Motor current interpretable diagnosis method based on dual-channel collaborative optimization deep convolutional neural network model

Abstract

The invention provides a motor current interpretable diagnosis method based on a dual-channel collaborative optimization deep convolutional neural network model. Firstly, time-frequency domain features of current signals are extracted through multi-stage wavelet decomposition and sideband energy analysis, and it is ensured that the physical significance of fault features is clear; secondly, a dual-output network structure is constructed, the dual-output network structure comprises a physical feature head and a classification head, and the physical feature head restrains the matching degree of hidden layer features and theoretical physical parameters through a spectrum consistency loss function; physical significance projection of a high-dimensional feature space is realized based on t-SNE, and topological consistency of feature distribution and a fault mode is verified through spectral clustering; and establishing a physical parameter inversion channel, carrying out error convergence analysis on a physical feature vector output by a network hidden layer and a theoretically calculated feature frequency, and ensuring that a diagnosis result simultaneously meets a data driving criterion and a physical mechanism constraint. According to the method, the advantage of high precision of deep learning is maintained, and meanwhile, the model decision interpretability and the reliability of a diagnosis result are improved.

Description

Technical Field

[0001] The present invention belongs to the field of electromechanical equipment status monitoring and fault diagnosis, and specifically relates to a motor current interpretable diagnosis method based on a dual-channel collaboratively optimized deep convolutional neural network model, which is used to achieve collaborative optimization of data-driven characterization of motor fault characteristics and physical mechanism explanation, thereby improving the reliability of motor fault diagnosis results and the interpretability of model decisions. Background Art

[0002] With the rapid development of industrial automation, motors, as core power equipment, have a significant impact on production stability and safety. Traditional fault diagnosis methods rely on periodic inspections and expert judgment. This manually-driven maintenance model is no longer able to meet the high reliability and real-time requirements of modern industrial systems. Against this backdrop, Intelligent Fault Diagnosis (IFD) technology offers a new approach to motor condition monitoring by deeply mining implicit features in equipment operating data. Motor current signals, as a direct indicator of physical failure processes such as winding heating and magnetic field distortion, have become core monitoring data for fault diagnosis due to their readily available availability.

[0003] Intelligent fault diagnosis (IFD) technology using current signals has demonstrated significant advantages in the field of complex system fault diagnosis. By automatically extracting deep features from equipment operating data through multi-layer neural networks, it achieves high-precision fault diagnosis in critical equipment such as rotating machinery. However, its inherent "black box" nature has become a bottleneck restricting its application in scenarios requiring high reliability. Specifically, while existing models can accurately identify fault types, their decision-making processes cannot effectively explain the explicit association between current signals and physical failure mechanisms (such as winding short circuits and broken rotor bars). This lack of interpretability results in a lack of physical root cause explanation for diagnostic results, potentially leading to unpredictable risks in high-reliability industrial scenarios, thus limiting its widespread application in practical engineering. To overcome this technical bottleneck and enhance the safety and reliability of IFD systems, it is imperative to establish an interpretable correlation mechanism between model outputs and equipment physical characteristics. Therefore, developing interpretable intelligent fault diagnosis models has become a key research focus.

[0004] To address these challenges, this paper proposes a data-physics collaboratively driven deep learning approach that systematically integrates domain knowledge-driven physical priors into a deep neural network architecture, creating a physically interpretable intelligent diagnostic framework. Specifically, in the data-driven dimension, a deep feature extractor with multiscale perception is constructed, utilizing stacked dilated convolutional layers to capture the nonlinear statistical regularities in current signals. In the physical-driven dimension, a physical feature parsing layer based on wavelet time-frequency analysis theory is constructed to quantitatively characterize the physical properties of fault features. A differentiable physical constraint layer is constructed to transform fault prior knowledge into spectral constraints. Combined with a physical feature regularization term, a loss function is constructed that embeds physical knowledge to improve model interpretability. In the forward propagation phase, physical feature parsing is used to achieve a physical explanation of diagnostic decisions. In the backpropagation phase, spectral constraints are used to dynamically modify the feature space distribution, forming a physically interpretable closed-loop architecture. This approach achieves a deep coupling of statistical regularities and physical mechanisms through the interaction of dual-domain features, establishing correlations between statistical and physical features. This not only imbues the model with the powerful representational capabilities of data-driven approaches, but also significantly improves the reliability of diagnostic results and the interpretability of model decisions. Summary of the Invention

[0005] To address the above issues, the present invention discloses a motor current interpretable diagnosis method based on a dual-channel collaboratively optimized deep convolutional neural network model, thereby resolving the problem of insufficient interpretability of traditional motor fault diagnosis models.

[0006] A motor current interpretable diagnosis method based on a dual-channel collaboratively optimized deep convolutional neural network model comprises the following steps: S1: Collect the motor's historical health and current signals during actual operation, and pre-process the collected signals; S2: Construct a network model for motor fault diagnosis; S3: Optimize and design network hyperparameters; S4: Evaluate model performance.

[0007] Furthermore, step S1 is specifically as follows: S1-1: The acquisition system consists of a high-precision current sensor and a synchronous data acquisition module. The sampling frequency is set to 12800Hz to obtain the current data during the healthy operation of the motor. , the measured current data during actual operation .

[0008] S1-2: pre-processing the collected current signal; Furthermore, step S2 is specifically as follows: S2-1: Divide the samples; S2-2: Construct a physically interpretable time-frequency feature tensor as model input; S2-3: Construct physical constraint module; S2-4: Construct a dual-channel collaborative optimization fault diagnosis module; S2-5: Build a t-SNE feature visualization module.

[0009] Furthermore, step S3 is specifically as follows: S3-1: Apply dynamics to decay learning rate; S3-2: Apply Selu activation function to control the distribution of output values; S3-3: Apply dropout technology to improve network overfitting.

[0010] Compared with the prior art, the present invention has the following effects: The basic design concept of the physics-guided deep convolutional neural network (PG-DCNN) model is as follows: To address the non-stationary nature of motor current signals, a multi-scale wavelet feature extraction module is first constructed. Through multi-level wavelet decomposition combined with time-frequency domain statistics calculation and normalization, a physically interpretable feature vector is extracted. Batch normalization is used to suppress covariate shift, and maximum pooling operations in convolutional and pooling layers are used to enhance the translation invariance of features. Furthermore, a spectral physical constraint module is introduced in the frequency domain branch to extract spectral features through FFT transforms, and spectral constraints are injected into the model as prior physical knowledge. To achieve coordinated optimization of data-driven and physical mechanisms, the model adopts a dual-channel collaborative output structure: the classification channel uses a softmax classifier based on global features extracted by the convolutional layers to identify fault modes. The physical channel adds a physical head, outputting physical feature predictions through a linear layer, and constraining network parameters through physical feature regularization. During training, the total loss function integrates cross-entropy loss, spectral constraint loss, and physical regularization terms. By dynamically adjusting the weights of each loss, it achieves joint optimization of physical mechanisms and data-driven approaches. Ultimately, based on gradient backpropagation and multi-task learning, classification accuracy and physical feature consistency are jointly optimized. This approach retains the strong representational capabilities of deep networks while incorporating domain knowledge priors, effectively overcoming the black-box limitations of traditional data-driven approaches and significantly improving the interpretability of motor fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Flowchart of the motor current interpretable diagnosis method based on dual-channel collaborative optimization deep convolutional neural network model.

[0012] Figure 2 Schematic diagram of the model structure of the deep convolutional neural network (PG-DCNN) guided by physical information.

[0013] Figure 3Fault diagnosis results of the physics-based deep convolutional neural network (PG-DCNN) model. DETAILED DESCRIPTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0015] like Figure 1-3 As shown, a motor current interpretable diagnosis method based on a dual-channel collaboratively optimized deep convolutional neural network model includes the following steps: S1: Collect the motor's historical health and current signals during actual operation, and pre-process the collected signals S1-1: The acquisition system consists of a high-precision current sensor and a synchronous data acquisition module. The sampling frequency is set to 12800Hz to obtain the current data during the healthy operation of the motor. , the measured current data during actual operation .

[0016] S1-2: pre-processing the collected current signal; The processing method is FIR filter, and the purpose of data preprocessing is to reduce noise in the signal.

[0017] S2: Building a network model for motor fault diagnosis This paper proposes a physics-guided deep convolutional neural network (PG-DCNN) intelligent motor fault diagnosis model. The model consists of five modules, each of which forms a closed loop through feature transfer and gradient backpropagation. Physical prior knowledge guides feature learning through a customized loss function, and a visualization module provides decision interpretability. The following details are provided: S2-1: Divide the samples The collected vibration and current signal datasets were divided into training, validation, and test sets in a ratio of 7:2:1. The training set is used to train the model and is the primary source of data features for the model to learn. The validation set is used to adjust the model's hyperparameters and evaluate its performance on unseen data to prevent overfitting. The test set is used for the final evaluation of the model's performance and to verify its practical applicability.

[0018] S2-2: Constructing physically interpretable time-frequency feature tensors as model input By recursively downsampling the wavelet basis function in scale space and detail space, the time-frequency localized features of the signal are extracted layer by layer. Each subband component is normalized by Z-score to eliminate dimensional differences between features. The normalized features are then concatenated using the Concatenate function to enhance the completeness of the feature representation and construct a physically interpretable time-frequency feature tensor as the model input. The details are as follows: S2-2-1: Extract wavelet features: The input current signal is decomposed into multiple layers of discrete wavelets using the Symlet wavelet basis function. The optimal scale is determined by optimizing the number of decomposition layers. The physical statistics of the approximate coefficients and detail coefficients are calculated at each scale, ultimately forming a multi-dimensional time domain feature vector. The decomposition process can be formally expressed as:

[0019]

[0020] In the formula represents the approximate coefficient, represents the detail coefficient, is a low-pass filter, is a high-pass filter.

[0021] S2-2-2: Standardize the features: Z-score normalization is used to regularize the feature space distribution, eliminate the dimensional differences between features at different scales, and facilitate cross-scale feature fusion and model learning. The formula is:

[0022] In the formula represents the normalized features, represents the original features, represents the mean of the feature, Represents the standard deviation of the feature.

[0023] S2-2-3: Feature Fusion: Use the Concatenate function to concatenate the standardized features to enhance the feature representation ability.

[0024]

[0025] In the formula represents the weighted feature tensor, Indicates in Dimensional splicing. A tensor representing the output.

[0026] S2-3: Building physical constraint modules To address the black-box nature of traditional intelligent fault diagnosis (IFD) systems, this study designed a fault diagnosis module with physical constraints. This module embeds domain knowledge into the model training process by constructing a loss function based on physical mechanisms, achieving interpretable physical constraints. In the design of the physical constraint module, frequency-domain energy spectral density analysis is used as the physical representation carrier. The frequency-domain characteristics of the motor current signal are extracted through a sliding window fast Fourier transform (FFT) method. A physical constraint mechanism is then constructed based on the energy distribution characteristics of the fault-sensitive frequency band. The details are as follows: S2-3-1: Quantify the energy of fault-sensitive frequency band: The fault will generate a corresponding fault frequency. The input current signal is used to obtain the spectrum through FFT (1). The spectrum is integrated in the frequency band related to the frequency conversion (2) to extract the characteristic amplitude and construct the frequency domain physical prior knowledge.

[0027]

[0028]

[0029] In the formula N Indicates the total number of sampling points of the time domain signal. Represents the value of the sampling point in the time domain, In the frequency domain k The complex amplitude of the frequency component. A It represents the sum of the amplitudes of each frequency component within a frequency band, reflecting the energy or intensity of the frequency band.

[0030] S2-3-2: Construct physical loss function: The total loss function consists of three parts: cross entropy loss (1), which ensures classification accuracy; spectrum constraint loss (2), which forces the physical channel output to be consistent with the physical prior of the fault category; and physical regularization (3), which prevents overfitting of physical features.

[0031]

[0032] (1) (2) (3) In the formula is the total loss term, is the cross entropy loss term, is the spectrum constraint loss term, is the physical regularization loss term. It is a sample i The true label, is the predicted probability. is the physical feature predicted by the model, is a category yi The known average spectral characteristics of .

[0033] S2-4: Building a dual-channel collaborative optimization fault diagnosis module like Figure 2 As shown in the figure, in order to achieve the deep integration of explainable physical constraints and data-driven, a dual-channel collaborative optimization architecture is designed. This architecture consists of two parallel branches: the diagnostic decision channel and the physical feature analysis channel. The diagnostic decision channel generates the classification probability distribution of the fault type through Softmax; the physical feature analysis channel outputs the energy spectrum feature vector constraint based on physical prior knowledge. The two channels realize feature coupling by sharing the underlying feature extraction layer of the convolutional neural network (CNN), and use the parameter sharing mechanism to establish the gradient propagation path across the channels, thereby guiding the network to meet the accuracy requirements of the diagnostic decision and the constraints of the physical laws during the feature learning process. Among them, SoftMax can be expressed as:

[0034] In the formula Refers to the category of classification, Refers to the output of the classification, Refers to the exponential function. This allows for motor fault diagnosis and classification.

[0035] S2-5: Building a t-SNE feature visualization module t-SNE (t-Distributed Stochastic Neighbor Embedding) uses KL divergence optimization to nonlinearly project high-dimensional feature spaces onto a two-dimensional physically interpretable coordinate space. This method preserves the local topological structure of fault features while revealing the relationship between deep features and the physical mechanisms of the fault. This method achieves physically meaningful projection of high-dimensional feature spaces and verifies the topological consistency of feature distributions and fault patterns through spectral clustering.

[0036] S3: Optimize and design network hyperparameters S3-1: Applying Dynamics to Decay Learning Rate Adjusting the dynamic decay learning rate can improve the stability of model training and effectively avoid model crashes. The dynamic decay learning rate decays exponentially, and as training progresses, the learning rate gradually decreases. The formula is:

[0037] In the formula represents the initial learning efficiency, represents the attenuation factor, step Indicates the cycle length.

[0038] S3-2: Applying Selu activation function to control the distribution of output values The activation function can complete the nonlinear mapping of data, which is a nonlinear transformation. Different from the linear operation model's ability to express features, it can improve the neural network's ability to learn complex data. In the process of backpropagation of errors to update parameters, continuous multiplication of activation function derivatives will appear, which is prone to gradient vanishing and gradient explosion problems. The activation function directly affects the convergence problem in the gradient descent process, and also directly affects the parameter update and the learning efficiency of the network. The Selu activation function effectively controls the distribution of output values ​​by introducing and in deep neural networks, avoiding the situation where the activation value deviates from zero, which is conducive to more stable training and prevents the model from falling into local optimality. The formula and derivative formula of the Selu activation function are as follows:

[0039]

[0040] In the formula represents the scaling factor, Indicates the coefficient for adjusting the negative part.

[0041] S3-3: Apply dropout technology to improve network overfitting like Figure 2 As shown in the figure, during the training process, the neural network may overfit due to excessive number of parameters, resulting in a decrease in network recognition accuracy. In order to avoid overfitting and ensure the speed of model recognition, the Dropout layer is introduced into the network to improve network overfitting. The formula is:

[0042]

[0043] In the formula r where represents a vector of independent Bernoulli random variables, each with probability 1, " " represents element-wise multiplication, Indicates input, Represents sparse input.

[0044] S4: Evaluate model performance In this implementation case, we take 6 sets of sample data, including those without faults and those with different degrees of broken bars (number of broken bars is 1, 3, 5, 7, and 9 respectively). According to step (S2-2), we extract the physical characteristics of each operating state, and according to step (S2-3-2), we construct a loss function that embeds physical prior knowledge. Figure 3The figure below shows the classification accuracy and loss function curves of the model in this case study, demonstrating high classification accuracy. This demonstrates that the physics-guided deep convolutional neural network (PG-DCNN) model has high diagnostic accuracy for motor faults with varying numbers of broken bars, significantly improving the interpretability of fault characteristics and the physical credibility of diagnostic conclusions.

[0045] This invention uses an interpretable intelligent model as a design factor and an interpretable decision-making process as an optimization goal. It constructs an interpretable feature extraction module and a loss function that constructs physical constraints based on physical prior knowledge. Through a dual-channel collaborative optimization architecture, the diagnostic decision channel generates a classification probability distribution of fault types using a softmax function, while the physical feature analysis channel outputs energy spectrum feature vector constraints based on physical prior knowledge. The two channels achieve feature coupling by sharing the underlying convolutional neural network (CNN) feature extraction layer. A parameter sharing mechanism is used to establish a cross-channel gradient propagation path, guiding the network to simultaneously meet the accuracy requirements of diagnostic decisions and the constraints of physical laws during feature learning. This achieves a deep fusion of interpretable physical constraints and data-driven approaches. Ultimately, an interpretable decision-making process is achieved, significantly improving the credibility and traceability of diagnostic results.

[0046] The present invention can adjust physical prior knowledge and physical constraints according to different motor fault types, making diagnostic results more reliable and significantly improving the interpretability of the model. It solves the problem of insufficient reliability caused by the black box characteristics of traditional IDF models.

[0047] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solutions and improved concepts of the present invention within the technical scope disclosed by the present invention, and these changes should be covered by the scope of protection of the present invention.

Claims

1. A motor current interpretable diagnosis method based on a dual-channel collaboratively optimized deep convolutional neural network model, characterized by: The method comprises the following steps: S1: Collect the motor's historical health and current signals during actual operation, and pre-process the collected signals; S2: Construct a network model for motor fault diagnosis; S3: Optimize and design network hyperparameters; S4: Evaluate model performance.

2. The motor current interpretable diagnosis method based on the dual-channel collaborative optimization deep convolutional neural network model according to claim 1 is characterized in that: Step S1 is specifically as follows: S1-1: The acquisition system consists of a high-precision current sensor and a synchronous data acquisition module. The sampling frequency is set to 12800Hz to obtain the current data during the healthy operation of the motor. , the measured current data during actual operation ; S1-2: Preprocess the collected current signal.

3. The motor current interpretable diagnosis method based on the dual-channel collaborative optimization deep convolutional neural network model according to claim 1 is characterized in that: Step S2 is specifically as follows: S2-1: Divide the samples; S2-2: Construct a physically interpretable time-frequency feature tensor as model input; S2-3: Construct physical constraint module; S2-4: Construct a dual-channel collaborative optimization fault diagnosis module; S2-5: Build a t-SNE feature visualization module.

4. The motor current interpretable diagnosis method based on the dual-channel collaborative optimization deep convolutional neural network model according to claim 1 is characterized in that: Step S3 is specifically as follows: S3-1: Apply dynamics to decay learning rate; S3-2: Apply Selu activation function to control the distribution of output values; S3-3: Apply dropout technology to improve network overfitting.

Images