A permanent magnet synchronous motor fault diagnosis method and system based on multi-feature fusion

Abstract

The application discloses a kind of permanent magnet synchronous motor fault diagnosis method and system based on multi-feature fusion, it is suitable for various need to carry out fault diagnosis, state monitoring to signal Application scene, such as motor fault detection, mechanical system state evaluation etc..Among them, the motor vibration signal is taken as fault diagnosis signal, from which the time domain feature and the texture feature of vibration signal corresponding time domain image are extracted, and feature fusion is carried out, finally fusion feature is used to carry out fault diagnosis.Among them, short-time energy entropy ratio is taken as time domain feature, and local binary pattern energy distribution and multi-scale gradient direction feature are introduced, composite feature vector containing global statistical characteristics, local texture and spatial direction information is constructed, comprehensive representation of fault signal time domain, frequency domain, time-frequency domain is realized, the distinguishing ability of feature to different fault modes is significantly improved, the high precision of model is guaranteed, and the problem of fault diagnosis under multiple working conditions is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet synchronous motor fault diagnosis technology, specifically a method and system for fault diagnosis of permanent magnet synchronous motors based on multi-feature fusion, which is applicable to fault detection under multiple operating conditions. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in deep-sea wind power systems and electric vehicles due to their high efficiency and high power density. However, due to prolonged overload operation and the influence of surrounding environmental factors, PMSMs may experience various faults during operation, affecting equipment performance and safety. Traditional fault diagnosis methods often rely on single characteristics, which are insufficient to meet the high-precision diagnostic requirements under complex operating conditions.

[0003] In modern industry and engineering, signal analysis is an important technique for monitoring and diagnosing the operating status of equipment. Traditional signal analysis methods mainly focus on extracting features in the time or frequency domain, such as the root mean square (RMS) value in the time domain and the power spectral density in the frequency domain. However, these single features often fail to fully reflect the complex characteristics of a signal, especially when facing complex fault modes, where their diagnostic accuracy and reliability are limited. In recent years, with the development of wavelet analysis technology, wavelet transform, as a method that can perform signal analysis in both the time and frequency domains simultaneously, has been gradually applied to the field of signal processing. Wavelet transform can decompose a signal into sub-band signals of different frequency bands and extract the features of each sub-band, thereby providing richer signal information. However, most current research based on wavelet transform only focuses on the extraction of single features, neglecting the fusion of other important features.

[0004] In summary, existing diagnostic methods have the following technical problems:

[0005] 1) Traditional methods only focus on the extraction of time-domain features, while ignoring the fusion of other important features, especially the influence of multiple operating conditions. The signals of the same motor will vary, making it difficult for single feature extraction to meet the accuracy requirements of fault diagnosis.

[0006] 2) Most current research based on wavelet transform focuses only on the extraction of a single feature, while ignoring the fusion of other important features. Summary of the Invention

[0007] This invention aims to improve the accuracy of fault diagnosis for permanent magnet synchronous motors (PMSMs), and thus provides a fault diagnosis method and system for PMSMs based on multi-feature fusion. The method of this invention uses the vibration signal of the faulty motor as the raw signal to conduct feature extraction and fault detection research, effectively solving the problem that single details and single feature extraction are insufficient for accurate fault diagnosis. Specifically, a multi-dimensional feature system is constructed based on temporal and texture features to extract fault signal features from different perspectives, improving fault diagnosis accuracy. This invention selects short-time energy entropy to more sensitively capture the periodic changes of temporal fault impacts; the local binary mode energy distribution in texture features characterizes the local texture structure and statistical properties of the temporal image; multi-scale gradient direction features describe the directional distribution characteristics of time-frequency energy at multiple scales to characterize the spatial evolution law of the fault signal; through the combination of the above features, a composite feature vector containing both global statistical characteristics and local texture and spatial direction information is constructed, achieving a comprehensive characterization of the fault signal in the temporal, frequency, and time-frequency domains, significantly improving the feature's ability to distinguish different fault modes.

[0008] Therefore, the present invention provides the following technical solution:

[0009] On the one hand, the present invention provides a fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion, which includes:

[0010] Step 1: Signal acquisition, acquiring vibration signals of the permanent magnet synchronous motor under various fault and / or healthy conditions;

[0011] Step 2: Feature extraction. Extract the temporal features of the vibration signal and the texture features of the corresponding temporal image of the vibration signal. The temporal feature is the short-time energy entropy ratio, which reflects the energy distribution and complexity of the vibration signal in different time windows. The texture features of the temporal image include local binary mode energy distribution features and multi-scale gradient direction features.

[0012] Step 3: Feature fusion, which fuses the extracted temporal and texture features to form a fused feature vector;

[0013] Step 4: Fault diagnosis. The fused feature vector is used as the input to the fault diagnosis classifier, and the fault type is used as the output of the fault diagnosis classifier.

[0014] Specifically, the vibration signals and fault tags of the permanent magnet synchronous motor under various fault and health conditions are used to process the data according to steps 1-3 to train the fault diagnosis classifier. The trained fault diagnosis classifier is used for fault diagnosis or fault monitoring of the permanent magnet synchronous motor.

[0015] Vibration signals can directly reflect the operating status of a motor. Their amplitude, frequency, phase and other characteristics are closely related to the type and degree of motor faults and can provide a wealth of fault characteristic information.

[0016] In some implementations, data preprocessing is also performed. That is, based on the range of motor fault characteristic frequencies, a bandpass filter is used to extract signals within a specific frequency range, and the vibration signal is filtered to normalize the amplitude of the vibration signal to a uniform range [0,1]. This effectively filters out high-frequency noise and low-frequency interference in the vibration signal. Normalization can eliminate the difference in amplitude variation of the vibration signal under different operating conditions, making the vibration signals collected by different sensors or vibration signals from different devices comparable, which is convenient for subsequent data processing and analysis.

[0017] Further, optionally, the short-time energy entropy ratio of the vibration signal is obtained using a short-time energy calculation method and a spectral entropy calculation method, as follows:

[0018] ;

[0019] In the formula, For the first The short-time energy entropy ratio of the frame signal; For the first Logarithmic energy of a frame signal; For the first The spectral entropy of the frame signal; wherein, the vibration signal is windowed and framed, with the frame length set to an integer multiple of the motor rotation period, to ensure that each frame signal contains complete periodic fault impact information, resulting in a series of short-time frame sequence signals in units of "frames", which satisfy:

[0020] ;

[0021] ;

[0022] In the formula, As a reference constant, For adjustment coefficients, To address the instantaneous frequency fluctuation variance, the energy sensitivity is automatically adjusted under variable speed conditions to maintain... The dynamic range is stable; For the first After the frame sequence signal is Fourier transformed, the first... The normalized spectral probability density of each frequency component; L is the signal length of one frame; For the first Frame sequence signal, where n is the number of sampling points.

[0023] The above model, for periodic faults, reintroduces the instantaneous frequency fluctuation variance as a dynamic adjustment coefficient to automatically adjust the energy sensitivity under variable speed conditions.

[0024] Further, optionally, the local binary mode energy distribution feature mentioned in step 2 is to divide the time-frequency image into... Each sub-block is then used to calculate the energy characteristics of each sub-block;

[0025] Specifically, for each pixel, a circular neighborhood is determined centered on the pixel. Based on this circular neighborhood, the local binary pattern encoding value of the pixel is calculated, and the encoding rule is as follows:

[0026] ;

[0027] ;

[0028] In the formula, center pixel The local binary pattern encoded value, i.e., the LBP encoded value. Let p be the coordinates of the center pixel, p be the number of pixels in the circular neighborhood, and r be the radius of the circular neighborhood. The grayscale value of the center pixel. Let be the grayscale value of the i-th pixel within the circular neighborhood. This is a threshold comparison function used to quantize the grayscale comparison result into 0 or 1; a , b For user-defined variable symbols;

[0029] The local binary pattern encoded values ​​are mapped to energy weights, and then an energy weight matrix is ​​generated based on the set of LBP encoded values ​​of all pixels within the entire sub-block. Finally, the variance of the sub-block is calculated based on the energy weight matrix. Entropy As an energy characteristic:

[0030] ;

[0031] ;

[0032] In the formula, Let K be the variance, K be the number of local binary pattern codes in the sub-block, and h be the frequency distribution (energy weight) of each local binary pattern code value in the sub-block. It is a tiny constant; Entropy.

[0033] Further, optionally, the extraction process of the multi-scale gradient direction features in step 2 is as follows:

[0034] First, a Gaussian pyramid is constructed based on the time-frequency image, and then the gradient direction features of each layer of the Gaussian pyramid are calculated, that is, the gradient magnitude and gradient direction of each image pixel in each layer of the image are obtained.

[0035] Then, the image is evenly divided into several non-overlapping units, and the gradient direction of all image pixels in each unit is counted. Weights are generated based on the gradient directions, and then the gradient magnitudes are weighted to obtain a gradient direction histogram.

[0036] Next, adjacent units are combined into blocks, and the gradient direction histograms of all units within each block are spliced ​​together to generate the block's descriptor vector.

[0037] Finally, by traversing all blocks in each layer of the image, the descriptor vectors of the normalized blocks are concatenated in order to obtain the layer feature vector of each layer of the image, and then the layer feature vectors are concatenated layer by layer to form a complete multi-scale gradient direction feature.

[0038] Further, optionally, the process of fusing temporal features and texture features in step 3 is as follows:

[0039] Three hierarchical sub-dictionaries were trained based on prior supervision information of fault type by utilizing temporal features, local binary mode energy distribution features, and multi-scale gradient orientation histogram features, respectively.

[0040] The complete hierarchical sub-dictionary of the m-th class features: , This represents the 1st, 2nd, and Vth sub-dictionaries, where m represents the feature category marker and V is the total number of fault categories;

[0041] Solving the joint sparse coding problem of the three features yields the sparse coefficient vector corresponding to each type of feature, so that the sparse representation of the three features in the dictionary uses the exact same set of atoms in the dictionary.

[0042] After solving the joint sparse coding problem, calculate the response energy for each type of fault. The response energy of each type of fault is compared one by one, and the fault category with the strongest response energy is determined as the dominant fault type, and its category number is recorded. This represents the sparse coefficient sub-block corresponding to the m-th class feature in the v-th sub-dictionary; the class index is used to locate the sparse coefficient sub-block corresponding to the hierarchical sub-dictionary. , The sparse coefficient sub-blocks corresponding to the three feature modes under the dominant fault type are extracted and weighted and summed to obtain the fused feature vector:

[0043]

[0044] In the formula, To fuse feature vectors, , This is a weight matrix for a single type of fault, corresponding to the three characteristic modes.

[0045] Further, optionally, the training process of the hierarchical sub-dictionary is as follows:

[0046] The vibration signal training samples are divided into V mutually exclusive subsets according to the fault type label, including healthy state and various fault types;

[0047] K-SVD dictionary learning is performed independently for each subset of fault types, resulting in V sub-dictionaries. Each sub-dictionary contains J atoms. The V sub-dictionaries are concatenated column by column to form a hierarchical sub-dictionary with complete m-th type features. , used to characterize the typical pattern of type V faults in the type m feature domain;

[0048] The sparse coefficient vector of the m-th class features The corresponding partition is divided into V sparse coefficient sub-blocks:

[0049]

[0050] in, Sub-dictionaries The sparse coefficient submatrix of the corresponding m-th feature, where T is the matrix transpose symbol.

[0051] Further, optionally, the joint sparse coding problem is expressed as:

[0052]

[0053]

[0054]

[0055] In the formula, These represent temporal features, local binary mode energy distribution features, and multi-scale gradient orientation histogram features, respectively. This represents the m-th type of feature representation of the input signal. Let be the sparse coefficient vector of the m-th class feature. It is the square norm of the sum of squares of the reconstruction errors. This is the sparse coefficient submatrix corresponding to the sub-dictionary of the v-th type of fault. It is a group sparsity induced norm that forces three features of the same fault type to activate the same dictionary atoms. For fault type mutual exclusion penalty items, This represents the Hadamard product, ensuring that sparse coefficient sub-blocks of different fault types cannot be non-zero simultaneously. The regularization coefficient is greater than zero, determined through cross-validation, and satisfies... ; superscript , All are fault type markers; T is the matrix transpose symbol, subscript j is the atom marker, and J is the total number of atoms in the sub-dictionary; For the sub-dictionary corresponding to the vth type of fault The sparse coefficient submatrix of class features.

[0056] Alternatively, the fault diagnosis may employ a deep random forest model classifier.

[0057] Secondly, the present invention provides a system based on the above-described fault diagnosis method, comprising:

[0058] The signal acquisition module is used to acquire vibration signals of the permanent magnet synchronous motor under various fault and / or healthy conditions.

[0059] The feature extraction module is used to extract the temporal features of the vibration signal and the texture features of the corresponding temporal image of the vibration signal. The temporal feature is the short-time energy entropy ratio, which reflects the energy distribution and complexity of the vibration signal in different time windows. The texture features of the temporal image include local binary mode energy distribution features and multi-scale gradient direction features.

[0060] The feature fusion module fuses the extracted temporal and texture features to form a fused feature vector;

[0061] The fault diagnosis classifier construction and training module is used to construct a fault diagnosis classifier and generate training sample data using vibration signals and fault labels of permanent magnet synchronous motors in various fault and health states, in order to train the fault diagnosis classifier.

[0062] The diagnostic module is used to take the fused feature vector as the input of the fault diagnosis classifier and the fault type as the output of the fault diagnosis classifier.

[0063] Thirdly, the present invention also provides a computer device including one or more processors;

[0064] A memory that stores one or more computer programs;

[0065] The processor invokes a computer program to implement the steps of a fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion.

[0066] Fourthly, the present invention also provides a computer-readable storage medium, characterized in that: it stores a computer program, which is called by a processor to implement: the steps of a fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion.

[0067] Compared with the prior art, the present invention achieves the following improvements:

[0068] The technical solution of this invention breaks through the limitations of traditional single feature extraction by fusing temporal and texture features to construct a multi-dimensional feature system. This method not only characterizes the transient energy and complexity of the signal from a temporal perspective, but also uses wavelet transform to extract local texture and structural information in the time-frequency domain, thereby extracting fault signal features from different angles and improving the accuracy of fault diagnosis.

[0069] This invention applies the short-time energy-entropy ratio to motor vibration fault diagnosis. Based on the physical characteristics of motor vibration signals (such as the periodicity and modulation of fault impacts and the non-stationarity of background noise), the entropy calculation of the short-time energy-entropy ratio is specifically optimized. Compared with existing short-time energy-entropy ratio methods (such as the energy / entropy ratio commonly used in speech signal processing), which often directly use the ratio of the original energy to entropy, the specific improvements include: introducing logarithmic energy transformation to compress the dynamic range and effectively suppress the excessive amplification of large-amplitude impacts, addressing the large differences in the impact amplitude of mechanical vibration signals; combining spectral entropy and utilizing the reduced spectral complexity at the moment of fault impact to enhance the identification of impact events; and smoothing the energy-entropy ratio through square root processing to improve the stability and detectability of the signal in the time domain, thereby achieving high-precision extraction of the start and end times of impact events in vibration signals and providing more robust temporal features for subsequent fault diagnosis.

[0070] For time-frequency images generated by wavelet transform, local binary mode energy distribution features and multi-scale gradient direction features are extracted. Based on the physical characteristics of the time-frequency image obtained after time-frequency transformation of motor vibration signal (such as local energy accumulation, texture abrupt change and random scattering caused by background noise in the time-frequency domain caused by fault impact), the feature extraction of local binary mode energy distribution is optimized. Compared to existing local binary mode methods (such as uniform LBP and rotation-invariant LBP commonly used in face recognition or texture classification), which often directly use encoded histograms as features, this method offers the following improvements: First, it introduces improved encoding rules, adding diagonal pixel difference comparisons to the traditional center-neighbor pixel comparisons, and introducing a global comparison of the mean values ​​of the center pixel and neighboring pixels to enhance the ability to capture directional texture abrupt changes and suppress random noise interference. Second, it abandons traditional encoded histogram statistics, constructs an energy weight matrix, and extracts intra-block variance and entropy as higher-order statistics to quantify the uniformity and contrast, complexity and randomness of the texture, respectively, to highlight the texture differences between faulty and non-faulty regions. Finally, it introduces a small constant in the entropy calculation for numerical stability optimization, avoiding the zero-probability problem caused by sparse encoding and ensuring stable feature output in low-energy background regions. The local binary mode energy distribution characterizes the fine structure of the fault signal by quantifying the texture uniformity and local contrast of the time-frequency energy map, while the multi-scale gradient direction features characterize the spatial evolution of the fault signal by describing the directional distribution characteristics of the time-frequency energy at multiple scales. By fusing multi-dimensional features, a composite feature vector containing both global statistical characteristics and local texture and spatial orientation information is constructed, which realizes a comprehensive representation of fault signals in the time domain, frequency domain, and time-frequency domain, and significantly improves the feature's ability to distinguish different fault modes. Attached Figure Description

[0071] Figure 1 This is a schematic diagram illustrating the technical concept of a fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion according to the present invention.

[0072] Figure 2 This is a schematic diagram of the model training process. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.

[0074] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0076] This invention provides a fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion, applicable to various application scenarios requiring fault diagnosis and status monitoring of signals, such as motor fault detection and mechanical system status assessment. The technical solution of this invention uses motor vibration signals as the raw signals for fault diagnosis. After preprocessing the collected vibration signals, they are fused, and finally, a classifier is constructed using the fused features. This effectively improves the accuracy of fault diagnosis results. The short-time energy entropy ratio reflects the energy distribution and complexity of the signal within different time windows. Wavelet transform extracts the local binary mode energy distribution and multi-scale gradient direction features from the image, capturing local details and multi-scale information of the signal. Feature vector fusion further enhances the effectiveness of the features, ensuring the accuracy of the model and enabling high-precision fault identification, effectively solving the problem of fault diagnosis under multiple operating conditions.

[0077] The present invention will be further described below with reference to specific embodiments:

[0078] The fault diagnosis and classification method for permanent magnet synchronous motors provided in this invention includes the following core steps: signal acquisition, preprocessing of the acquired signal, extraction and fusion of time-domain and texture features, and construction of a fault diagnosis classifier using the fused features. The details are as follows:

[0079] Step 1: Signal acquisition, acquiring vibration signals of the permanent magnet synchronous motor under various fault and / or healthy conditions;

[0080] Step 2: Feature extraction. After preprocessing the collected vibration signals, temporal and texture features are extracted.

[0081] Step 3: Feature fusion, which fuses the extracted temporal and texture features to form a fused feature vector;

[0082] Step 4: Fault diagnosis. The fused feature vector is used as the input to the fault diagnosis classifier, and the fault type is used as the output of the fault diagnosis classifier.

[0083] The process involves using vibration signals and fault tags from permanent magnet synchronous motors under various fault and health conditions, and processing the data according to steps 1-3 to train a fault diagnosis classifier. The trained fault diagnosis classifier is then used for fault diagnosis or fault monitoring of permanent magnet synchronous motors.

[0084] In some embodiments, after signal acquisition in step 1, a bandpass filter is used to extract signals within a specific frequency range based on the range of motor fault characteristic frequencies. The vibration signal is then filtered to normalize the amplitude of the vibration signal to a uniform range [0,1]. This effectively filters out high-frequency noise and low-frequency interference in the vibration signal. Normalization can eliminate the differences in amplitude variation of the vibration signal under different operating conditions, and at the same time make the vibration signals acquired by different sensors or vibration signals from different devices comparable, which is convenient for subsequent data processing and analysis.

[0085] To improve the diagnostic accuracy of the model, this invention found that fusing the short-time energy entropy ratio of the selected signal with features extracted from wavelet transform images can significantly improve diagnostic accuracy. Therefore, in this embodiment, the time-domain feature is the short-time energy entropy ratio, preferably obtained using both short-time energy calculation and spectral entropy calculation methods. Other feasible embodiments, using short-time energy calculation methods and empirical permutation entropy to calculate the short-time energy entropy ratio, also fall within the scope of this invention, but the effect is weaker than the former. The proposed short-time energy entropy ratio reflects the energy distribution and complexity of the signal within different time windows. Furthermore, the texture features extracted from the wavelet transform image are local binary mode energy distribution features and multi-scale gradient direction features extracted from the wavelet transform image, and preferably, the local binary mode energy distribution features are optimized according to the characteristics of the application object. Figure 1 As shown, the specific implementation process of the optimal embodiment is as follows:

[0086] First, the vibration signal is preprocessed by eliminating DC components and normalizing amplitude. Then, the preprocessed vibration signal is windowed and framed, and then the vibration signal is divided into multiple short time frames to obtain a series of time series signals in units of "frames".

[0087] Among them, the calculation involves windowing and frame segmentation to obtain the first... Frame sequence signal The model is as follows:

[0088]

[0089]

[0090] In the formula, For window functions, This represents the numerical value of a frame of sequence signal. For frame length, The frame shift length, This represents the total number of frames after the vibration signal is framed. The signal is the preprocessed vibration signal, and n is the number of sampling points.

[0091] Then, a Fourier transform is performed on each frame of the sequence signal to obtain its frequency domain representation, and the spectral probability density is extracted, as follows:

[0092] Definition of the first Frame number Normalized spectral probability density of each frequency component for:

[0093]

[0094] In the formula, To perform a Fourier transform on a given frame sequence signal, the th frame... The energy spectrum of each frequency component; L is the frame signal length.

[0095] Finally, the energy of each frame sequence signal is calculated, and the spectral entropy of each frame sequence signal is calculated based on the spectral probability density. The ratio of energy to spectral entropy is then used as the short-time energy-entropy ratio.

[0096] Definition of the first Spectral entropy of a frame The formula is as follows:

[0097]

[0098] Definition of the first The formula for the logarithmic energy of a frame is as follows:

[0099]

[0100] In the formula, It is a constant used to adjust the sensitivity to energy changes, logarithmic energy. This reflects the strength of the signal.

[0101] In some embodiments, further optimizations are made, specifically:

[0102] ;

[0103] In the formula, Reference constant, For adjustment coefficients, To address the instantaneous frequency fluctuation variance, the energy sensitivity is automatically adjusted under variable speed conditions to maintain... Its dynamic range is stable.

[0104] The formula for the short-time energy entropy ratio is as follows:

[0105]

[0106] In the formula, For the first The short-time energy entropy ratio of the frame signal.

[0107] The extraction process for the texture features of the time-domain image corresponding to the vibration signal is as follows:

[0108] First, a wavelet transform is performed on the vibration signal to generate a time-frequency image. In some embodiments, the db5 wavelet basis is used to generate the time-frequency image, which involves converting the three-dimensional wavelet coefficient matrix into a two-dimensional grayscale image. Specifically, the wavelet transform formula is as follows:

[0109]

[0110] In the formula, It is the input signal. These are wavelet basis functions, where a is the scaling parameter and b is the translation parameter; Wavelet coefficients are used to quantify the local characteristics of a signal at a specific frequency and time point.

[0111] Local binary mode energy distribution: dividing the time-frequency image into For each sub-block, the energy characteristics of each sub-block are preferably calculated using an improved local binary mode method;

[0112] For each pixel, a circular neighborhood is determined centered on the pixel, and the local binary pattern encoding value of the pixel is calculated based on the circular neighborhood. The encoding rule is as follows:

[0113] ;

[0114] ;

[0115] In the formula, center pixel Local binary pattern encoded value, Let P be the coordinates of the center pixel, P be the number of pixels in the circular neighborhood, and r be the radius of the circular neighborhood. The grayscale value of the center pixel. Let be the grayscale value of the i-th pixel within the circular neighborhood. This is a threshold comparison function used to quantize the grayscale comparison result into 0 or 1.

[0116] The local binary pattern encoded values ​​are mapped to energy weights, and then an energy weight matrix is ​​constructed based on the set of LBP encoded values ​​of all pixels within the entire sub-block. Finally, the variance of the sub-block is calculated based on the energy weight matrix. Entropy As an energy characteristic, the energy formula for a sub-block is as follows:

[0117]

[0118]

[0119] In the formula, Let K be the variance, K be the number of local binary pattern codes, and h be the frequency distribution of each local binary pattern code value in the sub-block. It is a tiny constant; Entropy.

[0120] Multi-scale gradient orientation features: Constructing a Gaussian pyramid based on time-frequency images , The layer index of the pyramid is used to calculate the gradient direction features of each layer image, which gives the gradient magnitude and gradient direction of each image pixel in each layer image.

[0121]

[0122]

[0123]

[0124]

[0125] In the formula, The horizontal position of the image. This represents the vertical position of the image. The gradient component of the image in the horizontal direction. This represents the gradient component of the image in the vertical direction. For gradient magnitude, The gradient direction is given by the formula. Appearing in is the two-dimensional coordinate index of the image pixels, where I represents the image.

[0126] Then, the image is evenly divided into several non-overlapping units, and the gradient direction of all image pixels in each unit is counted. Weights are generated based on the gradient directions, and then the gradient magnitudes are weighted to obtain a gradient direction histogram.

[0127]

[0128] In the formula, These are the accumulated values ​​of the gradient direction histogram, used to describe the gradient direction distribution characteristics of local regions of the image. This is the directional weight function, whose magnitude depends on the gradient direction at that point. , C represents the two-dimensional coordinate index of an image pixel, and C is the set of all pixels within a unit.

[0129] Next, adjacent units are combined into blocks, and the gradient direction histograms of all units within each block are spliced ​​together to generate the block's descriptor vector.

[0130] Finally, by traversing all blocks in each layer of the image, the descriptor vectors of the normalized blocks are concatenated in order to obtain the layer feature vectors of each layer, and then the layer feature vectors are concatenated layer by layer to form a complete multi-scale gradient direction feature.

[0131] In some embodiments, step 3, the process of fusing temporal features and image features, introduces sparse coding, specifically as follows:

[0132] Three hierarchical sub-dictionaries are trained based on prior information about fault types for temporal features, local binary pattern energy distribution features, and multi-scale gradient orientation histogram features, respectively. The joint sparse coding problem of these three features is solved to obtain the sparse coefficient vector corresponding to each feature class, ensuring that the sparse representations of the three features in the dictionary use the exact same set of atoms from the dictionary.

[0133] The training method for the hierarchical sub-dictionary is as follows: the vibration signal training samples are divided into V mutually exclusive subsets according to the fault type label, where V is the total number of fault categories, including healthy state and various fault types; K-SVD dictionary learning is performed independently on each fault type subset to obtain V sub-dictionaries. Each sub-dictionary contains J atoms. The V sub-dictionaries are concatenated column-wise to form a complete hierarchical dictionary for the m-th feature: , among which, sub-dictionary This represents the typical pattern of type V faults in the m-th characteristic domain.

[0134] The sparse coefficient vector of the m-th class features The corresponding partition is divided into V sparse coefficient sub-blocks:

[0135]

[0136] in, For sub-dictionary The corresponding sparse coefficient sub-block;

[0137] The joint sparse coding problem is represented as:

[0138]

[0139]

[0140]

[0141] In the formula, m represents the feature category marker. These represent temporal features, local binary mode energy distribution features, and multi-scale gradient orientation histogram features, respectively. This represents the m-th type of feature representation of the input signal. It is the square norm of the sum of squares of the reconstruction errors. This is the sparse coefficient submatrix corresponding to the sub-dictionary of the v-th type of fault. It is a group sparsity induced norm that forces three features of the same fault type to activate the same dictionary atoms. For fault type mutual exclusion penalty items, This represents the Hadamard product, ensuring that sparse coefficient sub-blocks of different fault types cannot be non-zero simultaneously. The regularization coefficient is greater than zero, determined through cross-validation, and satisfies... .

[0142] After solving, calculate the response energy for each type of fault. The response energy of each type of fault is compared one by one, and the fault category with the strongest response energy is determined as the dominant fault type, and its category number is recorded. This number is also used to locate the corresponding coefficient sub-block in the hierarchical dictionary. The corresponding coefficient sub-blocks of the three feature modes under the dominant fault type are extracted and weighted to obtain the fused feature vector.

[0143]

[0144] In the formula, To fuse feature vectors, , , These are sparse coefficient sub-blocks of the three feature modes obtained from solving the joint sparse coding problem. , , This refers to the weight matrix corresponding to the three characteristic modes for a single type of fault. It should be understood that the fusion method in the above embodiments is a preferred technique, but not the only one.

[0145] In some embodiments, the fault diagnosis model is a deep random forest algorithm. Each layer of the deep forest consists of multiple random forests (including ordinary random forests and fully random forests). This algorithm introduces a multi-layer cascaded structure on the basis of traditional random forests. Through layer-by-layer feature enhancement, it automatically learns the high-order representation of fused features, thereby further improving the accuracy of fault classification.

[0146] In this embodiment, the fused feature vector is used as input to train and diagnose the model classifier, including:

[0147] 1) Divide the fused feature vector dataset into a training set and a test set;

[0148] 2) Use the training set data to build a deep random forest model classifier;

[0149] 3) Use the validation set data for the optimized training of the deep random forest model classifier.

[0150] The model training process is as follows:

[0151] like Figure 2 For each input sample (the input to the first layer is the original fused feature vector), each forest outputs a class probability vector. The probability vectors from all forests are concatenated to obtain the output of that layer. To prevent information loss, this output is then concatenated with the original fused features and used as the input to the next layer. This process is repeated layer by layer to build deep random forests until the accuracy on the validation set no longer improves; at this point, the final multi-layered deep random forest is obtained. For the test sample, it is input into the trained multi-layered cascaded forest to obtain the class probability vectors of all forests in the last layer. The outputs of all forests in the last layer are concatenated and input into a logistic regression classifier to learn the optimal combination of weights to obtain the final predicted class.

[0152] Based on the above theoretical statements, a diagnostic model (fault diagnosis classifier) ​​that can be used to realize fault diagnosis of permanent magnet synchronous motors is constructed, and the motor can be fault diagnosed based on this fault diagnosis classifier.

[0153] In some embodiments, the present invention provides a system based on the above-described fault diagnosis method, which includes a signal acquisition module, a feature extraction module, a feature fusion module, a fault diagnosis classifier construction and training module, and a diagnosis module that are connected in sequence or to each other.

[0154] The signal acquisition module is used to acquire vibration signals of the permanent magnet synchronous motor when it is running under various fault and / or healthy conditions.

[0155] The feature extraction module is used to extract the temporal features of the vibration signal and the texture features of the corresponding temporal image of the vibration signal. The temporal feature is the short-time energy entropy ratio, which is used to reflect the energy distribution and complexity of the vibration signal in different time windows. The texture features of the temporal image include local binary mode energy distribution features and multi-scale gradient direction features.

[0156] The feature fusion module fuses the extracted temporal and texture features to form a fused feature vector.

[0157] The fault diagnosis classifier construction and training module is used to build a fault diagnosis classifier and generate training sample data using vibration signals and fault labels of permanent magnet synchronous motors in various fault and health states, in order to train the fault diagnosis classifier.

[0158] The diagnostic module is used to take the fused feature vector as the input of the fault diagnosis classifier and the fault type as the output of the fault diagnosis classifier.

[0159] In some embodiments, the present invention also provides a computer device including one or more processors and a memory storing one or more computer programs; wherein the processor invokes the computer programs to implement the steps of a method for fault diagnosis of permanent magnet synchronous motors based on multi-feature fusion.

[0160] In some embodiments, the electronic components of a computer device include:

[0161] The processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0162] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and the processor calls the algorithm program to execute the above methods.

[0163] Input / output interfaces are used to implement information input and output.

[0164] The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0165] A bus is used to transfer information between various components of a device, such as processors, memory, input / output interfaces, and communication interfaces.

[0166] The processor, memory, input / output interfaces, and communication interfaces communicate with each other within the device via a bus.

[0167] In some embodiments, the present invention also provides a computer-readable storage medium, characterized in that: it stores a computer program, which is invoked by a processor to implement: the steps of a method for fault diagnosis of a permanent magnet synchronous motor based on multi-feature fusion.

[0168] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the hardware and software device described in any of the foregoing embodiments, such as the hard drive or memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both internal storage units and external storage devices of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0169] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also protected by this invention.

Claims

1. A fault diagnosis method for permanent magnet synchronous motors based on multi-feature fusion, characterized in that: include: Step 1: Signal acquisition, acquiring vibration signals of the permanent magnet synchronous motor under various fault and / or healthy conditions; Step 2: Feature extraction. Extract the temporal features of the vibration signal and the texture features of the corresponding temporal image of the vibration signal. The temporal feature is the short-time energy entropy ratio, which reflects the energy distribution and complexity of the vibration signal in different time windows. The texture features of the temporal image include local binary mode energy distribution features and multi-scale gradient direction features. Step 3: Feature fusion, which fuses the extracted temporal and texture features to form a fused feature vector; Step 4: Fault diagnosis. The fused feature vector is used as the input to the fault diagnosis classifier, and the fault type is used as the output of the fault diagnosis classifier. Specifically, the vibration signals and fault tags of the permanent magnet synchronous motor under various fault and health conditions are used to process the data according to steps 1-3 to train the fault diagnosis classifier. The trained fault diagnosis classifier is used for fault diagnosis or fault monitoring of the permanent magnet synchronous motor.

2. The method according to claim 1, characterized in that: The short-time energy entropy ratio of the vibration signal is obtained using short-time energy calculation and spectral entropy calculation methods, as follows: ； In the formula, For the first The short-time energy entropy ratio of the frame signal; For the first Logarithmic energy of a frame signal; For the first The spectral entropy of the frame signal; wherein, the vibration signal is windowed and framed, with the frame length set to an integer multiple of the motor rotation period, to ensure that each frame signal contains complete periodic fault impact information, resulting in a series of short-time frame sequence signals in units of "frames", which satisfy: ； ； In the formula, As a reference constant, For adjustment coefficients, To address the instantaneous frequency fluctuation variance, the energy sensitivity is automatically adjusted under variable speed conditions to maintain... The dynamic range is stable; For the first After the frame sequence signal is Fourier transformed, the first... The normalized spectral probability density of each frequency component; L is the signal length of one frame; For the first Frame sequence signal, where n is the number of sampling points.

3. The method according to claim 1, characterized in that: The local binary mode energy distribution feature mentioned in step 2 is to divide the time-frequency image into... Each sub-block is then used to calculate the energy characteristics of each sub-block; Specifically, for each pixel, a circular neighborhood is determined with the pixel as the center, and the local binary pattern encoding value of the pixel is calculated based on the circular neighborhood; The local binary pattern encoded values ​​are mapped to energy weights, and then an energy weight matrix is ​​constructed based on the set of LBP encoded values ​​of all pixels within the entire sub-block. Finally, the variance of the sub-block is calculated based on the energy weight matrix. Entropy As an energy characteristic: ； ； In the formula, K represents the number of local binary pattern codes in the sub-block, and h represents the frequency distribution of each local binary pattern code value in the sub-block. It is a tiny constant.

4. The method according to claim 1, characterized in that: The extraction process of multi-scale gradient direction features in step 2 is as follows: First, a Gaussian pyramid is constructed based on the time-frequency image, and then the gradient direction features of each layer of the Gaussian pyramid are calculated, that is, the gradient magnitude and gradient direction of each image pixel in each layer of the image are obtained. Then, the image is evenly divided into several non-overlapping units, and the gradient direction of all image pixels in each unit is counted. Weights are generated based on the gradient directions, and then the gradient magnitudes are weighted to obtain a gradient direction histogram. Next, adjacent units are combined into blocks, and the gradient direction histograms of all units within each block are spliced ​​together to generate the block's descriptor vector. Finally, by traversing all blocks in each layer of the image, the descriptor vectors of the normalized blocks are concatenated in order to obtain the layer feature vector of each layer of the image, and then the layer feature vectors are concatenated layer by layer to form a complete multi-scale gradient direction feature.

5. The method according to claim 1, characterized in that: The process of fusing temporal and texture features in step 3 is as follows: Three hierarchical sub-dictionaries were trained based on prior supervision information of fault type by utilizing temporal features, local binary mode energy distribution features, and multi-scale gradient orientation histogram features, respectively. The complete hierarchical sub-dictionary of the m-th class features: , This represents the 1st, 2nd, and Vth sub-dictionaries, where m represents the feature category marker and V is the total number of fault categories; Solving the joint sparse coding problem of the three features yields the sparse coefficient vector corresponding to each type of feature, so that the sparse representation of the three features in the dictionary uses the exact same set of atoms in the dictionary. After solving the joint sparse coding problem, calculate the response energy for each type of fault. The response energy of each type of fault is compared one by one, and the fault category with the strongest response energy is determined as the dominant fault type, and its category number is recorded. This represents the sparse coefficient sub-block corresponding to the m-th class feature in the v-th sub-dictionary; the class index is used to locate the sparse coefficient sub-block corresponding to the hierarchical sub-dictionary. , The sparse coefficient sub-blocks corresponding to the three feature modes under the dominant fault type are extracted and weighted and summed to obtain the fused feature vector: ； In the formula, To fuse feature vectors, , , This is a weight matrix for a single type of fault, corresponding to the three characteristic modes.

6. The method according to claim 5, characterized in that: The training process for the hierarchical sub-dictionary is as follows: The vibration signal training samples are divided into V mutually exclusive subsets according to the fault type label, including healthy state and various fault types; K-SVD dictionary learning is performed independently for each subset of fault types, resulting in V sub-dictionaries. Each sub-dictionary contains J atoms. The V sub-dictionaries are concatenated column by column to form a hierarchical sub-dictionary with complete m-th type features. , used to characterize the typical pattern of type V faults in the type m feature domain; The sparse coefficient vector of the m-th class features The corresponding partition is divided into V sparse coefficient sub-blocks: ； in, Sub-dictionaries The sparse coefficient submatrix of the corresponding m-th feature, where T is the matrix transpose symbol.

7. The method according to claim 5, characterized in that: The joint sparse coding problem is represented as: ； ； ； In the formula, These represent temporal features, local binary mode energy distribution features, and multi-scale gradient orientation histogram features, respectively. This represents the m-th type of feature representation of the input signal. Let be the sparse coefficient vector of the m-th class feature. It is the square norm of the sum of squares of the reconstruction errors. This is the sparse coefficient submatrix corresponding to the sub-dictionary of the v-th type of fault. It is a group sparsity induced norm that forces three features of the same fault type to activate the same dictionary atoms. For fault type mutual exclusion penalty items, This represents the Hadamard product, ensuring that sparse coefficient sub-blocks of different fault types cannot be non-zero simultaneously. The regularization coefficient is greater than zero, determined through cross-validation, and satisfies... ; superscript , All are fault type markers; T is the matrix transpose symbol, subscript j is the atom marker, and J is the total number of atoms in the sub-dictionary; For the sub-dictionary corresponding to the vth type of fault The sparse coefficient submatrix of class features.

8. A diagnostic system based on the method of any one of claims 1-7, characterized in that: include: The signal acquisition module is used to acquire vibration signals of the permanent magnet synchronous motor under various fault and / or healthy conditions. The feature extraction module is used to extract the temporal features of the vibration signal and the texture features of the corresponding temporal image of the vibration signal. The temporal feature is the short-time energy entropy ratio, which reflects the energy distribution and complexity of the vibration signal in different time windows. The texture features of the temporal image include local binary mode energy distribution features and multi-scale gradient direction features. The feature fusion module fuses the extracted temporal and texture features to form a fused feature vector; The fault diagnosis classifier construction and training module is used to construct a fault diagnosis classifier and generate training sample data using vibration signals and fault labels of permanent magnet synchronous motors in various fault and health states, in order to train the fault diagnosis classifier. The diagnostic module is used to take the fused feature vector as the input of the fault diagnosis classifier and the fault type as the output of the fault diagnosis classifier.

9. A computer device, characterized in that: include: One or more processors; A memory that stores one or more computer programs; The processor invokes a computer program to achieve the following: The steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer program is stored and is invoked by the processor to implement: The steps of the method according to any one of claims 1-7.

Images