intelligent analysis is carried out based on the elasticity spectrum characteristic data of the LED crystal grains to generate the elasticity intelligent characteristic descriptor of the LED crystal grains, the holographic twin modeling is carried out on the LED crystal grains by utilizing the elasticity intelligent characteristic descriptor of the LED crystal grains and combining the holographic twin technology to generate the holographic twin model data of the LED crystal grains, and the elastic driving optimization processing is carried out according to the holographic twin model data of the LED crystal grains to generate the elastic driving optimization scheme of the LED crystal grains.

in the embodiment of the invention, micro-variable data distribution revealing processing is carried out on motor micro-variation marking data to generate a motor micro-variable data distribution diagram, heterogeneous network construction is carried out on the basis of the motor micro-variable data distribution diagram to generate motor state heterogeneous network data, deep association mining is carried out on the motor state heterogeneous network data by utilizing a GCN algorithm to generate motor state deep association data, state causal logic reasoning is carried out on the motor state deep association data by combining neural symbol learning and motor state deep association data to generate a motor state logic causal chain, and state causal chain integration processing is carried out on the basis of the motor state logic causal chain to generate motor state causal relationship data.

In the embodiment of the invention, state transition learning processing is carried out on motor state causal relationship data to generate motor state transition probability data, asynchronous state screening is carried out according to the motor state transition probability data to generate a motor key state transition path, state causal network reconstruction is carried out on the basis of the motor key state transition path to generate a motor state causal network reconstruction diagram, micro-variation state prediction processing is carried out on the motor state causal network reconstruction diagram to generate motor micro-variation state prediction data, semantic report programming is carried out on the basis of the motor micro-variation state prediction data to generate a motor state prediction report.

Preferably, step S1 comprises the following steps;

step S11: acquiring three-phase full-wave brushless motor operation data, and performing dynamic sensing positioning processing on the three-phase full-wave brushless motor operation data by utilizing a KNN algorithm to generate a sensor optimal coordinate set;

step S12: positioning and deploying the sensor based on the optimal coordinate set of the sensor, starting the sensor, and performing edge frequency sampling and filtering to generate a motor edge frequency data set;

Step S13: performing time sequence synchronization processing by using the motor edge frequency data set to generate a motor synchronization sound wave data set;

step S14: and carrying out multidimensional frequency spectrum reconstruction based on the motor synchronous sound wave data set to generate motor sound wave frequency spectrum data.

According to the invention, the operation data of the three-phase full-wave brushless motor is obtained, dynamic sensing and positioning processing is carried out on the operation data of the three-phase full-wave brushless motor by utilizing a KNN algorithm, so that an optimal coordinate set of the sensor is generated, and the adaptive positioning of the sensor is realized by taking the change of the motor state into consideration from a dynamic angle through the innovative application of the KNN algorithm. Therefore, the sensor can be optimally distributed according to the real-time changing requirement in the running process of the motor, so that the state change of the motor is more effectively captured. And (3) carrying out sensor positioning deployment based on the optimal coordinate set of the sensor, starting the sensor and carrying out edge frequency sampling and filtering to generate an edge frequency data set of the motor, wherein the sensor is arranged at a key part of the motor based on the optimal coordinate set of the sensor, and the direct connection between the position of each sensor and the running state of the motor is ensured by the well-designed deployment mode. Therefore, the data collected by different sensors can fully reflect the vibration and signal change in the motor, so that a rich information basis is provided for subsequent analysis. And (3) performing time sequence synchronization processing by using the motor edge frequency data set to generate a motor synchronization sound wave data set, wherein the efficient time sequence synchronization processing is adopted. And (3) establishing a synchronized time axis by selecting data of one reference sensor, and mapping data of other sensors to corresponding time points. The innovative synchronization method ensures that the data of different sensors have corresponding relations on the same time axis. Performing a multi-dimensional spectral reconstruction based on the motor-synchronized acoustic data set to generate motor acoustic spectral data,

As an example of the present invention, referring to fig. 2, a detailed implementation step flow diagram of step S1 in fig. 1 is shown, where step S1 includes:

in the embodiment of the invention, motor operation data including current, voltage, rotation speed and other information are collected from the operation process of the three-phase full-wave brushless motor. These data may be collected by an array of sensors, each measuring one or more parameters. These collected raw data are stored in time series. And adopting a K nearest neighbor (K-Nearest Neighbors, KNN) algorithm to perform dynamic sensing positioning processing on the acquired data so as to generate an optimal coordinate set of the sensor. The method comprises the following specific steps: preprocessing the collected original data, including denoising, outlier processing and the like. This can be achieved by filters and statistical methods to ensure data quality, extracting useful features from the pre-processed data, which may include spectral distribution of current and voltage waveforms, statistical features, etc. The features can be used for a subsequent KNN algorithm to perform standardized processing on the extracted features, and different features are ensured to have similar scales, so that the KNN algorithm can accurately calculate the distance between samples. A KNN algorithm is applied to the normalized features to determine K nearest neighbors for each data point. For each data point, the KNN algorithm will find the K data points most similar to it and determine their locations, and calculate the sensor optimal coordinates for the current data point based on the K nearest neighbor locations. The sensor optimal coordinates may be obtained by averaging or weighting the K nearest neighbor coordinates. The dynamic sensing positioning process is realized through the steps so as to generate the optimal coordinate set of the sensor.

in the embodiment of the invention, the sensor is deployed by utilizing the optimal coordinate set of the sensor, and then the sensor is started and edge frequency sampling and filtering are performed, so that a motor edge frequency data set is generated. The specific implementation details are as follows: the layout and the position of the sensor are determined based on the sensor optimal coordinate set generated in step S11. The sensors should be arranged at key parts of the motor so as to capture vibration and signals in operation, and the sensors are arranged on the motor according to the position information of the optimal coordinate set of the sensors, so that the sensors can accurately acquire data. And starting the sensor equipment to start to acquire motor operation data in real time, wherein the signals comprise current, voltage and the like. The acquired data are stored in a time series form, and edge frequency sampling filtering is carried out on the data acquired by each sensor. The purpose of this step is to capture the dither signals during operation of the motor, which are generally closely related to the operating state and faults of the motor, for which purpose it is first necessary to design a suitable bandpass filter for filtering out the low-frequency and high-frequency components, leaving only signals in a specific frequency range. This particular frequency range may be determined based on a priori knowledge or experimentation, and the data generated by each sensor after filtering by the edge frequency samples forms an edge frequency data sequence. And integrating the edge frequency data sequences generated by all the sensors to form a motor edge frequency data set. The edge frequency sampling filtering is achieved through the steps to generate a motor edge frequency dataset.

in the embodiment of the present invention, the key frequency component is extracted from the motor edge frequency data set obtained in step S12. This can be achieved by applying a fourier transform algorithm, converting the signal in the frequency domain into amplitude and phase information in the time domain. A particularly selected fourier transform method may be a Fast Fourier Transform (FFT). In order to time align the data from the different sensors, a synchronized timeline needs to be created. This can be done by selecting the data of one reference sensor as a reference, then mapping the data of the other sensors to corresponding points in time according to the sampling rate of that sensor, interpolating the data of each sensor according to the synchronized time axis. This is to ensure that the data of the different sensors have corresponding values at the same point in time for the subsequent synchronization process. By time-aligning and interpolating the synchronized data, synchronized data sets on the same time axis are obtained. These datasets now have the same temporal resolution and aligned time points. The motor edge frequency data can be converted into acoustic signals using the synchronized data set. This may be achieved by performing specific filtering, such as bandpass filtering, on the synchronized data to emphasize the acoustic signal. The filter design may use digital filter design techniques, such as Butterworth filter, that filter and process to obtain the motor-synchronized acoustic wave dataset. The time sequence synchronization processing is realized through the steps so as to generate the motor synchronization sound wave data set.

In the embodiment of the present invention, the data preprocessing is first performed in the motor synchronization acoustic data set obtained in step S13. This includes removing possible noise, filtering and normalization processes to ensure that the data remains high quality in subsequent processing, splitting the pre-processed data into short segments of short duration, each segment being referred to as a frame. This process, known as framing, helps capture the short-time varying features of the signal in the frequency domain by decomposing the data into small segments that are continuous in time. A Window function, such as Hanning Window, is applied to each frame. The window function can reduce the frequency spectrum leakage problem, so that the signal is smoother in the frequency domain, the frequency spectrum analysis is facilitated, and short-time Fourier transformation is performed on each frame data weighted by the window function. The STFT is capable of converting a signal from a time domain to a time-frequency domain, converting time domain information of each frame to spectrum information, and when the STFT is applied, a length of a window and a degree of overlap need to be selected. The window length affects the frequency resolution, longer windows may provide better frequency resolution, but may reduce the time resolution. Overlapping can reduce data loss and improve accuracy of spectrum analysis. For each frame, the STFT will generate a complex matrix containing spectral information. The amplitude spectrum of each frame can be calculated, which represents the intensities of the different frequency components. And adding the amplitude spectrums of all frames according to the corresponding time positions to obtain a multi-dimensional spectrum reconstruction result of the whole signal. Through the steps, multi-dimensional spectrum reconstruction is realized to generate motor sound wave spectrum data. The use of a window function can reduce the spectral leakage problem, but the window length affects the balance between frequency resolution and time resolution. Innovatively, according to the characteristics of the motor acoustic wave signals, the length of the selection window is weighted according to specific conditions, so that a better effect is achieved in spectrum analysis. The innovations herein are embodied in combining the selection of window functions with the actual need for motor acoustic signals to provide finer control over data processing.

Preferably, step S2 comprises the steps of:

step S21: step S21: calculating the motor acoustic wave frequency spectrum data by using a motor acoustic wave entropy calculation formula to generate motor frequency spectrum entropy score data;

step S22: performing entropy threshold scaling processing through the motor spectrum entropy score data to generate an entropy threshold data set;

step S23: entropy threshold data sets are utilized, and entropy driving spectrum screening is carried out on motor sound wave spectrum data through an edge computing technology, so that motor high-entropy spectrum data are generated;

step S24: and performing distributed acoustic remodeling on the motor high-entropy frequency spectrum data to generate motor remodelling acoustic data.

According to the invention, entropy threshold calibration processing is carried out through the motor frequency spectrum entropy value data so as to generate an entropy threshold data set, and a proper frequency range is divided according to the structure and the operation characteristics of the motor. The personalized division considers the special properties of the motor, so that the entropy threshold calibration can be more accurately adapted to different motor types and working conditions, the analysis precision of fault characteristics is improved, and the sub-set divided by each frequency range is subjected to statistical calculation. And obtaining the distribution condition of the spectrum entropy scores in different areas by calculating statistics such as the mean value, standard deviation and the like of the spectrum entropy scores. This statistical analysis method increases the multidimensional understanding of the spectral entropy score data, making the determination of the entropy threshold more targeted and reliable. Entropy threshold data sets are utilized, entropy driving spectrum screening is conducted on motor sound wave spectrum data through an edge computing technology, so that motor high-entropy spectrum data are generated, and in order to capture local features in the spectrum data more finely, a small window dividing mode is introduced. The locality judgment of the spectrum data is realized by locality screening, namely, the local entropy of the high-entropy spectrum component is calculated and compared in a small window. The innovative method avoids the excessive processing of the whole frequency spectrum data while retaining the high entropy characteristic, thereby capturing the tiny change of the motor state more accurately. And performing distributed acoustic remodeling on the motor high-entropy frequency spectrum data to generate motor remodelling acoustic data. And carrying out inverse transformation on the sub-signals obtained by each decomposition level, so as to realize the remolding of the frequency spectrum data. The innovative inverse transform recombines the decomposed spectral sub-signals, producing remodeled spectral data. By mapping back to the time domain, an acoustic signal corresponding to each sub-signal is obtained. The wavelet packet decomposition algorithm is adopted to process the high-entropy frequency spectrum data, and multi-scale signal decomposition is creatively introduced in the step. By decomposing the spectral data into sub-signals of different frequency bands and scales, fine changes in the frequency spectrum can be captured more finely, revealing information contained within the fine changes in the motor. This multi-scale decomposition helps to dissect complex spectral data into more interpretable components.

As an example of the present invention, referring to fig. 3, a detailed implementation step flow diagram of step S2 in fig. 1 is shown, where step S2 includes:

step S21: calculating the motor acoustic wave frequency spectrum data by using a motor acoustic wave entropy calculation formula to generate motor frequency spectrum entropy score data;

preferably, the motor acoustic wave entropy calculation formula in step S21 is as follows:

wherein H refers to motor frequency spectrum entropy score data, deltaf refers to frequency interval, f _min Refers to the minimum frequency of the motor acoustic wave signal, f _max Refers to the maximum frequency of the sound wave signal of the motor, f refers to the sound wave frequency spectrum data of the motor, p (f) refers to the probability distribution of the sound wave signal under f, t refers to the time variable, N refers to the number of sound wave characteristics, N refers to the index value of the sound wave characteristics, a _n Refers to the amplitude adjustment coefficient, b _n Refers to the frequency offset coefficient, c refers to the phase offset coefficient, d _n Refers to the attenuation coefficient. A step of

The invention creates a motor acoustic wave entropy calculation formula, wherein p (f) is a basic element in motor acoustic wave frequency spectrum data, and reflects acoustic wave intensity or occurrence probability under certain frequency. Its product with natural logarithm, i.e. -p (f) log ₂ (p (f)) used in the information theory to calculate entropy, it can quantify the uncertainty of the signal. This portion represents the rate of change in frequency of the acoustic wave signal over time. The time dimension and the frequency dimension are combined, so that the formula not only considers static sound wave characteristics, but also can capture dynamic change information. Since the formula considers a plurality of characteristics and parameters of the acoustic wave signal, the sensitivity is very high, and micro faults which can be ignored by the traditional method can be detected, and the acoustic wave signal of the motor is detectedThe number is often affected by a variety of acoustic characteristics, such as frequency, amplitude, phase, etc. By introducing parameters such as amplitude adjustment coefficients, frequency offset coefficients, phase offset coefficients and the like, the formula can capture the multidimensional characteristics of the acoustic wave signals. This allows the spectral entropy score data to reflect not only the spectral distribution of the acoustic signal, but also the interrelation between the different characteristics, thus providing more comprehensive motor state information. The mathematical operation part of the formula contains nonlinear functions such as logarithmic functions and sinusoidal functions. These nonlinear operations can capture nonlinear features in the acoustic wave signal, which are often associated with complex changes in motor state and potential faults. By applying these nonlinear operations, the spectral entropy score data can more accurately reflect the diversity of the internal states of the motor.

in the embodiment of the present invention, the frequency range division is first determined from the motor spectrum entropy score data obtained in step S21. The division is based on the characteristics of the acoustic signals and can be determined according to the motor structure, the operating characteristics and the like. The spectral entropy score data is divided into subsets according to this frequency range. For each subset of spectral entropy score data, a statistical calculation is performed. And calculating statistics such as a mean value, a standard deviation and the like of the spectrum entropy scores in each subset to obtain the overall distribution condition of the spectrum entropy scores. Based on the results of the statistical analysis, an innovative method is employed to generate the entropy threshold. The threshold may be the mean plus a standard deviation of a certain multiple, or may be other thresholds designed according to the data distribution situation. This threshold will serve as a basis for determining the complexity of the spectral data. The generated entropy threshold is applied to the spectral entropy score data for all subsets. For each subset, its spectral entropy score is compared to a corresponding entropy threshold. If the spectral entropy score exceeds the entropy threshold, the subset is considered to represent a high entropy region, and the high entropy region labels for each subset are integrated into an entropy threshold dataset. This dataset records the location and distribution of the high entropy regions in different frequency ranges. The entropy threshold scaling process is implemented by the above steps to generate an entropy threshold dataset.

in the embodiment of the present invention, for each frequency band, the entropy threshold value obtained in step S22 is compared with the entropy score of the corresponding frequency band. For bands with entropy scores above the entropy threshold, they are considered to have a higher complexity and randomness and are therefore selected as candidates for high entropy spectral components. To capture local features in the spectral data more finely, the spectral data is partitioned into small windows. The windows are locally adjacent in frequency, and a specific division mode can adopt a beam division (Beamforming) or sliding window-based method, and local entropy of the selected high-entropy frequency spectrum component is calculated in each small window by using a shannon entropy calculation method. This local entropy measures the complexity of the spectral data within the window, i.e. the disorder and randomness of the data. For each portlet, a filtering is performed based on the value of the local entropy. If the local entropy exceeds a preset threshold, the spectral data in the window are marked as high entropy data. The screening method realizes the locality judgment of the spectrum data, more pertinently reserves the high entropy characteristic, and reconstructs the spectrum data marked as high entropy. For each high entropy spectral component, the complete spectral data can be reconstructed in the time and frequency dimensions by interpolation or fitting, etc. The reconstructed spectrum data are the high-entropy spectrum data of the motor. Through the steps, entropy driving spectrum screening is carried out on the motor sound wave spectrum data so as to generate motor high-entropy spectrum data.

In the embodiment of the present invention, first, a wavelet packet decomposition algorithm is applied to each reserved high-entropy spectrum data band. The wavelet packet decomposition is a multi-scale signal decomposition method, which can decompose signals into sub-signals with different frequency bands and different scales, select wavelet basis functions suitable for the characteristics of acoustic signals, such as Daubechies wavelet, and select a proper decomposition layer number according to the distribution condition of frequency spectrum data. The decomposition of each layer decomposes the signal into sub-signals of different frequency ranges. And reconstructing the sub-signals obtained by each decomposition layer. This reconstruction process involves an inverse transform of the wavelet packet that reassembles the individual sub-signals into reconstructed spectral data in which each sub-signal represents a spectral component of a particular frequency range. In order to generate motor remodelling sound wave data, each sub-signal is mapped back to the time domain to obtain a corresponding sound wave signal. All remodeled acoustic signals are time domain synthesized and added according to the time corresponding positions. Thus, final motor remodelling acoustic data can be obtained.

Preferably, step S3 comprises the steps of:

step S31: performing data granularity division on the motor remodelling sound wave data to generate a motor sound wave particle data set;

step S32: acquiring real-time current and voltage data of the three-phase full-wave brushless motor, and extracting motor dynamic characteristics from the real-time current and voltage data of the three-phase full-wave brushless motor to generate motor dynamic characteristic data;

step S33: homomorphic encryption technology is utilized to carry out homomorphic encryption coding on the motor acoustic particle data set and the motor dynamic characteristic data so as to generate a post-secret particle data set and post-secret characteristic data;

step S34: and carrying out heterogeneous data fusion on the dense particle data set and the dense characteristic data to generate motor sound wave particle fusion data.

According to the method, the motor remodelling sound wave data are subjected to data granularity division to generate the motor sound wave particle data set, and the interested frequency sub-band can be selected according to the time-frequency domain analysis result. By further dividing the data, the data within each frequency sub-band is sliced into smaller acoustic particles. This fine division allows finer details of small vibrations inside the motor to be captured, providing a stronger support for early prediction of motor failure. Real-time current and voltage data of the three-phase full-wave brushless motor are obtained, motor dynamic characteristic extraction is carried out on the real-time current and voltage data of the three-phase full-wave brushless motor, so that motor dynamic characteristic data are generated, and the running state of the motor can be deeply informed through obtaining the real-time current and voltage data. The changes in current and voltage contain rich operational information such as vibration of the motor, load changes, etc. The information has important significance in fault diagnosis, performance evaluation and the like. Through application of domain knowledge and data analysis technology, dynamic characteristics suitable for motor state monitoring are selected. These features may cover various aspects of motor operation, such as vibration, imbalance, bearing wear, etc. The innovative choice of these features is critical, since they must be able to accurately reflect the actual functioning of the motor. And the homomorphic encryption technology is utilized to carry out homomorphic encryption coding on the motor acoustic particle data set and the motor dynamic characteristic data so as to generate a post-secret particle data set and post-secret characteristic data, and the Paillier homomorphic encryption algorithm is adopted to realize the encryption coding on the motor acoustic particle data set and the dynamic characteristic data, so that the data privacy is effectively protected. This encryption scheme allows the data to be calculated in an encrypted state without exposing the actual value, thereby ensuring confidentiality of the data. Heterogeneous data fusion is carried out on the dense particle dataset and the dense characteristic data to generate motor sound wave particle fusion data, and the depth self-encoder is introduced to carry out unsupervised characteristic learning and data reconstruction, so that the method is a core innovation of the step. By the joint action of the encoder and the decoder, the depth self-encoder can learn the low-dimensional representation of the most dynamic feature from the post-dense feature data, providing more representative and meaningful features for the subsequent fusion step.

in the embodiment of the present invention, first, data preprocessing is performed from the motor synchronization acoustic data set obtained in step S13. This includes removing noise, filtering and normalizing to reduce interference and enhance the useful signal, slicing the pre-processed acoustic data into short-period frames. Each frame represents an acoustic signal within a small time segment. A windowing function, such as a hamming window, is applied to each frame to reduce spectral leakage problems, and a Fast Fourier Transform (FFT) is performed on the windowed data for each frame. This will transform each frame from the time domain to the frequency domain, producing spectral information, and perform a time-frequency domain analysis on the spectral information of each frame. This involves decomposing the spectrum further into smaller time segments to obtain the frequency variation over time, and selecting the frequency sub-bands of interest based on time-frequency domain analysis. These sub-bands may correspond to the vibration frequency ranges of particular components of the motor, with the data within each frequency sub-band being further divided into smaller data granularity, i.e., acoustic particles. This may be accomplished by selecting a smaller frequency range in the spectral data, or further slicing the data in the time domain, summing the acoustic particle data in all frequency sub-bands to form a motor acoustic particle dataset. The data granularity division is realized through the steps so as to generate the motor acoustic particle data set.

in the embodiment of the invention, real-time current and voltage data of the three-phase full-wave brushless motor are acquired through devices such as a sensor and the like. These data record the current and voltage changes during motor operation. Dynamic characteristics of motor state monitoring are selected. These features should be able to reflect the operation of the motor, such as vibrations, unbalance, bearing wear, etc. The innovative selection of features can be achieved by combining domain knowledge and data analysis, where there are often various types of noise and interference in the actual motor data. By applying digital filtering techniques, such as low pass filtering or band pass filtering, unnecessary high frequency noise can be removed, preserving critical frequency information. In actual motor data, there are often various types of noise and interference. Unnecessary high frequency noise can be removed by applying digital filtering techniques, such as low pass filtering or band pass filtering, key frequency information is retained, and time domain and frequency domain features are calculated on the basis of instantaneous parameters. The time domain features may include mean, standard deviation, etc. The frequency domain features may then be used to calculate spectral energy distribution, frequency peaks, etc. by applying fourier transforms or wavelet transforms. And judging the running state of the motor according to the extracted dynamic characteristics. This may require prior training of a decision model, such as a Support Vector Machine (SVM) or a deep learning model. The model is capable of mapping dynamic characteristics to categories of motor operating conditions. The dynamic characteristic extraction of the motor can be realized through the steps so as to generate dynamic characteristic data of the motor.

in the embodiment of the invention, the Paillier homomorphic encryption algorithm is selected as a specific encryption scheme. The Paillier algorithm is a public key encryption algorithm based on the assumption of large integer decomposition difficulty, supports homomorphic addition and homomorphic multiplication operations, is suitable for encrypting data, and allows calculation in a ciphertext state. First, the public and private keys required by the Paillier algorithm are generated. The public key is used to encrypt data and the private key is used to decrypt data. This step is related to the initialization of the encryption scheme, and is not directly related to the data, and each acoustic particle in the acoustic particle dataset of the motor is encrypted one by one. The actual value of the acoustic wave particles is mapped to an integer form, for example using a fixed point number representation. This is to ensure that the addition and multiplication operations can be performed within the encryption domain, using the Paillier public key, to encrypt each acoustic particle in integer form. The encrypted ciphertext represents the value of the acoustic wave particles within the encryption domain. Each feature in the motor dynamic feature data set is encrypted one by one, the actual value of the dynamic feature is mapped into an integer form, the fixed point number representation is adopted, the same Paillier public key is used for encrypting the dynamic feature data in each integer form, a corresponding ciphertext is generated, and in the encryption process, the ciphertext forms of sound wave particles and the dynamic feature data are obtained. The homomorphic encryption coding is realized through the steps so as to generate a fine particle data set after encryption and characteristic data after encryption.

Step S34: heterogeneous data fusion is carried out on the dense particle data set and the dense characteristic data so as to generate motor sound wave particle fusion data;

in the embodiment of the invention, firstly, homomorphic decryption operation is carried out on the fine particle data set after encryption and the characteristic data after encryption, and the fine particle data set after encryption and the characteristic data after encryption are converted from a ciphertext form to original plaintext data. Here, the decryption is performed using the Paillier homomorphic encryption algorithm to ensure the privacy of the data. The decrypted microparticle data set and the feature data are subjected to proper preprocessing, such as normalization and data type conversion, so that the microparticle data set and the feature data are suitable for being input by a depth self-encoder, and the depth self-encoder is introduced, so that the neural network capable of performing unsupervised feature learning and data reconstruction is provided. The network is composed of an encoder and a decoder. In the encoder, gradually reducing the number of neurons, mapping the raw data to a low-dimensional representation; in the decoder, the number of neurons is gradually increased, mapping the low-dimensional representation back to the original data space. The encoder portion of the depth self-encoder is trained using the post-dense feature data as training data. This will result in the encoder learning a low-dimensional representation that best represents the dynamic features, inputting the decrypted microparticle dataset into the trained depth self-encoder, and mapping the low-dimensional representation back to the original space of the microparticle data by the decoder. This step enables efficient reconstruction of the particulate data. The trained encoder is applied to the decrypted particulate data to obtain a low dimensional representation thereof. And splicing the low-dimensional features with the particle data reconstruction result obtained from the decoder to form motor sound wave particle fusion data. The heterogeneous data fusion can be realized through the steps, so that motor acoustic particle fusion data can be generated.

Preferably, step S4 comprises the steps of:

step S41: performing particle space nonlinear remodeling on motor acoustic particle fusion data to generate remolded particle space structure data;

step S42: calculating the remolded particle space structure data by utilizing an electromagnetic force source point intensity positioning formula so as to generate electromagnetic force field source point data;

step S43: modeling the electromagnetic field intensity based on the electromagnetic field source point data to generate a motor electromagnetic intensity model;

step S44: carrying out dynamic force field path prediction by combining a time sequence prediction technology and a motor electromagnetic intensity model so as to generate motor dynamic electromagnetic path data;

step S45: and performing electromagnetic standing space rendering processing based on the dynamic electromagnetic path data of the motor to generate an electromagnetic force field diagram of the motor.

According to the invention, the particle space nonlinear remodeling is carried out on the motor acoustic particle fusion data to generate the remolded particle space structure data, and the logarithmic mapping is carried out on the acoustic particle fusion data to substantially readjust the distribution of the data. This operation can make the data more sensitive in low amplitude regions while providing a greater dynamic range in high amplitude regions. Therefore, the remodeled data can more prominently show small changes, and the identifiability of the data characteristics is further enhanced. Electromagnetic field strength modeling is performed based on electromagnetic field source point data to generate an electromagnetic strength model of the motor, the electromagnetic field distribution of the motor generally presents complex nonlinear characteristics, and the traditional linear model is often difficult to accurately describe. By selecting the radial basis function neural network, the nonlinear distribution of the electromagnetic force field can be better adapted, so that the fitting capacity and precision of the model are improved, and the modeling of the electromagnetic force field strength involves data interpolation and function approximation in a multidimensional space. The radial basis function neural network is excellent in multidimensional data modeling, can effectively model source point data with different dimensions, and captures the relation between different coordinate points, so that the distribution condition of an electromagnetic force field can be reflected more accurately. And carrying out dynamic force field path prediction by combining a time sequence prediction technology and a motor electromagnetic intensity model to generate motor dynamic electromagnetic path data, integrating the data generated by the motor electromagnetic intensity model and historical electromagnetic field data into time sequence data, and ensuring that the data are orderly arranged according to a time sequence. The method is not only beneficial to the training of a follow-up prediction model, but also can fully reflect the time sequence evolution of the electromagnetic field, so that the prediction result is more time-efficient and reliable. And carrying out electromagnetic force field space rendering processing based on motor dynamic electromagnetic path data to generate a motor electromagnetic force field map, and seamlessly fusing the space-time characteristics of the electromagnetic force field by discretizing continuous time sequence data so as to correlate time steps with specific positions of the motor. The integrated treatment can highlight the time sequence evolution of the electromagnetic force field of the motor, and the color enhancement is visually presented.

in the embodiment of the invention, the logarithmic mapping is selectively applied. The mapping is a common nonlinear transformation mode, and can map the amplitude range of the data to a larger value range, so that the resolution of the data features is enhanced. Regularized processing is carried out on the motor acoustic particle fusion data, so that the value range of the motor acoustic particle fusion data is between 0 and 1. This step is intended to ensure that the data is in the proper range for subsequent nonlinear mapping, applying a logarithmic mapping to the regularized data. The specific implementation steps are as follows: for each regularized particle data x, a log-mapped value y=log (1+x) is calculated, and the amplitude range of the particle data is mapped to a larger value range through log mapping, so that the expression capability of the particle data in a particle space is enhanced. In the logarithmic mapping process, the choice of parameters has an impact on the effect of non-linear remodeling. Parameters in the log-map are adjusted by experimentation and data analysis to obtain the optimal nonlinear remodeling effect. The above steps can realize the nonlinear remodeling of the particle space so as to generate the remodeled particle space structure data.

preferably, the electromagnetic force source point intensity positioning formula in step S42 is as follows:

wherein L refers to electromagnetic force field source point data, alpha refers to standard deviation variable of Gaussian mixture model, M refers to total data point number in RPS, RPS _(j) Refers to the jth data point in the remodeled particle space structure data, μ refers to the mean of the gaussian mixture model, RPS refers to the remodeled particle space structure data, σ refers to the desired distribution value, and γ (α) refers to a nonlinear function that accounts for external disturbances.

The invention creates an electromagnetic force source point intensity positioning formula, RPS _(j) Represents the jth data point in the remodeled microparticle spatial structure data. By combining the gaussian mixture model, analysis of the data points can be performed from the overall distribution of the RPS. By calculating the source point data of the electromagnetic force field, a quantified index can be obtained to describe the source and distribution characteristics of the electromagnetic field. This provides us with a more intuitive and accurate way to understand the behaviour of the electromagnetic field. The core of the formula is to perform Gaussian mixture model calculation on the space structure data of the particles after the reconstruction. By considering the weights and distribution of the different data points, the formula can locate the different source points in the particle space and calculate the corresponding electromagnetic field strengths. The positioning method can improve the positioning precision of the particles in space, so that the source point data of the electromagnetic force field can reflect the distribution of the particles more accurately. The nonlinear function gamma (alpha) is introduced into the formula, and the introduction of the nonlinear characteristic enables the source point data of the electromagnetic force field to better capture nonlinear changes in the spatial structure of the particles. The nonlinear feature extraction can make the electromagnetic force field source point data more expressive and more suitable for the diversity of particle space structures. Alpha is a variable and can be adaptively adjusted according to the particle space structure data in the process of continuously approaching 0, so that the optimal result can be obtained under various conditions.

in an embodiment of the invention, a radial basis function neural network (Radial Basis Function Neural Network, RBFNN) is selected for modeling the electromagnetic field strength distribution of the motor. RBFNN is a neural network suitable for nonlinear pattern modeling that is very efficient for sample interpolation and function approximation in multidimensional space. Preprocessing the source point data of the electromagnetic force field, including data normalization and feature extraction. This may make the data more suitable for modeling before input to the neural network while avoiding unnecessary interference, training on the selected RBFNN model to build the electromagnetic strength model of the motor. During training, the space coordinates of the source points are input, and the corresponding electromagnetic force field intensity is output. Some key parameters including the number, center, width, etc. of radial basis functions need to be set before training RBFNN. The setting of these parameters directly affects the performance and accuracy of the model. The model weight and parameters are adjusted by minimizing the error function of the model by using an optimization algorithm such as a gradient descent method so as to achieve the aim of enabling the model to fit source point data as accurately as possible, and after training is completed, the model is evaluated, and the generalization capability and stability of the model can be checked by using methods such as cross verification and the like. The modeling of the electromagnetic force field strength is realized through the steps so as to generate a motor electromagnetic strength model.

in the embodiment of the invention, the data generated by the motor electromagnetic intensity model and the historical electromagnetic field data are integrated into time series data. Ensure that the data is arranged in chronological order and that any outliers or missing data is handled. To predict future electromagnetic field paths, a suitable time series prediction model is selected. Long and short term memory networks (LSTM), a neural network model that is capable of capturing long-term dependencies in a time series, are employed. The time series data is divided into a training set and a test set. Typically, data over a period of time may be used as a training set, and then tested using subsequent data. An LSTM model is constructed having an input layer, a plurality of LSTM layers, and an output layer. In order for the LSTM to understand the historical dependencies of the data, an appropriate time step (time window) needs to be set, and the LSTM model is trained using a training set. During training, the model minimizes the error between the predicted value and the actual value by adjusting weights and biases. Here, the number of iterations and learning rate of training need to be specified, and the test set is used to verify the predictive effect of the trained LSTM model. By comparing the predicted value of the model with the actual value, the accuracy and generalization ability of the model can be assessed, and after the model is verified, the LSTM model predicts the future electromagnetic field path using known historical data as input. The model generates a predicted path over a period of time in the future based on historical trends and dependencies. And converting the path data predicted by the LSTM model into actual electromagnetic field path data. The dynamic force field path prediction is realized through the steps so as to generate motor dynamic electromagnetic path data.

Step S45: and performing electromagnetic force field space rendering processing based on the motor dynamic electromagnetic path data to generate a motor electromagnetic force field diagram.

In the embodiment of the present invention, key information is extracted from the motor dynamic electromagnetic path data obtained in step S44. Such information may include time series data and corresponding electromagnetic force field strength values. These data are used as input to a rendering process, which data-formats the electromagnetic force field path data, discretizes the continuous time series into a series of time steps. Then, a three-dimensional grid is generated, wherein each grid point corresponds to a specific position of the motor and is associated with a corresponding time step and electromagnetic force field strength value. An innovative Rendering algorithm is employed, such as Volume Rendering (Isosurface Rendering) or isosurface Rendering. The algorithms can extract the distribution and variation of the electromagnetic force field from the grid data to form a visual result. And applying a rendering algorithm to the tidied data grid. Specifically, the volume rendering may convert electromagnetic force field intensity values into color and transparency information by a ray tracing method, and generate a realistic three-dimensional electromagnetic force field map. The isosurface rendering may extract a series of isosurfaces from the electromagnetic force field data for displaying the intensity distribution of the electromagnetic force field. The electromagnetic force field space rendering processing is realized through the steps so as to generate the electromagnetic force field diagram of the motor.

Preferably, step S5 comprises the steps of:

step S51: performing electromagnetic complex deformation interpretation on the electromagnetic field diagram of the motor to generate electromagnetic complex characteristic data of the motor;

step S52: based on the generated countermeasure network, carrying out fault microcosmic correlation structure identification on the electromagnetic complex characteristic data of the motor so as to generate a motor fault early warning correlation diagram;

step S53: carrying out variant behavior deep association analysis through a motor fault early warning association graph to generate a microscopic variant behavior recognition strategy set;

step S54: and performing micro-variation self-adaptive marking processing based on the micro-variation behavior recognition strategy set to generate motor micro-variation marking data.

According to the invention, electromagnetic complex deformation interpretation is carried out on the electromagnetic force field diagram of the motor to generate electromagnetic complex transformation characteristic data of the motor, the electromagnetic force field diagram data of the motor is regarded as a function on a complex plane, complex transformation is introduced, and the data are mapped. The processing mode is to better capture the phase information of the data, and the phase has important significance in the vibration and operation processes of the motor. In this way, more abundant phase features can be extracted from the data, providing more dimensions for subsequent feature analysis. Based on the generation of the countermeasure network, fault microcosmic correlation structure identification is carried out on the electromagnetic complex characteristic data of the motor so as to generate a motor fault early warning correlation diagram, and the motor fault is often accompanied with tiny and complex electromagnetic changes. The generators in the generation countermeasure network have strong representation capability and can capture the weak but important characteristics, so that the microstructure of motor faults is reflected better. This capability is difficult to achieve with conventional methods. And carrying out variant behavior deep association analysis through the motor fault early-warning association diagram so as to generate a microscopic variant behavior recognition strategy set, and capturing association relations among various microscopic fault early-warning features through the motor fault early-warning association diagram. And introducing a frequency spectrum clustering algorithm to divide the fault early warning characteristics into different clusters, wherein each cluster represents a similar variation behavior. Therefore, the microscopic variation behaviors of the motor can be subjected to deep correlation analysis from multiple dimensions, and a more accurate identification strategy is provided for various fault modes. And carrying out micro-variation self-adaptive marking processing based on the micro-variation behavior recognition strategy set to generate motor micro-variation marking data, and carrying out accurate analysis on each node in the motor fault early-warning association diagram by a method based on the micro-variation behavior recognition strategy set. Each node represents a specific microscopic associative feature that typically exhibits minor but critical changes in motor fault conditions. By accurate positioning of these features, deeper fault feature capture is achieved.

in the embodiment of the invention, the electromagnetic force field diagram data of the motor are regarded as functions on complex planes. Complex transform is introduced to map data on the real number domain to the complex number domain, so that phase information can be better processed. Specifically, for each data point, it is represented in complex form, i.e. real and imaginary parts correspond to data values and times, respectively. And applying the Rockwell series expansion to the complex transformed data. This is a method of expanding complex functions to an infinite number of steps on a complex plane. By taking the series within a certain convergence range, the distribution of the original data on the complex plane can be approximated. The aim of the method is to capture the complex form of the electromagnetic force field of the motor, including the characteristics of harmonic wave, oscillation and the like, and to carry out spectrum analysis on the data after the Rockwell series expansion. The data is converted from the time domain to the frequency domain using discrete fourier transform (Discrete Fourier Transform, DFT) or the like. In the frequency domain, frequency characteristics of the electromagnetic force field of the motor, such as fundamental frequency and harmonic components, can be identified. In the frequency domain, according to the distribution and amplitude of harmonic components, the electromagnetic complex characteristic data of the motor are extracted. The electromagnetic complex deformation interpretation can be realized through the steps so as to generate the electromagnetic complex characteristic data of the motor.

in the embodiment of the present invention, the electromagnetic complex characteristic data of the motor generated in step S51 is used as input. These data may represent weak features of vibrations, frequency distribution, phase changes, etc. inside the motor. These feature data are the basis for identifying fault associations. A generator is designed to generate realistic fault-related characteristic data from random noise. The input to the generator is a random noise vector, typically following a gaussian distribution, and the network structure of the generator typically employs a deep neural network comprising a plurality of hidden layers. The details of the neuron number, the activation function and the like of each hidden layer are required to be selected according to the data characteristics and tasks, random noise is taken as input by a generator, multiple transformation and mapping are carried out in a network, and the output of the microcosmic association characteristic data similar to motor faults is generated step by step. The key is the training generator, which enables it to generate realistic, features that are similar to the actual data. The input of the discriminator comprises real motor fault microcosmic associated characteristic data and data generated by the generator, and the discriminator evaluates the input data to judge whether the input data is real data or generated data. Through training, the discriminators gradually improve the recognition capability of the real data. During the training process, the generator and the arbiter perform resistance learning. The goal of the generator is to generate realistic data that can spoof the discriminant, while the goal of the discriminant is to accurately distinguish between the realistic data and the generated data. The process is carried out iteratively, the parameters of the generator and the discriminator are repeatedly optimized, so that the data generated by the generator gradually approaches to real data distribution, and after repeated iterative training, the generator can generate vivid fault associated characteristic data. Through the steps, the fault microcosmic correlation structure identification is realized, so that a motor fault early warning correlation diagram is generated.

Step S53: and carrying out variant behavior deep correlation analysis through the motor fault early warning correlation diagram so as to generate a microscopic variant behavior recognition strategy set.

In the embodiment of the present invention, the motor fault early warning correlation map generated in step S52 is used. The association graph is a graph composed of nodes and edges, wherein the nodes represent micro-fault early-warning features, and the edges represent association relations among the features. And introducing an innovative frequency spectrum clustering algorithm on the motor fault early warning correlation diagram. The algorithm combines graph theory and cluster analysis, can divide the nodes into different clusters, and each cluster represents a similar microscopic variation behavior. First, a graph Laplacian matrix of a motor fault early warning association graph is calculated. The graph Laplace matrix reflects the connection relation among the nodes, can retain the similarity information among the nodes in the cluster analysis, and performs feature vector decomposition on the graph Laplace matrix to obtain feature values and corresponding feature vectors. The feature vectors represent the embedding space of the graph and can reveal the topology between nodes. And projecting the feature vectors into a low-dimensional space, and clustering the projected feature vectors by using a K-means clustering algorithm. K represents the number of clusters, which need to be determined according to practical situations, and each cluster represents a similar microscopic variation behavior. And regarding the nodes in each cluster as similar variant behaviors, and generating a corresponding microscopic variant behavior identification strategy. Through the steps, variant behavior deep correlation analysis is realized, so that a microscopic variant behavior recognition strategy set is generated.

Step S54: performing micro-variation self-adaptive marking processing based on the micro-variation behavior recognition strategy set to generate motor micro-variation marking data;

in the embodiment of the invention, based on a microscopic variation behavior identification strategy set, the strategies are applied to a motor fault early warning association diagram. And determining whether to add a mark on the node according to the microcosmic association characteristic represented by the node and the strategy in the microcosmic mutation behavior identification strategy set for the node in each fault early warning association graph. If the characteristics of the node are matched with a certain strategy, a corresponding micro variation mark is added for the node. Based on the microscopic variation behavior identification strategy set, the strategies are applied to the motor fault early warning correlation diagram. And determining whether to add a mark on the node according to the microcosmic association characteristic represented by the node and the strategy in the microcosmic mutation behavior identification strategy set for the node in each fault early warning association graph. If the characteristics of the node are matched with a certain strategy, a corresponding micro variation mark is added for the node. According to the micro variation marking data generated in the adaptive marking process, the micro variation adaptive marking process can be realized through the steps so as to generate motor micro variation marking data.

Preferably, step S6 comprises the steps of:

step S61: performing micro-variable data distribution revealing processing on the motor micro-variation marking data to generate a motor micro-variable data distribution map;

step S62: heterogeneous network construction is carried out based on the motor micro-data distribution diagram so as to generate motor state heterogeneous network data;

step S63: performing depth association mining on the motor state heterogeneous network data by using a GCN algorithm to generate motor state depth association data;

step S64: performing state causal logic reasoning by combining neural symbol learning and motor state depth association data to generate a motor state logic causal chain;

step S65: and carrying out state causal chain integration processing based on the motor state logic causal chain to generate motor state causal relationship data.

According to the invention, the motor micro-variation marking data is subjected to micro-variable data distribution revealing processing to generate a motor micro-variable data distribution map, and the probability density distribution of the micro-variation data in each dimension is mapped onto a two-dimensional plane to generate the motor micro-variable data distribution map. This graph not only shows the data distribution in each dimension, but also reveals the correlation between the different dimensions. The visualization of this multidimensional data is innovative in that it presents the high-dimensional nature of the data in an intuitive and easily understood image. Heterogeneous network construction is performed based on the motor micro-data distribution diagram to generate motor state heterogeneous network data, and for each node, not only the dimension is taken as the attribute thereof, but also the feature is extracted from the distribution as the attribute of the node, such as a parameter using probability distribution. The method combines the property of probability distribution with the node attribute innovatively, so that the node attribute is richer and diversified, more information of data distribution can be captured, and the association degree between the nodes is explicitly represented by the probability value in the heterogeneous network. An association degree calculating method based on information entropy is innovatively introduced, and the association degree between nodes is quantized by calculating the cross entropy of probability distribution. This not only highlights the relevance between nodes, but also provides more accurate relevance information for subsequent analysis. And performing depth association mining on the motor state heterogeneous network data by using a GCN algorithm to generate motor state depth association data, and iterating a multi-layer graph convolution layer to realize layer-by-layer information transfer and association mining of motor states. The graph rolling operation of each layer combines the features of the node with the features of its neighbors to form a more expressive representation of the features. With the increase of the layer number, the characteristics of the nodes are fused with the information of the more distant neighbors, and the iterative property enables the deep association mining of the motor state to be realized. State causal logic reasoning is performed in combination with neural symbol learning and motor state depth related data to generate a motor state logical causal chain, and complex data relationships can be converted into interpretable symbol logic forms by mapping the motor state depth related data to a neural symbol representation space. This symbolization process makes the inherent links between the data clearer, providing a solid basis for subsequent causal reasoning. And (3) carrying out state causal chain integration processing based on the motor state logic causal chain to generate motor state causal relationship data, and adopting a complex association analysis algorithm such as an Apriori algorithm to comprehensively and systematically analyze the association between different causal chains. Such innovative approaches enable the revealing of common relationship elements between those potential, different causal links, whether or not they are directly apparent, thus enabling a more comprehensive understanding of the causal relationships of motor states.

in the embodiment of the invention, firstly, the motor micro variation mark data is subjected to data preprocessing. This includes removing possible noise and outliers, and normalizing the data to ensure reliability and consistency of the data, and after data preprocessing, the distribution of the micro-variant data in each dimension needs to be calculated. Here, a kernel density estimation method, in particular a gaussian kernel density estimation algorithm, is used. Gaussian kernel density estimation uses a gaussian distribution function to estimate the probability density distribution of the micro-variation data in each dimension. For each dimension, the appropriate bandwidth parameters are selected to ensure the smoothness and accuracy of the estimation. And mapping probability density distribution in each dimension onto a two-dimensional plane by using the result of Gaussian kernel density estimation to generate a motor micro-data distribution map, and firstly, generating grid points with certain density on the two-dimensional plane as an abscissa. Each grid point represents a distribution of potential micro-variable data, and for each grid point, the probability density of the micro-variable data at that point is calculated using the result of the gaussian kernel density estimation. This can be obtained by interpolating probability densities of nearby micro-variable data, and assigning a corresponding color to each point on the motor micro-variable data distribution map according to the micro-variable data probability density of each grid point. The shades represent the distribution density of the micro-variation data. The micro-variable data distribution revealing process is realized through the steps so as to generate a motor micro-variable data distribution map.

in the embodiment of the invention, aiming at each dimension in the motor micro-data distribution diagram, the distribution condition of each dimension is taken as a node, and the characteristics are extracted from the node as the attributes of the node. Here, the parameters of the probability distribution are used as node attributes, for example, for gaussian distribution, the mean and variance may be used as node attributes. In the probability map model, the degree of association between nodes may be represented by probability values. An innovative approach may be used, such as a relevance calculation based on information entropy. In particular, cross entropy of probability distributions between nodes can be calculated to measure the degree of association between nodes. After the node association degree is calculated, the nodes with the association degree higher than a certain threshold value are connected, and the edges in the network are constructed. These edges represent the correlation of the micro-variant data in different dimensions, and the nodes in different dimensions and the edges between them are combined to form a heterogeneous network. Nodes in this network represent the distribution characteristics of the micro-variant data, and edges represent the relevance in different dimensions. And (5) giving weight to the edges according to the association degree between the nodes while constructing the edges. The weight can be obtained through the normalized value of the association degree, and the heterogeneous network construction can be realized through the steps so as to generate motor state heterogeneous network data.

in the embodiment of the present invention, the motor state heterogeneous network data generated in step S62 is converted into a representation of a graph, where each node represents a distribution feature of micro-variable data, and the edges represent an association relationship between the micro-variable data. By using the core idea of GCN, a multi-layer picture volume lamination layer is built. The picture scroll laminate for each layer consists of two main steps: information aggregation and feature update for each node, the GCN updates the feature representation of the current node by aggregating the feature information of its neighbor nodes. An innovative spectral convolution operation is used here to weight sum the features of neighboring nodes and add self-loop terms to preserve the node's own information. The feature representation of the node is updated by a weighted transformation matrix using the aggregated information. The transformation matrix is determined by the structure and the characteristics of the graph, the association relation among the nodes is fully considered, and through multi-layer GCN layer iteration, each layer can acquire node association information of more layers. This is where the GCN algorithm is innovative, allowing deep information transfer and association mining between nodes. Through the iteration of the multi-layer GCN, the characteristic representation of each node gradually fuses the information of the more distant neighbors to form a richer representation. The new characteristic representation of the nodes is motor state depth association data. Depth-related mining is achieved through the steps to generate motor state depth-related data.

Step S64: and carrying out state causal logic reasoning by combining the neural symbol learning and the motor state depth correlation data to generate a motor state logic causal chain.

In an embodiment of the invention, motor state depth association data is mapped to a neuro-symbolic representation space. This embedding process aims to translate the deep association of data into the form of symbolic logic. A symbol embedded network (Symbol Embedding Network) is selected to achieve this. In the neuro-symbolic representation space, causal logic rules are introduced. These rules describe causal relationships between different motor state variables, such as what effect a certain state might have if another state were to change. To learn these rules, a causal neural network (Causal Neural Networks) approach is employed. And reasoning the motor state depth association data based on the learned causal logic rules. Specifically, causal rules are applied to predict possible outcomes based on changes in current state. This reasoning process is similar to logical reasoning but combines the capabilities of neural networks to deal with complex nonlinear relationships. And reasoning the motor state depth association data based on the learned causal logic rules. Specifically, causal rules are applied to predict possible outcomes based on changes in current state. This reasoning process is similar to logical reasoning but combines the capabilities of neural networks to deal with complex nonlinear relationships. The state cause and effect logic reasoning is realized through the steps so as to generate a motor state logic cause and effect chain.

In an embodiment of the present invention, a plurality of motor state logic causal links have been obtained, which represent different causal relationships of motor operating states. In order to integrate them into one comprehensive state causal relationship data, the association and overlap between chains needs to be considered. First, for different causal links, a complexity correlation analysis algorithm, such as the Apriori algorithm, is used to analyze the correlation between links. This helps identify causal relationship elements shared between chains, as well as cross-correlation conditions between chains. For shared causal elements, they are integrated into a new causal chain. This may be done by identifying overlapping causal chain elements and merging according to their weights. Innovatively, fuzzy logic approaches have also been introduced to handle uncertainty between causal chain elements. For cross-correlation between chains, an innovative algorithm in graph theory, such as a maximum flow minimum cut algorithm, is adopted to judge the cross influence intensity between different chains. This helps to determine which causal links have greater impact and thus give higher weight in the integration, and eventually, the causal links that have undergone correlation analysis, element integration and cross-correlation processing are integrated into one comprehensive logical causal link. The state causal link integration processing can be realized through the steps so as to generate motor state causal relation data.

Preferably, step S7 comprises the steps of:

step S71: performing state transition learning processing on the motor state causal relationship data to generate motor state transition probability data;

step S72: asynchronous state screening is carried out according to the motor state transition probability data so as to generate a motor key state transition path;

step S73: performing state causal network reconstruction based on the motor critical state transition path to generate a motor state causal network reconstruction graph;

step S74: carrying out micro-variation prediction processing by utilizing a motor state causal network reconstruction graph to generate motor micro-variation prediction data;

step S75: and carrying out semantic report programming based on the motor micro-variation state prediction data to generate a motor state prediction report.

According to the invention, state transition learning processing is carried out on the motor state causal relationship data to generate motor state transition probability data, and state transition learning is applied to pay attention to not only the occurrence times of each state of the motor, but also the transition relationship among the states. The method creatively converts the motor state causal relationship data into state transition probability data, can capture the dynamic relationship between states more accurately, and provides a more powerful basis for motor state prediction. Asynchronous state screening is performed according to the motor state transition probability data to generate a motor key state transition path, and the motor state transition probability data is calculated to capture the possibility of direct transition between states and accurately reflect the non-uniformity of state change. The change of state of the motor may be irregular under different operating conditions, and conventional methods may ignore such dynamic characteristics. Through the step, the distribution rule of state transition can be better understood, so that a more accurate basis is provided for analysis of the running state of the motor. And carrying out state causal network reconstruction based on the motor key state transition path to generate a motor state causal network reconstruction graph, and introducing a Bayesian network, a strong probability graph model for more accurately reflecting causal relations among states. Bayesian networks are capable of innovatively revealing conditional dependencies between states, i.e., the probability of occurrence of subsequent states given some pre-states. The method breaks through the limitation of traditional association analysis, and further deepens the association between states into probability dependence. And carrying out micro-variation state prediction processing by using the motor state causal network reconstruction graph to generate motor micro-variation state prediction data, and based on the motor state causal network reconstruction graph, not only bringing the transition relation between states into prediction, but also fusing the attribute information of the nodes. The fusion of the multidimensional information can consider more factors in prediction, so that the comprehensive performance of a prediction model is improved, nodes and edges are extracted from a key state transition path of a motor, and the nodes and edges are used for micro-variation prediction processing. The method innovatively applies the key information of the state causal network to the prediction task directly, so that the prediction model is effectively guided to pay attention to the state changes related to the key paths, and the accuracy and relevance of prediction are improved. Semantic reporting is performed based on the motor micro-variation prediction data to generate a motor state prediction report, and small changes inside the motor, which are often early signs of motor faults, are emphasized from the micro-variation prediction data obtained in step S74. Through careful feature extraction, the frequently ignored minute signals can be captured, a powerful clue is provided for early diagnosis of faults, and the LSTM model is individually adapted to the operation characteristics of the motor by model training through historical micro-variation state data. The model can better fit the change rule of micro variation state data by adjusting the weight in the training process, and more accurate support is provided for the prediction of the individual motor state.

in the embodiment of the invention, the required information is extracted from the motor state causal relationship data. Each state variable and its causal relationship will be represented as a feature vector. These feature vectors include the state variables themselves and their associated attributes, such as time stamps, state values, etc. By carefully selecting features, the relationship between states can be better captured in subsequent state transition learning. A markov chain model is applied to model transitions of motor states. To this end, a state transition matrix will be established in which each element represents a probability of transitioning from one state to another. The markov chain is chosen here because it can be based on the current state, taking into account its relation to the previous state, thus characterizing the law of dynamic transitions between states. For the construction of the state transition matrix, the probability of each state transition needs to be calculated. This innovative step involves considering the correlation between state sequences, not just the mere number of state occurrences. Specifically, the probability of a state transition is calculated using frequency statistics and smoothing methods, such as laplace smoothing. And obtaining a transition probability data set between motor states according to the state transition matrix. The state transition learning process is realized through the steps so as to generate motor state transition probability data.

in the embodiment of the present invention, the probability of each state to other states is calculated based on the motor state transition probability data generated in step S71. These probabilities represent the possibility of transitions between different states of the motor, including direct and asynchronous transitions. Based on the state probability, an asynchronous state screening algorithm, specifically "Hidden Semi-Markov Model" (HSMM), is introduced. HSMM is an extension of the hidden markov model, allowing the duration of the state to be variable, consistent with the non-uniformity of state changes in the motor system. The innovation in HSMM is that it considers not only the transition probabilities between states, but also the duration of the states. In motor systems, where different states may last for different times, HSMM allows to model the change of state in a more accurate way, applying the HSMM algorithm, the "lifetime" of each state, i.e. the period of time from occurrence to disappearance of the state, is calculated from the transition probability and the duration distribution of each state. In this process, the asynchronous transition of the state is taken into account, thereby describing the state change more accurately. Based on the results of the HSMM, those state transition paths can be identified that have a critical impact in the motor system. These paths include not only direct state transitions, but also cover the case of asynchronous transitions. And realizing asynchronous state screening through the steps to generate a motor key state transition path.

in an embodiment of the present invention, a dataset is prepared that includes critical state transition paths. Each critical path contains a series of motor states that transition in the motor system in a sequence, each state being considered a node in the network. And determining the causal relationship between the states according to the key state transition paths. For example, if one state occurs before another state in the path, a directed edge may be established that points from the previous state to the next state. For each state node, the probability distribution of the state given its parent node (preamble state) is estimated from the critical path dataset. This is to capture the conditional dependence between the motor states. A bayesian network based on conditional probability tables is employed here. The directed edge relationships between the state nodes are learned from the data using a structure learning algorithm of the bayesian network, such as a maximum likelihood estimation based method or a constraint based method. These edge relationships represent causal relationships between states. And further learning probability parameters for the learned network structure. By observing the data, probability distribution parameters of each node, and conditional probability tables of each edge are estimated. The state causal network reconstruction is realized through the steps so as to generate a motor state causal network reconstruction diagram.

in the embodiment of the present invention, starting from the motor critical state transition paths obtained in step S72, the nodes and edges related to these paths are extracted from the state causal network reconstruction graph. The transition information between these critical state paths and their corresponding states will be used as a basis for prediction. For each critical state path, useful features are innovatively extracted from the transition relationships between the attributes and states of the nodes. These characteristics may include transition probabilities between states, time delays, state attributes, and so forth. These features will serve as inputs for the prediction of micro-variation. And selecting a proper prediction model according to the characteristics of the motor state prediction task. For example, LSTM is suitable for modeling sequence data, while TCN has advantages in processing time series data, using historical state data as a training set, selected predictive models are trained by optimization algorithms (e.g., adam, SGD, etc.). The goal of the training is to enable the model to capture the nonlinear relationship between states, thereby enabling prediction of micro-variant states. And on the basis of the prediction model after training, the extracted features are used as input to predict the motor micro variation state. The prediction of each critical state path can obtain a prediction result of micro-variation, and the prediction results are integrated to generate a motor micro-variation prediction data set. The micro-variation prediction processing can be realized through the steps so as to generate motor micro-variation prediction data.

In the embodiment of the present invention, data preparation is first performed in the motor micro-variation prediction data obtained in step S74. These data may include minor changes in vibration, sound, current, etc. of the motor in various states, which may be early signs of motor failure, feature extraction is performed for each point in time of the micro-variant data to capture key information. For example, frequency domain features, time domain statistics, etc. for each point in time may be calculated. These features will be used as inputs to the predictive model. An appropriate predictive model is selected. Considering the characteristics of micro-variant data, a Long Short-Term Memory (LSTM) model is selected to be used for prediction. The LSTM model can capture long-term dependency in a time sequence and is suitable for prediction of small changes. The LSTM model is trained using historical microvariant data as a training set. In the training process, the model optimizes the weight through a back propagation algorithm, so that the model can better fit the change rule of the micro variation state data. After training is completed, the trained LSTM model is used for predicting micro-variation state data at future time points. In the process, the input is the feature extracted before, the model outputs the predicted tiny change condition, and a semantic report is compiled based on the tiny variation state prediction data generated by the model. The semantic report preparation is realized through the steps so as to generate a motor state prediction report.

The foregoing is only a specific embodiment of the invention to enable those skilled in the art to understand or practice the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for evaluating the state of a three-phase full-wave brushless motor, comprising the steps of:

2. The method according to claim 1, wherein the specific steps of step S1 are:

3. The method according to claim 1, wherein the specific steps of step S2 are:

4. A method according to claim 3, wherein the motor acoustic wave entropy calculation formula in step S21 is specifically:

wherein H refers to motor frequency spectrum entropy score data, deltaf refers to frequency interval, f _min Refers to the minimum frequency of the motor acoustic wave signal, f _max Refers to the maximum frequency of the sound wave signal of the motor, f refers to the sound wave frequency spectrum data of the motor, p (f) refers to the probability distribution of the sound wave signal under f, t refers to the time variable, N refers to the number of sound wave characteristics, N refers to the index value of the sound wave characteristics, a _n Refers to the amplitude adjustment coefficient, b _n Refers to the frequency offset coefficient, c refers to the phase offset coefficient, d _n Refers to the attenuation coefficient.

5. The method according to claim 1, wherein the specific step of step S3 is:

6. The method according to claim 1, wherein the specific step of step S4 is:

7. The method according to claim 6, wherein the electromagnetic force source point intensity locating formula in step S42 is specifically:

8. The method according to claim 1, wherein step S5 is specifically:

9. The method according to claim 1, wherein the specific step of step S6 is:

10. The method according to claim 1, wherein the state prediction evaluation process includes a state transition learning process, an asynchronous state screening process, a state causal network reconstruction process, a micro-variant state prediction process, and a semantic report preparation process, and step S7 is specifically: