Power harmonic frequency spectrum motor fault diagnosis system combined with deep learning

Abstract

The invention relates to the technical field of motor fault diagnosis, and discloses a power harmonic frequency spectrum motor fault diagnosis system combined with deep learning, which comprises a signal acquisition and preprocessing module used for acquiring a power signal from a motor and preprocessing the power signal, including denoising and normalization processing; a harmonic spectrum analysis and feature extraction module which is connected with the signal acquisition and preprocessing module and is used for carrying out Fourier transform and wavelet transform time-frequency analysis on the preprocessed electric power signal and extracting the features of the motor fault; and the deep learning feature optimization and classification module is connected with the harmonic spectrum analysis and feature extraction module. Through deep learning, multi-task learning and physical constraint optimization, efficient and accurate motor fault classification is realized, model robustness is enhanced, dependence on large-scale labeled data is reduced, and stability and adaptability of a diagnosis system under complex working conditions are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly to a power harmonic spectrum motor fault diagnosis system combined with deep learning. Background Art

[0002] With the rapid development of industrial automation, motors are widely used in various fields such as production equipment and household appliances. As an important power source of mechanical systems, the stability of motor operation directly affects the overall performance and safety of equipment. In order to ensure the efficient operation of motors and extend their service life, it is particularly important to perform motor fault diagnosis in a timely and effective manner.

[0003] In the prior art, common motor fault diagnosis methods include traditional signal processing methods based on vibration signal analysis, sound signal analysis, and current signal analysis. These methods mainly rely on signal processing technologies such as Fourier transform and wavelet transform to perform spectrum analysis and feature extraction on motor fault signals. Through these features, the system can detect potential faults of the motor. In addition, there are also some methods based on manual experience, which usually rely on the knowledge and experience of experts to judge the fault types and are applicable to some simple and common fault situations. With the popularization of machine learning, more and more research has begun to adopt data-driven diagnosis methods, which classify by analyzing a large amount of historical fault data and have achieved good results.

[0004] However, the prior art still has some deficiencies; firstly, traditional signal processing methods are often unable to cope when dealing with complex and diverse motor faults, especially when the signals are affected by noise interference or there are a large number of fault types, and the accuracy and robustness are greatly reduced; secondly, the fault diagnosis methods relying on manual experience are inefficient and lack consistency. Especially when facing complex working conditions, the knowledge of experts cannot comprehensively cover all possible fault types; in addition, data-driven methods usually require a large amount of labeled data, and in practical applications, the labeled data may be severely insufficient, and these methods are not easily adaptable to new or uncommon fault types; more importantly, the applications of existing physical constraint models are mostly limited to specific fault types and lack generality and flexibility. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a power harmonic spectrum motor fault diagnosis system combined with deep learning, which solves the problems of insufficient accuracy, poor robustness, and excessive dependence on data in existing motor fault diagnosis methods in complex fault scenarios.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A power harmonic spectrum motor fault diagnosis system combined with deep learning, comprising:

[0007] A signal acquisition and preprocessing module, which is used to acquire power signals from the motor and perform preprocessing, including denoising and normalization processing;

[0008] A harmonic spectrum analysis and feature extraction module, which is connected to the signal acquisition and preprocessing module, and is used to perform Fourier transform and wavelet transform time-frequency analysis on the preprocessed power signals, and extract the features of motor faults;

[0009] A deep learning feature optimization and classification module, which is connected to the harmonic spectrum analysis and feature extraction module, and is used to optimize the extracted features through a deep neural network and perform fault type classification, and output the classification results of motor faults;

[0010] A multi-task learning and knowledge transfer module, which is connected to the deep learning feature optimization and classification module, and is used to perform joint training on multiple fault types through a multi-task learning framework, and optimize the learning ability of new fault types through knowledge distillation;

[0011] A physical constraint and robustness enhancement module, which is connected to the deep learning feature optimization and classification module, and is used to combine the physical characteristics of the motor to perform constraint optimization on the deep learning model, and enhance the stability and robustness of the model;

[0012] A diagnosis result output module, which is connected to the physical constraint and robustness enhancement module, and is used to judge the fault type according to the fault classification result output by the deep learning feature optimization and classification module, and display the fault diagnosis result of the motor through a visualization interface.

[0013] Preferably, the signal acquisition and preprocessing module includes:

[0014] A current signal acquisition unit, which is used to acquire the current signal of the motor in real time;

[0015] A noise removal unit, which is used to remove high-frequency noise and low-frequency interference in the motor signal through a band-pass filter;

[0016] A normalization unit, which is used to perform standardization processing on the acquired power signals to eliminate the influence caused by signal amplitude differences.

[0017] Preferably, the harmonic spectrum analysis and feature extraction module includes:

[0018] A Fourier transform unit, which is used to convert the time-domain signal into a frequency-domain signal to obtain the spectral characteristics of the motor signal;

[0019] A wavelet transform unit, which is used to perform time-frequency analysis on the power signal and extract the multi-scale features of the motor fault signal;

[0020] A feature extraction unit, which is used to extract the features of harmonic amplitude, phase and spectral density from the spectral signal.

[0021] Preferably, the deep learning feature optimization and classification module includes:

[0022] A deep neural network unit for learning and optimizing the features of the power harmonic spectrum signal;

[0023] An L1 norm regularization unit for sparsifying and optimizing the feature selection of the network to reduce the influence of redundant features;

[0024] A classification unit for classifying the extracted features through a deep learning model to determine the fault type of the motor.

[0025] Preferably, the multi-task learning and knowledge transfer module includes:

[0026] A multi-task learning unit for knowledge transfer between different fault modes by sharing intermediate layer features;

[0027] A knowledge distillation unit for transferring the existing model knowledge to a new model to optimize the learning of new fault types.

[0028] Preferably, the physical constraint and robustness enhancement module includes:

[0029] A physical modeling unit for constructing physical constraints according to the physical characteristics of the motor;

[0030] A constraint optimization unit for introducing the physical constraints of the motor into the deep learning model to optimize the learning process of the model;

[0031] A Bayesian optimization unit for optimizing the hyperparameters of the model to enhance the robustness of the model under different working conditions.

[0032] Preferably, the diagnostic result output module includes:

[0033] A fault judgment unit for judging the operating state of the motor and determining whether there is a fault according to the output result of the deep learning model;

[0034] A visualization display unit for graphically displaying the diagnostic results to facilitate the operator to view the fault type and health status of the motor.

[0035] Preferably, the deep neural network unit in the deep learning feature optimization and classification module includes:

[0036] A convolutional neural network unit for extracting hierarchical features from the power harmonic spectrum image;

[0037] A recurrent neural network unit for processing the temporal features of the motor signal;

[0038] Pooling layer, used for performing feature dimensionality reduction operations in a convolutional neural network, reducing computational complexity and preventing overfitting.

[0039] Preferably, the classification unit in the deep learning feature optimization and classification module further includes:

[0040] Support vector machine classification unit, used for classifying fault types of the features optimized by deep learning;

[0041] Random forest classification unit, used to make fault diagnosis more accurate and avoid overfitting.

[0042] Preferably, the Bayesian optimization unit in the physical constraint and robustness enhancement module includes:

[0043] Gaussian process regression unit, used for modeling and optimizing hyperparameters in the motor fault diagnosis process;

[0044] Sampling strategy unit, used for selecting the optimal combination of hyperparameters to accelerate the speed of model training.

[0045] The present invention provides a power harmonic spectrum motor fault diagnosis system combined with deep learning. It has the following beneficial effects:

[0046] 1. Through the deep learning model and multi-task learning technology, the present invention realizes accurate classification of motor fault types. Compared with traditional manual feature extraction methods, the system can automatically identify and classify fault types from complex signals, improving the classification accuracy and solving the problem that traditional methods are difficult to handle complex or diverse fault types.

[0047] 2. Introducing the constraints of motor physical laws makes the model output conform to the physical conditions in the actual working environment. This not only enhances the rationality of prediction but also avoids the problem of the model overfitting the data. Compared with the traditional pure data-driven method, the system can better ensure that the predicted results of the output are physically consistent, solving the deficiency that the model prediction may not conform to the actual physical state.

[0048] 3. Through technologies such as adversarial training and data augmentation, the present invention strengthens the robustness of the model, especially being able to provide stable diagnostic results even in noisy and changing environments. Compared with the prior art, this robustness design ensures that the system can maintain efficient operation under various interferences during motor operation, avoiding the risk of traditional methods failing under unstable conditions. Description of the Drawings

[0049] Figure 1 It is the system structure diagram of the present invention;

[0050] Figure 2 It is the module architecture diagram of the harmonic spectrum analysis and feature extraction module of the present invention;

[0051] Figure 3 This is the module architecture diagram of the deep learning feature optimization and classification module of the present invention;

[0052] Figure 4 This is the module architecture diagram of the multi-task learning and knowledge transfer module of the present invention;

[0053] Figure 5 This is the module architecture diagram of the physical constraint and robustness enhancement module of the present invention;

[0054] Figure 6 This is the module architecture diagram of the diagnostic result output module of the present invention. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the specification of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0056] Please refer to the attached Figure 1 - attached Figure 6 , the embodiments of the present invention provide a power harmonic spectrum motor fault diagnosis system combined with deep learning, including:

[0057] A signal acquisition and preprocessing module, configured to acquire power signals from the motor and perform preprocessing, including denoising and normalization processing;

[0058] The signal acquisition and preprocessing module is used to acquire power signals from the motor and perform necessary processing to ensure that the subsequent feature extraction and fault classification modules can operate efficiently based on clean and standardized signals. The core tasks of the signal acquisition and preprocessing module include: signal acquisition, noise removal, signal normalization, etc.

[0059] The power signals of the motor mainly include current signals and voltage signals. Generally, signal acquisition is achieved by installing current sensors and voltage sensors in the current loop and voltage loop of the motor. These sensors can collect the running signals of the motor in real time. Due to vibrations, load changes, and external environmental interference during the operation of the motor, the signals collected by the sensors often contain a large amount of noise, which may reduce the accuracy of the signals. Therefore, further processing of the signals is required.

[0060] In some embodiments, the sampling frequency of the signal acquisition unit generally depends on the operating frequency of the motor and the required diagnostic accuracy. According to the load and operating state of the motor, the system can dynamically adjust the sampling frequency. In the case of a heavy motor load, the sampling frequency is usually high to ensure that the collected signals are more detailed; in the case of a light load, the sampling frequency can be appropriately reduced to save computing resources.

[0061] Since various noises and interferences are generated during the operation of the motor, the original signal usually contains unwanted high-frequency components, low-frequency interferences, and external electromagnetic noises, etc. Therefore, the primary task of signal preprocessing is to remove these noises to ensure the signal quality for subsequent processing.

[0062] Generally, a band-pass filter is used for noise removal. The band-pass filter can remove the low-frequency and high-frequency noise components and only retain the useful frequency components in the motor signal. The working principle of the band-pass filter is to select a frequency range such that the signals within this range can pass through, while the signals outside this frequency range are attenuated or filtered out.

[0063] Specifically, the band-pass filter can be defined as the following convolution operation:

[0064]

[0065] where: y(t) is the filtered signal; h(t - τ) is the impulse response function of the band-pass filter, representing the response of the filter to the time delay τ; τ is the time delay; dτ is the infinitesimal increment of the integration variable, representing the infinitesimal time step used in calculating the integral.

[0066] Through this formula, the band-pass filter performs a time-domain convolution operation on the signal to filter out the noise components whose frequencies are not within the set range. In motor fault diagnosis, a working frequency range (for example, 50Hz to 500Hz) is usually selected to ensure that only the frequency signals related to the motor operation are retained.

[0067] During the signal acquisition and preprocessing process, the differences in signal amplitude may affect the subsequent feature extraction and classification processing. The signal amplitudes generated by different motors under different load conditions may vary greatly, so it is necessary to standardize the signal so that the subsequent analysis can be based on a unified scale.

[0068] The standardization method usually adopts Z-score standardization, which can adjust the signal to a standard normal distribution with a mean of zero and a variance of one. The specific standardization process is as follows:

[0069]

[0070] Where: z(t) is the normalized signal; x(t) is the original signal; μ is the mean of the original signal; σ is the standard deviation of the original signal.

[0071] This method converts all signal data to a standard format with zero mean and unit variance, eliminating the effects of varying amplitudes. This standardization helps with subsequent deep learning model training, ensuring that the model doesn’t favor specific signals due to large amplitude variations.

[0072] Through the aforementioned noise removal and normalization, the signal acquisition and preprocessing module provides a clean, standardized signal for use by subsequent modules. During fault diagnosis, the accuracy of the signal acquisition and preprocessing module directly impacts the effectiveness of the subsequent analysis modules. Using bandpass filtering, the system effectively removes noise and interference from the motor signal, ensuring the accuracy of subsequent analysis results. Furthermore, signal normalization ensures consistent signal amplitude, facilitating subsequent feature extraction and classification.

[0073] In some embodiments, in addition to bandpass filters, adaptive filters can also be used. Adaptive filters dynamically adjust filter parameters based on the characteristics of the input signal, making the filtering effect more flexible and adaptable to different types of noise. This is particularly effective for the highly variable noise generated by motors in different operating environments.

[0074] Adaptive filters usually use the least mean square error (LMS) algorithm, which minimizes the output error by continuously adjusting the filter weights. The operating formula of the adaptive filter is as follows:

[0075] y(t)=w(t)·x(t);

[0076] Where: y(t) is the filter output; w(t) is the filter weight; x(t) is the input signal.

[0077] This method can adaptively adjust the filter parameters when the signal environment changes significantly, providing a more accurate signal denoising effect.

[0078] Alternatively, the signal normalization method can use the minimum-maximum normalization method to map the minimum and maximum values of the signal to a predetermined range (e.g., [0, 1]). This method is suitable for signals with large amplitude variations, especially when the motor load varies greatly. For example, the minimum-maximum normalization formula is:

[0079]

[0080] Where: x norm (t) is the normalized signal; x(t) is the original signal; xmin and x max are the minimum and maximum values of the signal respectively.

[0081] By compressing the signal range to between [0, 1], this method makes the signal amplitudes of different motors no longer affect the subsequent processing.

[0082] The signal acquisition and preprocessing module removes noise through a band-pass filter and adjusts the signal amplitude through normalization, providing clear and standardized signal data for the subsequent modules. Through this module, the system can effectively filter out common noises and interferences during motor operation and eliminate the influence of amplitude differences on feature extraction and classification.

[0083] The harmonic spectrum analysis and feature extraction module, which is connected to the signal acquisition and preprocessing module, is used to perform Fourier transform and wavelet transform time-frequency analysis on the preprocessed power signal and extract the features of motor faults;

[0084] The main task of the harmonic spectrum analysis and feature extraction module is to further analyze the power signal processed by the signal acquisition and preprocessing module and extract the key spectral features related to motor faults. These features can provide an effective basis for subsequent fault diagnosis and classification systems to determine faults.

[0085] Through time-frequency analysis methods such as Fourier transform and wavelet transform, this module can extract frequency-domain information and time-frequency information from the power signal that helps identify motor faults. Fourier transform can reveal the distribution of different frequency components in the signal, while wavelet transform provides more detailed time-frequency local features, which is very effective for detecting non-stationary signals (such as signal changes caused by motor faults).

[0086] Generally, after the power signal is collected and preprocessed, Fourier transform is first used to perform frequency-domain analysis on the signal and convert it into a frequency-domain signal. The formula for Fourier transform is as follows:

[0087]

[0088] where: X(f) is the frequency-domain signal, representing the amplitude and phase of the signal at frequency f; x(t)1 is the time-domain signal, representing the original current or voltage signal collected by the motor; e -j2πft is the complex exponential function, representing the basis function in Fourier transform; f is the frequency, representing the frequency component in the frequency domain; j is the imaginary unit, representing the phase information of the frequency component; t1 is the time, representing the change of the signal over time; dt is the differential element of integration, representing the integration over all possible values of time t1.

[0089] Role of Fourier transform: Through Fourier transform, the original time-domain signal is converted into a frequency-domain signal. By analyzing the frequency-domain signal, the performance of the motor at different frequencies can be observed. Especially when the motor fails, the amplitude or phase of certain frequency components in the spectrum will change, which provides an important basis for fault diagnosis.

[0090] The difference between wavelet transform and Fourier transform is that wavelet transform can perform analysis in both the time and frequency domains simultaneously. Therefore, when processing motor fault signals, especially when the signals have time-local characteristics, wavelet transform can more accurately capture the changing characteristics of the signals.

[0091] The formula for wavelet transform is:

[0092]

[0093] where: W(a, b) is the wavelet transform coefficient, representing the characteristics of the signal at different scales (a) and positions (b); x(t2)2 is the time-domain signal, representing the original current or voltage signal collected from the motor sensor; ψ * is the complex conjugate of the mother wavelet, representing the basis function used for signal decomposition; is the complex conjugate of the wavelet basis function. Here, ψ(t) is the mother wavelet function, and ψ * (t) is its complex conjugate function, usually used to extract the characteristics of the signal; a is the scale factor, controlling the stretching of the wavelet; b is the translation factor, controlling the displacement of the wavelet; t2 is the time, representing the time axis of the signal; dt represents the integration variable, and in the integration process, the signals at all times t2 are weighted and summed.

[0094] Role of wavelet transform: Wavelet transform can decompose the signal at different scales, obtaining the local frequency information of the signal at each time point. This is particularly effective for the analysis of non-stationary signals. In motor fault diagnosis, the fault signals are usually non-stationary, so wavelet transform can better capture the time-varying characteristics of the signals, thereby improving the accuracy of diagnosis.

[0095] Through Fourier transform and wavelet transform, the system can extract the frequency-domain characteristics and time-frequency characteristics of the power signal. These characteristics include but are not limited to harmonic amplitude, harmonic phase, spectral density, etc.

[0096] Generally speaking, the harmonic amplitude is the amplitude at each frequency point in the spectrum, reflecting the intensity of that frequency component. It can be calculated through Fourier transform, specifically as:

[0097] A(f) = |X(f)|;

[0098] Where: A(f) is the amplitude at frequency f in the spectrum, representing the intensity of that frequency component; |X(f)| is the amplitude of the frequency-domain signal, representing the amplitude of that frequency component.

[0099] Function: By calculating the amplitude in the spectrum, the main frequency components during the operation of the motor can be revealed. Especially when a fault occurs, the amplitude of certain frequency components changes significantly, providing important characteristics of the motor fault.

[0100] The harmonic phase refers to the phase information of each frequency component in the spectrum, which reflects the time delay of the signal. The phase calculation can be obtained through the phase part of the Fourier transform:

[0101] φ(f) = arg(X(f));

[0102] Where: φ(f) is the phase at frequency f in the spectrum, representing the phase information of that frequency component; arg(X(f)) is the phase of the frequency-domain signal, representing the phase angle of that frequency component.

[0103] Function: The harmonic phase can provide time delay information for fault diagnosis. For motor faults, the phase change of the signal can be an important characteristic to help identify different types of faults.

[0104] The spectral density represents the energy distribution of the signal in different frequency ranges, which can reflect the energy distribution of the motor signal. The spectral density is usually calculated by squaring the spectrum amplitude:

[0105] S(f) = |X(f)| 2 ;

[0106] Where: S(f) is the spectral density at frequency f in the spectrum, representing the energy of that frequency component; |X(f)| 2 is the square of the amplitude of the frequency-domain signal, representing the intensity of the signal energy.

[0107] Function: By calculating the spectral density, the system can reveal the energy distribution of the signal, helping to identify the frequency intervals where the energy is concentrated in the signal, which is usually the most significant manifestation when the motor fails.

[0108] After extracting features through Fourier transform and wavelet transform, the system will obtain the frequency-domain features and time-frequency features of the motor signal. These features will be further transmitted to the subsequent deep learning feature optimization and classification module for fault type classification.

[0109] Specifically, through the extracted features such as harmonic amplitude, harmonic phase, and spectral density, etc., the system can compare the motor signal with known fault types, thus achieving efficient and accurate fault diagnosis.

[0110] In some embodiments, in addition to using Fourier transform and wavelet transform, the system can also incorporate the Empirical Mode Decomposition (EMD) method. EMD is an adaptive decomposition method based on the characteristics of the signal itself. It can decompose the signal into multiple Intrinsic Mode Functions (IMFs), and each IMF represents different frequency components of the signal. By analyzing these IMFs, local features in the motor signal can be extracted more meticulously, further improving the accuracy of diagnosis.

[0111] As another option, the system can also incorporate the Short-Time Fourier Transform (STFT). By sliding a window over the signal and performing Fourier transform within each window, a time-frequency representation of the signal is obtained. This method can provide more refined spectral information than the Fourier transform, and is particularly suitable for situations where the frequency components of the signal vary over time.

[0112] By calculating features such as harmonic amplitude and harmonic phase, the system can reveal abnormal behaviors during the operation of the motor, thereby providing important basis for subsequent fault diagnosis and classification.

[0113] The deep learning feature optimization and classification module, which is connected to the harmonic spectrum analysis and feature extraction module, is used to optimize the extracted features through a deep neural network and perform fault type classification, and output the classification result of the motor fault;

[0114] The deep learning feature optimization and classification module is a core component of the power harmonic spectrum motor fault diagnosis system. It is responsible for learning meaningful patterns from the signal features extracted from the previous module and performing fault classification. This module classifies the input signal using a deep neural network (especially the Convolutional Neural Network CNN) to accurately determine whether the motor has a fault and its type. This module optimizes the features to improve the accuracy of classification, while reducing redundant features and improving the efficiency of the system.

[0115] In the previous steps, the signal has been denoised and normalized by the signal acquisition and preprocessing module, and frequency domain features and time-frequency features have been extracted by the harmonic spectrum analysis and feature extraction module. These features provide crucial information for subsequent fault classification. The task of the deep learning feature optimization and classification module is to further optimize these features and perform classification, thereby achieving accurate diagnosis of motor faults.

[0116] In some embodiments, the system uses a Convolutional Neural Network (CNN) to process the extracted feature data. CNN is particularly suitable for processing data with structural or spatial features, such as spectrograms. Through operations such as convolution, pooling, and fully connected layers, CNN can efficiently extract and learn local and global features in the signal.

[0117] The core operation process of CNN includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer performs convolution operations on the input data, the pooling layer reduces the dimension of the convolution result, reduces the computational amount, and enhances the generalization ability of the model, and the fully connected layer is used for the final fault type classification.

[0118] Specifically, the convolution operation in CNN can be expressed as:

[0119]

[0120] where: y(t3)1 is the output signal after convolution, representing the feature at time t3. The result of the convolution operation is used to capture the local features of the signal at this time point; x(t3 + k) is the value of the input signal at time t3 + k, representing the offset of the input signal in the convolution window; w(k) is the weight of the convolution kernel, representing the weighting coefficient of each signal element in the convolution operation. The convolution kernel is automatically adjusted according to the learning process to extract useful signal features; k is the time delay in the convolution operation, controlling the sliding position of the convolution kernel on the input signal. It is the offset of the input signal required to calculate each convolution output.

[0121] This formula indicates that the convolution operation extracts features by weighted summing different parts of the input signal, thereby generating a local representation of the signal.

[0122] The pooling layer is used to reduce the dimension of the output of the convolutional layer, thereby reducing the computational amount and preventing overfitting. The max pooling operation selects the maximum value in each pooling window, and the formula is as follows:

[0123] y(t4)2 = max(x(t4), x(t4 + 1), …, x(t4 + n2));

[0124] where: y(t4)2 is the output after pooling, representing the signal feature after the pooling operation. It is the maximum value within the pooling window; x(t4), x(t4 + 1), …, x(t4 + n) are the signal values within the pooling window, representing all the input signal values within the current pooling window; t4 is the position of the pooling window, representing the starting position of the pooling window; n2 is the size of the pooling window, representing the number of signal data points included in the window.

[0125] This formula indicates that max pooling retains the most significant features by selecting the maximum value within the pooling window while reducing the dimension of the input data.

[0126] In the fully connected layer of the neural network, each input node is connected to each node in the output layer. The calculation formula of the fully connected layer is:

[0127]

[0128] Among them: y is the output of the fully connected layer, representing the final classification result. It obtains the classification result by weighted summing the input features and performing activation; x i is the input feature, representing the feature data passed from the previous layer. These features are the signals processed by the convolutional layer and the pooling layer; w i is the weight, representing the weight coefficient of the input feature. Each feature value x i is multiplied by its corresponding weight w i to calculate the weighted sum; b is the bias term, representing the offset of the output. It is used to adjust the output result to ensure more flexibility in the learning process of the network; n3 is the number of input features, representing the number of features passed from the previous layer to the current layer.

[0129] This formula generates the final classification result by multiplying the input features by the corresponding weights and then adding the bias term.

[0130] L1 regularization is used to optimize the feature selection of the model and encourage the network to select sparse features. The loss function formula of L1 regularization is as follows:

[0131]

[0132] Among them: is the total loss function, representing the overall loss of the deep learning model. It is the sum of the data loss and the regularization loss; is the data loss, representing the error of the model on the training data. It measures the fitting degree of the model to the training data; λ is the regularization coefficient, controlling the strength of L1 regularization. A larger λ value will result in stronger sparsity, making more weights approach zero; ∥W∥1 is the L1 norm, representing the L1 norm of the weight matrix W. It is the sum of the absolute values of all weights in the weight matrix and has the effect of sparse feature selection; L1 regularization improves the accuracy of the model by constraining the L1 norm of the weight matrix and prompting the neural network to select important features and reduce unnecessary features.

[0133] The Softmax activation function is used to convert the network output into the probability distribution of each category, so that the prediction result of each category can be represented as a probability. The formula of the Softmax function is as follows:

[0134]

[0135] Among them: P(y=k∣x) represents the probability that the motor fault belongs to the k-th category, giving the prediction probability of each fault type; z kThe score corresponding to class k in the model output layer, representing the predicted score of the network for class k. A higher score indicates a greater predicted probability for that class; e is the base of the natural logarithm, used to calculate the exponential function; K is the total number of classes, representing the number of fault types. This value determines the number of classes in the classification task; z j The scores of other classes in the model output layer, representing the predicted scores of other classes.

[0136] The Softmax function transforms the raw output of the network into a probability distribution, making the sum of probabilities for each class equal to 1, thus enabling effective classification of fault types.

[0137] Through the convolution operation, pooling operation, and fully connected layers in the Convolutional Neural Network (CNN), as well as the L1 norm regularization and Softmax activation function, the deep learning feature optimization and classification module can efficiently extract and optimize the features of motor signals and perform accurate fault classification.

[0138] The multi-task learning and knowledge transfer module, which is connected to the deep learning feature optimization and classification module, is used to jointly train multiple fault types through a multi-task learning framework and optimize the learning ability of new fault types through knowledge distillation;

[0139] The core purpose of the multi-task learning and knowledge transfer module is to enhance the learning ability of the model in multiple fault diagnosis tasks and optimize the learning efficiency of new fault types through multi-task learning (MTL) and knowledge distillation (Knowledge Distillation) techniques. This module learns in multiple fault type tasks by sharing some weights of the neural network, thereby improving the overall diagnostic accuracy and further enhancing the generalization ability of the model through knowledge distillation.

[0140] In this embodiment, the multi-task learning (MTL) technique is used to simultaneously train the classification tasks of multiple fault types. Motor faults usually have multiple types, such as rotor faults, stator faults, bearing faults, etc. Through multi-task learning, the model can share the weights of the intermediate layer to improve the learning ability in these tasks and promote information sharing between different tasks. Each task has its own loss function, and by means of a weighted total loss function, the system can balance the contributions of each task to model training.

[0141] Generally, the total loss function of multi-task learning can be expressed as:

[0142]

[0143] Where: is the total loss function, representing the sum of the weighted losses of all tasks in multi-task learning; T is the number of tasks, representing the total number of fault types processed by the system; is the loss function for the i-th task, representing the classification loss related to fault type i; λ i is the weighting coefficient for task i, controlling the contribution of this task to the total loss.

[0144] The task weighting coefficient λ i is usually adjusted according to the importance and difficulty of each task. For motor fault diagnosis, some fault types may be more common or easier to detect than others, so their weighting coefficients may be higher.

[0145] By sharing some weights of the neural network, the models of different tasks can learn common features, improving the overall performance and efficiency. The weighted loss function ensures that the system focuses more on important tasks during training while avoiding neglecting difficult tasks.

[0146] Knowledge distillation technology is mainly used to improve the learning ability of small models. Generally, knowledge distillation transfers knowledge from a large model (teacher model) to a small model (student model), enabling the student model to achieve a performance similar to that of the teacher model by mimicking the output of the teacher model. The teacher model is usually a large model that has been trained on multiple tasks and can effectively classify multiple fault types, while the student model is trained by learning the output of the teacher model (instead of the original labels).

[0147] The loss function of knowledge distillation can be expressed as:

[0148]

[0149] where: is the distillation loss function, representing the total loss of the student model optimized through knowledge distillation; is the data loss, representing the error of the student model on the training data, usually the traditional cross-entropy loss for classifying labels; α is the balancing coefficient; is the divergence, representing the difference between the probability distributions of the outputs of the teacher model and the student model. KL divergence measures the relative entropy between two probability distributions and is usually used to measure the similarity of the prediction results of the teacher model and the student model. The specific formula is:

[0150]

[0151] where: is the KL divergence, measuring the difference between the probability distributions of the outputs of the teacher model and the student model; P teacher (k2): the prediction probability of the teacher model for class k2; P student (k2) is the prediction probability of the student model for class k2; k2 is the total number of classes, representing the total number of fault types.

[0152] Function: Through knowledge distillation, the student model can learn more detailed knowledge from the output of the teacher model, thereby enhancing its classification ability for new fault types. The teacher model is trained on a large dataset and can provide high-quality knowledge support for the student model, enabling the student model to achieve better learning results with limited data.

[0153] To better balance the influence of each task in multi-task learning, the task weighting coefficient λ i is crucial to design. Generally, the task weighting coefficient is dynamically adjusted according to the loss value of each task. In the motor fault diagnosis task, different fault types may have different learning difficulties. The classification tasks of some fault types may be more challenging, while the data of some fault types may be more abundant.

[0154] The task weighting coefficient λ i can be dynamically adjusted through the following formula:

[0155]

[0156] where: λ i is the weighting coefficient of task i, used to dynamically adjust the weight of the task; is the loss value of task i, representing the loss of task i in the current training stage. Tasks with smaller loss values will be assigned larger weighting coefficients, and vice versa.

[0157] By dynamically adjusting the task weighting coefficient, the system can preferentially process more complex or data-scarce tasks during training, thereby improving the accuracy of fault classification and the training efficiency of the model.

[0158] As an option, in this embodiment, transfer learning can also be combined to further improve the learning efficiency and accuracy of the model. Transfer learning can help the small model quickly adapt to new tasks in the case of scarce data by transferring the knowledge in the large model to the small model.

[0159] Transfer learning is represented by the following formula:

[0160]

[0161] where: is the total loss of transfer learning, representing the loss of the student model optimized under the guidance of the teacher model's knowledge; is the training loss of the student model, representing the classification error of the student model; is the loss of the teacher model, representing the guidance loss of the teacher model to the student model; β is the coefficient of transfer learning, controlling the degree of integration of the student model and the teacher model's knowledge.

[0162] Through transfer learning, the student model can accelerate the learning of new fault types by imitating the knowledge of the teacher model.

[0163] Through task weighting, transfer learning, and knowledge sharing, the system can improve classification accuracy, reduce training time, and enable the model to work effectively even in the case of scarce data. This module provides stronger learning ability and higher diagnostic accuracy for motor fault diagnosis, ensuring the reliability and stability of the system.

[0164] The physical constraint and robustness enhancement module, which is connected to the deep learning feature optimization and classification module, is used to combine the physical characteristics of the motor to constrain and optimize the deep learning model, enhancing the stability and robustness of the model;

[0165] The physical constraint and robustness enhancement module combines the physical characteristics of the motor with robustness enhancement techniques to optimize the training process of the deep learning model. The introduction of physical constraints helps to ensure that the model output conforms to the actual working laws of the motor, while robustness enhancement can improve the stability and anti-interference ability of the model in different working environments. Specifically, this module enables the fault diagnosis model to not only have better generalization ability but also maintain high accuracy and reliability in the actual industrial environment through physical constraint optimization and robustness enhancement means.

[0166] In the foregoing steps, the system optimizes the signal features through the deep learning model and performs fault classification. However, in the actual application of motor fault diagnosis, the signal is usually affected by factors such as noise, vibration, and temperature changes. To cope with these uncertainties, the physical constraint and robustness enhancement module enhances the robustness of the model by combining the known physical laws of the motor with the deep learning model, thereby improving its performance in complex environments.

[0167] Generally, the working state and fault types of the motor are closely related to some known physical laws. For example, there are specific physical relationships between the current and voltage signals of the motor and the fault types. Therefore, physical constraints ensure that the prediction results of the model always conform to the actual physical characteristics of the motor by incorporating these physical laws into the loss function.

[0168] In some embodiments, the physical constraint loss function can be expressed as:

[0169]

[0170] Where: is the physical constraint loss, representing the difference between the model output and the physical laws of the motor. This loss term encourages the model output to be consistent with the known physical characteristics; N is the number of samples, representing the number of samples processed by the model during training; is a function of actual physical laws, representing the true physical state (such as current, voltage, etc.) of the motor under the i-th sample; is the physical state predicted by the model, representing the physical state of the motor predicted by the model.

[0171] By introducing physical constraints, the model can consider the limitations of the motor's physical characteristics during training, avoiding prediction results that violate physical laws. For example, the current or vibration signals of the motor in a specific fault mode have certain regularities, and the output of the model must conform to these known regularities.

[0172] In actual industrial applications, motor fault diagnosis often faces interference from noise, incomplete data, and other external factors. To enable the model to remain efficient and accurate in the face of these uncertain factors, robustness enhancement techniques improve the model's adaptability to noise through various means.

[0173] Robustness enhancement methods include but are not limited to adversarial training, data augmentation, and regularization techniques. For example, common regularization techniques include Dropout and L2 regularization, which can prevent the model from overfitting to noise and data perturbations and enhance the robustness of the model.

[0174] In some embodiments, the robustness-enhanced loss function can be expressed as:

[0175]

[0176] Where: is the robustness-enhanced loss, representing the total loss obtained by enhancing the robustness; is the data loss, representing the error of the model on the training data, usually the classification loss; is the regularization loss, representing constraining the complexity of the model through regularization methods (such as L2 regularization), thereby enhancing the robustness of the model to noise; α2 is the regularization coefficient, controlling the influence of the regularization loss on the total loss. A larger α2 value means a larger regularization weight.

[0177] Regularization loss can be represented by the L2 norm:

[0178]

[0179] Where: is the regularization loss, representing constraining the complexity of the model through L2 regularization; λ3 is the regularization coefficient, controlling the influence of the regularization term on the loss function; w i3 is the model parameter, representing the weight of the i3-th neuron; m is the number of model parameters, representing the total number of parameters in the neural network.

[0180] Through L2 regularization, the weights of the model are restricted within a certain range, thereby reducing the sensitivity of the model to noise and small perturbations and improving its robustness.

[0181] To enhance the stability and accuracy of the motor fault diagnosis model, this embodiment combines physical constraints with robustness enhancement and optimizes both simultaneously during the training process. By jointly optimizing physical constraints and robustness enhancement, the system can not only ensure that the model follows the physical laws of the motor but also improve its adaptability under noise and perturbations.

[0182] The combined loss function can be expressed as:

[0183]

[0184] Where: is the combined loss function, representing the total loss that combines data loss, physical constraint loss, and robustness enhancement loss; is the data loss, representing the error of the model on the training data; is the physical constraint loss, representing the difference between the model output and the physical laws of the motor; The robustness enhancement loss represents the loss of enhancing the model's robustness through regularization; λ4 is the weighting coefficient of the physical constraint, controlling the contribution of the physical constraint to the total loss; α3 is the weighting coefficient of the robustness enhancement, controlling the impact of the regularization loss on the total loss.

[0185] Through this joint optimization, the model can enhance its adaptability to noise and perturbations while satisfying physical constraints, thereby improving the stability and diagnostic accuracy of the system.

[0186] Through physical constraints, the model output conforms to the actual physical behavior of the motor, enhancing the credibility of the model; through robustness enhancement, the model can adapt to the noise and changes in the data, improving its stability in practical applications. The method of jointly optimizing physical constraints and robustness enhancement enables the system to still perform motor fault diagnosis stably and accurately in complex environments, providing stronger diagnostic capabilities and higher reliability.

[0187] The diagnostic result output module, which is connected to the physical constraint and robustness enhancement module, is used to judge the fault type based on the fault classification result output by the deep learning feature optimization and classification module and display the motor fault diagnosis result through a visualization interface;

[0188] The diagnostic result output module is the last link of the power harmonic spectrum motor fault diagnosis system, which undertakes the task of converting the fault diagnosis results obtained through the previous modules (such as feature extraction, deep learning classification, multi-task learning and knowledge transfer modules) into actionable fault information and analysis results. This module not only includes the classification output of fault types, but also can provide information such as the location of the fault, the severity assessment of the fault, and the recommended solution. Through effective display means, the diagnostic results of motor faults are made easy to understand and corresponding maintenance measures can be taken.

[0189] In the aforementioned module, the system has completed the signal processing, feature extraction and optimization through the deep learning model, and has carried out fault classification. The main task of the diagnostic result output module is to integrate these results, generate the final diagnostic result, and display it to the user through visualization means, so that the motor maintenance personnel can accurately judge the type, location and severity of the motor fault and make corresponding decisions.

[0190] In this embodiment, the core task of the diagnostic result output module is to convert the results (i.e., fault types) processed and classified by the previous module into clear and easy-to-understand fault information. The output includes but is not limited to content such as fault type, fault occurrence location, fault severity, and recommended maintenance operations. Specifically, the output content can be generated through the following steps:

[0191] After being processed by the deep learning network, the model classifies the input signal and outputs the probability value of each fault type. The Softmax activation function is used to convert the network output into a probability distribution, indicating the prediction probability of each fault type. Based on these probabilities, the system can determine the most likely fault type.

[0192] Generally, the following formula is used for probability calculation:

[0193]

[0194] Where: P(y = k3|x) represents the probability that the motor fault belongs to the k3th class, and gives the prediction probability of each fault type; is the score of the k3th class in the network output layer, indicating the prediction score of the model for the class k3. A higher score means a higher probability for this class; e is the base of the natural logarithm, used to calculate the exponential function; k3 is the total number of classes, indicating the number of all possible fault types; z j is the score of other classes, indicating the scores of the network output layer for other classes.

[0195] The function of this formula is to convert the original scores output by the network into the probabilities of each class, and the system can output the most likely fault type according to these probabilities.

[0196] In addition to the classification of fault types, the diagnostic result output module also evaluates the severity of the fault based on the fault type and its characteristics. The severity evaluation can be carried out according to factors such as the predicted probability of the fault category and the degree of influence of the fault mode on the normal operation of the motor. Specifically, the model can calculate the severity of the fault by learning historical data or expert rules, based on the known fault types and the operating state of the motor.

[0197] In some embodiments, the severity evaluation may be calculated by the following formula:

[0198]

[0199] Where: is the severity score of fault type k4, representing the severity of the fault type; P(y = k4|x) is the predicted probability of fault type k4, coming from the Softmax classifier; C k is the common influence factor of fault category k4, representing the potential influence of the motor fault type on the motor performance. For example, a short circuit in the stator winding may cause serious damage to the motor, while a rotor fault may lead to a decrease in motor efficiency.

[0200] By calculating the severity score, the system can evaluate the influence degree of different fault types, thus providing a basis for subsequent maintenance decisions.

[0201] The determination of the fault location usually depends on the sensor data of the motor and the output of the classification model. In some embodiments, the fault location can be judged by combining the input data of multiple sensors. For example, if the sensor data shows abnormal rotor temperature, it may indicate that a rotor fault has occurred, and the system can output "rotor fault" as the fault location.

[0202] According to the fault type and the occurrence location, the system can automatically recommend corresponding solutions, such as suggesting component replacement, adjustment or further detection, etc.

[0203] To enhance the user experience and ensure that operators can quickly obtain key information, the diagnostic result output module also displays the diagnostic results through visualization means. Generally, the system will display the results in the following forms:

[0204] Pie chart or bar chart: Display the classification probabilities of different fault types to help users quickly identify the most likely fault type.

[0205] Heat map: Display the fault probabilities or severities of various parts of the motor to help users identify the parts with the most serious problems in the motor.

[0206] Trend chart: Display the evolution trend of faults during the operation of the motor to help users predict the possible paths of fault development.

[0207] For example, the trend of motor faults can be shown through a time series graph, in which the trend of the fault probability changing over time can be marked to help users understand whether the fault is intensifying.

[0208] In addition to the graphical display, the system can also generate a detailed fault diagnosis report, which includes:

[0209] Fault description: motor fault type, occurrence location, severity, etc.;

[0210] Diagnostic analysis: the process of the model's judgment of the fault type, including eigenvalue, probability distribution, etc.;

[0211] Repair suggestions: repair solutions for different fault types and suggestions for further inspection.

[0212] Through the classification output of fault types, severity assessment, judgment of fault locations, and suggestions for solution measures, the system can help users quickly locate problems and take effective measures. In addition, combining functions such as graphical display, trend analysis, and fault diagnosis reports, the system can provide more comprehensive and accurate fault diagnosis information to ensure the efficient operation of the motor.

[0213] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power harmonic spectrum motor fault diagnosis system combined with deep learning, characterized in that, It includes: A signal acquisition and preprocessing module, which is used to acquire power signals from the motor and perform preprocessing, including denoising and normalization processing; A harmonic spectrum analysis and feature extraction module, which is connected to the signal acquisition and preprocessing module, and is used to perform Fourier transform and wavelet transform time-frequency analysis on the preprocessed power signals, and extract the characteristics of motor faults; A deep learning feature optimization and classification module, which is connected to the harmonic spectrum analysis and feature extraction module, and is used to optimize the extracted features through a deep neural network and perform fault type classification, and output the classification result of motor faults; A multi-task learning and knowledge transfer module, which is connected to the deep learning feature optimization and classification module, and is used to perform joint training on multiple fault types through a multi-task learning framework, and optimize the learning ability of new fault types through knowledge distillation; A physical constraint and robustness enhancement module, which is connected to the deep learning feature optimization and classification module, and is used to combine the physical characteristics of the motor to constrain and optimize the deep learning model, and enhance the stability and robustness of the model; A diagnostic result output module, which is connected to the physical constraint and robustness enhancement module, and is used to judge the fault type according to the fault classification result output by the deep learning feature optimization and classification module, and display the fault diagnosis result of the motor through a visualization interface.

2. The power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, wherein The signal acquisition and preprocessing module includes: A current signal acquisition unit, which is used to acquire the current signal of the motor in real time; A noise removal unit, which is used to remove high-frequency noise and low-frequency interference in the motor signal through a band-pass filter; A normalization unit, which is used to perform standardization processing on the acquired power signals to eliminate the influence caused by signal amplitude differences.

3. The power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, wherein The harmonic spectrum analysis and feature extraction module includes: A Fourier transform unit, which is used to convert the time-domain signal into a frequency-domain signal to obtain the spectrum characteristics of the motor signal; A wavelet transform unit, which is used to perform time-frequency analysis on the power signal and extract the multi-scale characteristics of the motor fault signal; A feature extraction unit, which is used to extract the characteristics of harmonic amplitude, phase and spectral density from the spectrum signal.

4. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The deep learning feature optimization and classification module includes: A deep neural network unit, which is used to learn and optimize the characteristics of the power harmonic spectrum signal; An L1 norm regularization unit, which is used to perform sparse optimization on the feature selection of the network to reduce the influence of redundant features; A classification unit, which is used to classify the extracted features through a deep learning model to judge the fault type of the motor.

5. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The multi-task learning and knowledge transfer module includes: A multi-task learning unit, which is used to transfer knowledge between different fault modes by sharing intermediate layer features; A knowledge distillation unit, which is used to transfer the existing model knowledge to the new model to optimize the learning of new fault types.

6. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The physical constraint and robustness enhancement module includes: A physical modeling unit, which is used to construct physical constraints according to the physical characteristics of the motor; A constraint optimization unit, which is used to introduce the physical constraints of the motor into the deep learning model to optimize the learning process of the model; A Bayesian optimization unit, which is used to optimize the hyperparameters of the model to enhance the robustness of the model under different working conditions.

7. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The diagnostic result output module includes: A fault judgment unit, configured to judge the operating state of the motor according to the output result of the deep learning model and determine whether there is a fault; A visualization display unit, configured to graphically display the diagnosis result, facilitating the operator to view the fault type and health status of the motor.

8. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The deep neural network unit in the deep learning feature optimization and classification module includes: A convolutional neural network unit, configured to extract hierarchical features from the power harmonic spectrum image; A recurrent neural network unit, configured to process the temporal features of the motor signal; A pooling layer, configured to perform feature dimensionality reduction operations in the convolutional neural network, reducing the computational complexity and preventing overfitting.

9. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that The classification unit in the deep learning feature optimization and classification module further includes: A support vector machine classification unit, configured to classify the fault types of the features optimized by deep learning; A random forest classification unit, configured to make the fault diagnosis more accurate and avoid overfitting.

10. A power harmonic spectrum motor fault diagnosis system combined with deep learning according to claim 1, characterized in that, The Bayesian optimization unit in the physical constraint and robustness enhancement module includes: A Gaussian process regression unit, configured to model and optimize the hyperparameters in the motor fault diagnosis process; a sampling strategy unit, configured to select the optimal hyperparameter combination to accelerate the model training speed.

Images