An artificial intelligence-based motor fault early warning method, device and storage medium
By combining frequency domain and time-frequency characteristic analysis of motor operating data and vibration signals, a fault diagnosis model with a dual-branch structure was adopted to solve the problem of misjudging early faults in motors under different operating conditions, thus achieving higher fault diagnosis accuracy and stability.
Patent Information
- Application Number
- CN202610815776.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-08
- Publication Date
- 2026-08-25
AI Technical Summary
Existing motor fault detection methods are prone to misjudging early faults under different operating conditions, resulting in insufficient identification accuracy, especially since it is difficult to distinguish weak features in the early stages of a fault.
By combining motor operating data and analyzing frequency domain and time-frequency features, a fault diagnosis model with a dual-branch structure is adopted. The cross-attention mechanism is used to enhance the features. Combined with the denoising processing of vibration signals, complementary features are extracted and the feature weights are adaptively adjusted.
It improves the separability of early fault characteristics, reduces the probability of misdiagnosis and missed detection, enhances the accuracy and stability of motor fault diagnosis, and improves the reliability of equipment operation and maintenance efficiency.
Smart Images
Figure CN122634306A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of motor fault detection technology, specifically to a motor fault early warning method, device, and storage medium based on artificial intelligence. Background Technology
[0002] With the continuous development of industrial automation and intelligent manufacturing technologies, motors, as core power components in various industrial equipment, are widely used in manufacturing, energy and power, rail transportation, petrochemicals, and metallurgical mining. The operating status of motors directly affects the stability and reliability of production systems. Once a fault occurs, it can not only lead to a decline in equipment performance but also potentially cause production interruptions, product quality fluctuations, and significant economic losses. Therefore, monitoring the operating status of motors and promptly identifying potential faults has become an important aspect of equipment operation and maintenance management.
[0003] Currently, monitoring the operating status of motors mainly relies on operational data such as vibration signals, temperature signals, and current signals. Among these, vibration signals are widely used because they can sensitively reflect abnormal conditions such as bearing wear, rotor imbalance, shaft eccentricity, and mechanical loosening. Existing technologies typically analyze the collected vibration data to extract feature information and determine the motor's operating status accordingly.
[0004] However, motors typically operate under varying speeds, loads, and environmental conditions during practical applications, and these changes significantly impact vibration signal characteristics. Existing fault detection methods primarily focus on analyzing the vibration signal itself, neglecting to utilize operational information reflecting the equipment's running state and failing to fully consider the changing patterns of signal characteristics under different operating conditions. When operating conditions change, the characteristic boundary between normal and fault states can easily shift, leading to fluctuations in detection results. Especially in the early stages of a fault, abnormal characteristics are usually weak; relying solely on vibration signals for state identification can easily misinterpret characteristic fluctuations caused by changes in operating conditions as fault characteristics, resulting in insufficient accuracy in early fault identification. Summary of the Invention
[0005] The purpose of this invention is to solve the problem of accuracy in early fault identification mentioned in the background art, and to propose a motor fault early warning method, device and storage medium based on artificial intelligence.
[0006] A first aspect of this invention provides a motor fault early warning method based on artificial intelligence, the method comprising: Acquire operating condition data and vibration signals of the motor during operation; The vibration signal is denoised to obtain the target vector; The target vector is transformed in the frequency domain to obtain the first frequency domain feature; Perform time-frequency analysis on the target vector to obtain the first time-frequency feature; The operating condition data, the first frequency domain feature, and the first time-frequency feature are input into a pre-trained fault diagnosis model; the fault diagnosis model uses two branches to extract features from the first frequency domain feature and the first time-frequency feature respectively, to obtain a second frequency domain feature and a second time-frequency feature; and enhances the second frequency domain feature and the second time-frequency feature according to the operating condition data; and performs fault diagnosis based on the enhanced features. Obtain the diagnostic results output by the fault diagnosis model; Based on the diagnostic results, an early warning message is generated.
[0007] Optionally, the operating data includes rotational speed, load, and bearing housing temperature.
[0008] Optionally, the denoising process of the vibration signal to obtain the target vector includes: Variational mode decomposition is performed on the vibration signal to obtain multiple modal components; Calculate the sample entropy of each modal component and its correlation coefficient with the vibration signal; Modal components whose sample entropy is less than a preset entropy threshold and whose correlation coefficient is greater than a preset correlation threshold are classified as signal-dominant components, and the remaining modal components are classified as noise-dominant components. Wavelet threshold denoising is applied to the noise-dominant component to obtain the denoised noise-dominant component; The dominant signal component and the denoised noise component are combined and reconstructed to obtain the denoised vibration signal, which is used as the target vector.
[0009] Optionally, the number of decomposition layers and the penalty factor of the variational mode decomposition are determined offline through a metaheuristic optimization algorithm; specifically, the optimization is performed with the goal of minimizing the average envelope entropy of the modal components.
[0010] Optionally, the target vector can be frequency domain transformed using a fast Fourier transform, and the amplitude spectrum can be used as the first frequency domain feature.
[0011] Optionally, the target vector can be subjected to time-frequency analysis using continuous wavelet transform to obtain the first time-frequency feature.
[0012] Optionally, the fault diagnosis model includes a frequency domain feature extraction branch, a time-frequency feature extraction branch, a working condition feature extraction module, a feature enhancement and fusion module, and a classification module; wherein: The frequency domain feature extraction branch uses a one-dimensional convolutional network and a bidirectional long short-term memory network to process the first frequency domain features to obtain the second frequency domain features; The time-frequency feature extraction branch uses a visual Transformer to process the first time-frequency feature to obtain the second time-frequency feature; The operating condition feature extraction module uses a linear layer to map the operating condition data into a high-dimensional operating condition embedding vector; the operating condition embedding vector, the second frequency domain feature, and the second time-frequency feature have the same feature dimension; The feature enhancement and fusion module employs a cross-attention mechanism to enhance the second frequency domain feature and the second time-frequency feature based on the working condition embedding vector, thereby obtaining a third frequency domain feature and a third time-frequency feature; the third frequency domain feature and the third time-frequency feature are then concatenated to obtain a fused feature. The classification module performs global average pooling on the fused features to obtain a global feature vector; the global feature vector is mapped to a probability distribution under each preset state through a fully connected layer, and the state corresponding to the maximum value in the probability distribution is used as the diagnostic result of the motor.
[0013] Optionally, the operation process of the feature enhancement fusion module includes: Different linear layers are used to map the working condition embedding vector to query Q, the second frequency domain feature to key K1 and value V1, and the second time-frequency feature to key K2 and value V2, respectively. We employ a scaled dot product attention mechanism for weighted feature enhancement: ; Among them, FD3 and TF3 are the third frequency domain feature and the third time-frequency feature, respectively; d1 and d2 are the feature dimensions of K1 and K2, respectively; Softmax is a normalization function.
[0014] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based motor fault early warning method.
[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described artificial intelligence-based motor fault early warning method.
[0016] The beneficial effects of this invention are: 1. By introducing motor operating condition data and combining it with the frequency domain and time-frequency characteristics of vibration signals for joint analysis, the model can distinguish between changes in operating conditions and actual fault characteristics. Simultaneously, a dual-branch structure is used to extract complementary features, and operating condition information is used to adaptively enhance the features, reducing feature shifts caused by different operating conditions. This method effectively improves the separability of early, weak fault features, reduces the probability of misjudgment and missed detection, and improves the accuracy, stability, and timeliness of motor fault diagnosis, thereby enhancing equipment reliability and maintenance efficiency.
[0017] 2. When denoising vibration signals, the modal components are divided into signal-dominant components and noise-dominant components. Only the noise-dominant components are de-denoised by wavelet thresholding and then reconstructed. This can efficiently remove environmental and structural noise while preserving key fault information, thereby improving the signal-to-noise ratio and fault characterization capabilities of subsequent feature extraction.
[0018] 3. A cross-attention mechanism was introduced into the fault diagnosis model. The frequency domain features and time-frequency features were adaptively enhanced by the operating condition embedding vector, which enabled the model to dynamically adjust the feature weights according to the current operating conditions, thereby improving the model's generalization ability and diagnostic stability under different operating conditions. Attached Figure Description
[0019] Figure 1 A flowchart illustrating an artificial intelligence-based motor fault early warning method provided in an embodiment of the present invention; Figure 2 This is a network architecture diagram of a fault diagnosis model provided in an embodiment of the present invention. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.
[0021] This invention provides a method for early warning of motor faults based on artificial intelligence. See also... Figure 1 The method includes the following steps: S101 acquires operating condition data and vibration signals of the motor during operation.
[0022] S102, the vibration signal is denoised to obtain the target vector.
[0023] S103, perform frequency domain transformation on the target vector to obtain the first frequency domain feature.
[0024] S104, Perform time-frequency analysis on the target vector to obtain the first time-frequency feature.
[0025] S105, input the operating condition data, the first frequency domain feature and the first time-frequency feature into the pre-trained fault diagnosis model.
[0026] S106, Obtain the diagnostic results output by the fault diagnosis model.
[0027] S107, Based on the diagnostic results, generate early warning information.
[0028] The fault diagnosis model uses two branches to extract features from the first frequency domain feature and the first time-frequency feature respectively, to obtain the second frequency domain feature and the second time-frequency feature. The second frequency domain feature and the second time-frequency feature are enhanced based on the operating condition data. Fault diagnosis is performed based on the enhanced features, and the diagnosis result is output.
[0029] In one implementation, the operating data includes rotational speed, load, and bearing housing temperature. Rotational speed and load are core operating parameters that determine the mechanical stress and electromagnetic state of the motor, while bearing housing temperature directly reflects frictional losses and heat dissipation conditions. Using these three parameters in combination allows for the acquisition of operating information strongly correlated with vibration signal changes at a relatively low cost, providing effective prior input for subsequent feature enhancement.
[0030] In one embodiment, step S102, denoising the vibration signal to obtain the target vector, includes: Step 1: Perform variational mode decomposition on the vibration signal to obtain multiple modal components.
[0031] Step two: Calculate the sample entropy of each modal component and its correlation coefficient with the vibration signal. The Pearson correlation coefficient is used for the correlation coefficient.
[0032] Step 3: Classify the modal components whose sample entropy is less than a preset entropy threshold and whose correlation coefficient is greater than a preset correlation threshold as signal dominant components, and classify the remaining modal components as noise dominant components.
[0033] Step 4: Perform wavelet threshold denoising on the noise dominant component to obtain the denoised noise dominant component.
[0034] Step 5: Combine and reconstruct the dominant signal component and the denoised noise component to obtain the denoised vibration signal, which serves as the target vector.
[0035] In one implementation, a minimum-maximum normalization algorithm is used to normalize the sample entropy and similarity coefficient of each modal component. The maximum and minimum values of the current modal component are used to map the sample entropy range to the [0, 1] interval. Then, the normalized values are compared with preset thresholds. The entropy threshold and correlation threshold are determined by analyzing the modal component distribution of historical data. For example, the entropy threshold can be set to a value between 0.6 and 0.8, and the correlation threshold can be set to a value between 0.3 and 0.6. Preferably, the entropy threshold is 0.6, and the correlation threshold is 0.5.
[0036] This embodiment decomposes the original vibration signal into multiple modal components through variational mode decomposition. Sample entropy is then used to measure the regularity of each modal component, and correlation coefficients are used to measure the degree of correlation between each modal component and the original vibration signal, thereby achieving automatic identification and classification of the dominant signal and noise components. Subsequently, wavelet threshold denoising is performed only on the noise-dominant component, and it is reconstructed together with the dominant signal component to obtain the target vector. This effectively suppresses random noise and environmental interference while retaining fault impact characteristics and other useful information, avoiding excessive weakening of weak fault features by traditional overall denoising methods. Therefore, the signal-to-noise ratio and feature integrity of the target vector can be improved, providing a more reliable data foundation for subsequent frequency domain feature extraction, time-frequency feature extraction, and fault diagnosis, thereby improving the accuracy of fault early warning.
[0037] In one implementation, the number of decomposition levels and the penalty factor of the variational mode decomposition are determined offline using a metaheuristic optimization algorithm, such as particle swarm optimization or whale optimization. Specifically, the optimization aims to minimize the average envelope entropy of the modal components.
[0038] The number of decomposition layers and the penalty factor have a significant impact on the variational mode decomposition (VMD) effect. Empirical methods of setting these values are insufficient to adapt to the vibration signal characteristics of different motor models, operating conditions, and acquisition environments. By optimizing the number of decomposition layers and the penalty factor with the goal of minimizing the average envelope entropy of the modal components, the frequency band division of each modal component can be made more reasonable, reducing mode aliasing and improving the separation ability of effective features from noise components. This enhances the adaptability and automation of the vibration signal denoising process, providing a more reliable data foundation for subsequent fault feature extraction and fault diagnosis.
[0039] In one implementation, a Fast Fourier Transform (FFT) is used to transform the target vector in the frequency domain, with the amplitude spectrum serving as the first frequency domain feature. The FFT is computationally efficient, converting time-domain vibration signals into frequency-domain signals, and the amplitude spectrum directly reflects the energy distribution of the vibration signal at different frequencies. For common motor faults such as bearing wear, rotor imbalance, and gear failure, their corresponding characteristic frequencies and harmonic components can be effectively characterized in the amplitude spectrum, thus highlighting the fault-related frequency features and providing effective input for subsequent deep learning models to extract frequency-domain fault features and perform fault identification.
[0040] In one implementation, continuous wavelet transform is used to perform time-frequency analysis on the target vector to obtain the first time-frequency feature. Continuous wavelet transform can analyze signals simultaneously in both time and frequency dimensions, and has a strong ability to characterize non-stationary features such as transient impacts and amplitude modulation / frequency modulation in vibration signals. Using the time-frequency information output by continuous wavelet transform as the first time-frequency feature can compensate for the deficiency of pure frequency domain analysis in reflecting the time-varying laws of features, and is beneficial for capturing short-term, weak abnormal features that appear in the early stages of a fault, thereby improving the fault feature extraction effect and the accuracy of fault diagnosis.
[0041] In one embodiment, see Figure 2 In the figure, FD1, TF1, and OCD represent the first frequency domain feature, the first time-frequency feature, and the operating condition data, respectively. The fault diagnosis model includes a frequency domain feature extraction branch, a time-frequency feature extraction branch, an operating condition feature extraction module, a feature enhancement and fusion module, and a classification module; among which: The frequency domain feature extraction branch uses a one-dimensional convolutional network and a bidirectional long short-term memory network (1D CNN+BiLSTM) to process the first frequency domain features and obtain the second frequency domain features.
[0042] The time-frequency feature extraction branch uses a visual Transformer (ViT) to process the first time-frequency feature to obtain the second time-frequency feature.
[0043] The working condition feature extraction module uses a linear layer to map the working condition data into a high-dimensional working condition embedding vector; the working condition embedding vector, the second frequency domain feature, and the second time-frequency feature have the same feature dimension.
[0044] The feature enhancement and fusion module adopts a cross-attention mechanism to enhance the second frequency domain feature and the second time-frequency feature according to the working condition embedding vector, respectively, to obtain the third frequency domain feature and the third time-frequency feature; the third frequency domain feature and the third time-frequency feature are concatenated to obtain the fused feature.
[0045] The classification module performs global average pooling on the fused features to obtain a global feature vector. The global feature vector is then mapped to a probability distribution under each preset state through a fully connected layer. The state corresponding to the maximum value in the probability distribution is used as the diagnostic result of the motor.
[0046] 1D CNN+BiLSTM can simultaneously handle local spectral feature extraction and global frequency correlation modeling; the visual Transformer can fully exploit global time-frequency correlation features in the time-frequency plot. By adopting adapted network structures for different features, the representation ability of different types of fault features can be improved, providing more accurate and richer feature information for subsequent fault diagnosis. Furthermore, operating condition information is used to enhance and fuse the two types of features, enabling the model to simultaneously utilize frequency information, time-frequency evolution information, and operating condition information for fault analysis. This highlights fault features related to the current operating state, reduces interference from changes in operating conditions, and improves the accuracy, robustness, and early fault identification capability of motor fault diagnosis under complex operating conditions.
[0047] In one implementation, the computation process of the feature enhancement fusion module includes: Different linear layers are used to map the working condition embedding vector to query Q, the second frequency domain features to key K1 and value V1, and the second time-frequency features to key K2 and value V2, respectively; a scaled dot product attention mechanism is used to weighted enhance the features. ; Among them, FD3 and TF3 are the third frequency domain feature and the third time-frequency feature, respectively; d1 and d2 are the feature dimensions of K1 and K2, respectively; Softmax is a normalization function.
[0048] This embodiment uses a cross-attention mechanism, with the operating condition embedding vector as the query and frequency domain features and time-frequency features as key-value pairs, to establish the correlation between operating condition information and vibration characteristics. This allows the model to adaptively adjust the importance of different features based on the current operating state, highlighting fault-related features and suppressing interference caused by changes in operating conditions. Simultaneously, weighted enhancement of frequency domain features and time-frequency features helps to fully extract effective information from both types of features, improving the discriminative power of the fused features and thus enhancing the accuracy and robustness of motor fault diagnosis under complex operating conditions.
[0049] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the artificial intelligence-based motor fault early warning methods described in the above embodiments.
[0050] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the artificial intelligence-based motor fault early warning methods described in the above embodiments.
[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.
Claims
1. A method for early warning of motor faults based on artificial intelligence, characterized in that, The method includes: Acquire operating condition data and vibration signals of the motor during operation; The vibration signal is denoised to obtain the target vector; The target vector is transformed in the frequency domain to obtain the first frequency domain feature; Perform time-frequency analysis on the target vector to obtain the first time-frequency feature; The operating condition data, the first frequency domain feature, and the first time-frequency feature are input into a pre-trained fault diagnosis model; the fault diagnosis model uses two branches to extract features from the first frequency domain feature and the first time-frequency feature respectively, to obtain a second frequency domain feature and a second time-frequency feature; and enhances the second frequency domain feature and the second time-frequency feature according to the operating condition data; and performs fault diagnosis based on the enhanced features. Obtain the diagnostic results output by the fault diagnosis model; Based on the diagnostic results, an early warning message is generated.
2. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The operating data includes speed, load, and bearing housing temperature.
3. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The denoising process of the vibration signal to obtain the target vector includes: Variational mode decomposition is performed on the vibration signal to obtain multiple modal components; Calculate the sample entropy of each modal component and its correlation coefficient with the vibration signal; Modal components whose sample entropy is less than a preset entropy threshold and whose correlation coefficient is greater than a preset correlation threshold are classified as signal-dominant components, and the remaining modal components are classified as noise-dominant components. Wavelet threshold denoising is applied to the noise-dominant component to obtain the denoised noise-dominant component; The dominant signal component and the denoised noise component are combined and reconstructed to obtain the denoised vibration signal, which is used as the target vector.
4. The method for early warning of motor faults based on artificial intelligence according to claim 3, characterized in that, The number of decomposition levels and the penalty factor of the variational mode decomposition are determined offline through a metaheuristic optimization algorithm; specifically, the optimization aims to minimize the average envelope entropy of the modal components.
5. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The target vector is transformed in the frequency domain using a fast Fourier transform, and the amplitude spectrum is used as the first frequency domain feature.
6. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The target vector is subjected to time-frequency analysis using continuous wavelet transform to obtain the first time-frequency feature.
7. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The fault diagnosis model includes a frequency domain feature extraction branch, a time-frequency feature extraction branch, a working condition feature extraction module, a feature enhancement and fusion module, and a classification module; wherein: The frequency domain feature extraction branch uses a one-dimensional convolutional network and a bidirectional long short-term memory network to process the first frequency domain features to obtain the second frequency domain features; The time-frequency feature extraction branch uses a visual Transformer to process the first time-frequency feature to obtain the second time-frequency feature; The operating condition feature extraction module uses a linear layer to map the operating condition data into a high-dimensional operating condition embedding vector; the operating condition embedding vector, the second frequency domain feature, and the second time-frequency feature have the same feature dimension; The feature fusion module employs a cross-attention mechanism to enhance the second frequency domain feature and the second time-frequency feature based on the working condition embedding vector, thereby obtaining a third frequency domain feature and a third time-frequency feature; the third frequency domain feature and the third time-frequency feature are then concatenated to obtain the fused feature. The classification module performs global average pooling on the fused features to obtain a global feature vector; the global feature vector is mapped to a probability distribution under each preset state through a fully connected layer, and the state corresponding to the maximum value in the probability distribution is used as the diagnostic result of the motor.
8. The method for early warning of motor faults based on artificial intelligence according to claim 1, characterized in that, The operation process of the feature enhancement fusion module includes: Different linear layers are used to map the working condition embedding vector to query Q, the second frequency domain feature to key K1 and value V1, and the second time-frequency feature to key K2 and value V2, respectively. We employ a scaled dot product attention mechanism for weighted feature enhancement: ； Among them, FD3 and TF3 are the third frequency domain feature and the third time-frequency feature, respectively; d1 and d2 are the feature dimensions of K1 and K2, respectively; Softmax is a normalization function.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.