VMD-CWT-Transformer-based sound and vibration fusion fault diagnosis method for industrial rotating machinery
By using the cascaded processing method of VMD-CWT-Transformer, the problem of fault diagnosis under complex working conditions and data imbalance in traditional methods is solved. It achieves high-accuracy fault identification and health assessment, is highly adaptable, and supports multi-task expansion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU SOUND TECH TECH CO LTD
- Filing Date
- 2025-12-19
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional fault diagnosis methods for industrial rotating machinery have weak generalization ability when faced with complex operating conditions and data imbalance, making it difficult to adapt to changes in speed and load. Furthermore, the scarcity of fault samples leads to poor diagnostic results.
The VMD-CWT-Transformer cascaded processing method is adopted, which combines variational mode decomposition, continuous wavelet transform and Transformer network to achieve adaptive signal decomposition, multi-scale time-frequency representation and cross-modal deep fusion. Fault classification and health assessment are performed through multilayer perceptron.
It improves the accuracy of fault identification under complex operating conditions, enhances the robustness and adaptability of the model, maintains the stability of diagnostic results under strong background noise and variable speed and load conditions, and supports multi-task extensions such as fault category identification, health assessment and life prediction.
Smart Images
Figure CN122045910A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial machinery and equipment anomaly detection technology, and in particular, it is a method for fusion of acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer. Background Technology
[0002] Large industrial rotating machinery, as core production equipment in fields such as power, chemical, and metallurgy (e.g., 1000kW motors, 10,000 cubic meter centrifugal fans), directly determines production line capacity and safety through its continuous operation. According to data from the Manufacturing Equipment Management White Paper, unplanned equipment downtime accounts for 8%-15% of an enterprise's annual output value, with downtime caused by failures in key components such as bearings and rotors accounting for over 60%. With the advancement of Industry 4.0 and intelligent manufacturing, the traditional "post-incident maintenance" model can no longer meet the modern industrial demands for "cost reduction, efficiency improvement, and controllable safety." Large industrial rotating machinery is the "heart" of industrial production, and its operational stability directly determines production efficiency and safety levels.
[0003] Currently, the industrial sector mainly relies on the following methods for diagnosing faults in rotating machinery, which have obvious technical shortcomings.
[0004] 1. Reliance on traditional signal processing methods and expert experience, resulting in weak generalization ability: Traditional methods are mostly based on manually designed signal features (such as time-domain RMS, kurtosis, peak factor, frequency-domain gear side frequencies, envelope spectrum features, etc.), and the feature extraction process depends on expert knowledge and experience rules. Once the operating environment changes (changes in speed, load, lubrication status), the feature extraction effect drops significantly, making it difficult to adapt to complex working conditions and new equipment.
[0005] 2. Poor adaptability to different operating conditions and difficulty in cross-equipment adaptation: Industrial sites often present challenges such as speed fluctuations, coupled operating conditions, noise interference, and inconsistent sensor placement. Models are often trained for single-machine types and single operating conditions, and need to be remodeled or recalibrated when transferred to other equipment, lacking robustness.
[0006] 3. Fault samples are scarce and the data is unbalanced: The probability of real faults is low, dangerous and uncontrollable, and a large amount of data comes from normal operating conditions.
[0007] Traditional classification models require a rich number of fault samples to generalize, but in real factories, there are often: many normal samples and few fault samples; incomplete fault types and lack of labels; and the inability to label early and minor faults, making it difficult to train supervised diagnostic models. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art and propose a fault diagnosis method for acoustic and vibration fusion of industrial rotating machinery based on VMD-CWT-Transformer. Through the cascaded processing of VMD-CWT-Transformer, it realizes adaptive signal decomposition, multi-scale time-frequency characterization, and cross-modal deep fusion. Combined with a switchable output structure, it supports various industrial application needs such as fault classification, health assessment, and life prediction.
[0009] The technical problem solved by this invention is achieved through the following technical solution: A method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer includes the following steps: Step 1: Synchronously acquire vibration and sound signals from rotating machinery and perform preprocessing; Step 2: Perform variational mode decomposition on the preprocessed signal to obtain the intrinsic mode function components; Step 3: Perform continuous wavelet transform on each intrinsic mode function component to obtain multi-scale time-frequency features; Step 4: Input the multi-scale time-frequency features into the Transformer network for acoustic-vibration fusion and deep feature encoding to obtain the fused representation vector; Step 5: Based on the task requirements, input the fused representation vector into the multilayer perceptron classification head or regression head, and output the fault diagnosis or health assessment results.
[0010] Furthermore, the synchronous acquisition includes the following steps: Step 1.1: Install an accelerometer at the mechanical housing or bearing seat to collect vibration signals; Step 1.2: Deploy condenser or MEMS microphones to collect sound signals; Step 1.3: Synchronously sample the vibration signal a(t) and the sound signal s(t) according to a unified clock.
[0011] Furthermore, the preprocessing in step 1 includes bandpass filtering and signal segmentation.
[0012] Moreover, in step 2, the number K of intrinsic mode function components in variational mode decomposition is adaptively determined based on the signal characteristics.
[0013] Moreover, in step 3, the continuous wavelet transform uses the Morlet wavelet as the mother wavelet.
[0014] Furthermore, in step 4, the Transformer network includes an encoder and a decoder structure, and vibration and sound features are respectively input into the encoder and decoder for fusion.
[0015] Furthermore, in step 5, the multilayer perceptron classification head is used to output the probability distribution of fault categories, and the loss function is cross-entropy or Focal Loss; the multilayer perceptron regression head is used to output the health index, fault degree estimate or remaining life prediction value, and the loss function is mean squared error or mean absolute error.
[0016] Furthermore, step 5 of the Transformer network also includes the use of data augmentation strategies and regularization mechanisms during the training phase to improve the robustness of the model under imbalanced samples and changing operating conditions.
[0017] The advantages and positive effects of this invention are: 1. High Diagnostic Accuracy Through Deep Multidimensional Feature Fusion: This invention overcomes the limitations of single-sensor information by fusing acoustic and vibration signals. Vibration signals directly reflect the rigid connection state of internal mechanical structures, while acoustic signals capture non-contact, airborne fault information. Deep complementarity between the two modes is achieved through the attention mechanism of the Transformer network, significantly improving the accuracy of fault identification under complex operating conditions.
[0018] 2. Superior time-frequency analysis capability and signal decoupling performance: Variational mode decomposition (VMD) effectively overcomes the mode aliasing problem of traditional decomposition methods (such as EMD), and can adaptively extract stationary components in non-stationary signals; combined with the multi-scale characteristics of continuous wavelet transform (CWT), it further extracts extremely subtle local time-frequency features, providing higher quality feature inputs for subsequent deep neural networks.
[0019] 3. The powerful long-range modeling and global feature capture capabilities of the Transformer model: Compared with traditional CNN or RNN, the Transformer network used in this invention utilizes the self-attention mechanism, which can more effectively capture the temporal correlation and cross-modal intrinsic correlation of acoustic and vibration signals in long sequences, greatly enhancing the model's ability to encode deep features of complex fault modes.
[0020] 4. It exhibits strong adaptability and robustness under various operating conditions. Through data augmentation strategies and regularization mechanisms, it addresses the problem of sample imbalance (few faulty samples and many normal samples) in real-world industrial environments. Furthermore, the VMD-CWT preprocessing combination possesses inherent noise reduction characteristics, enabling the model to maintain the stability of diagnostic results even under harsh industrial environments with strong background noise, variable speed, and variable load.
[0021] 5. Possesses multi-task scalability and wide applicability. This solution is not limited to fault category identification. By flexibly configuring the classification head or regression head of the multilayer sensor, it can simultaneously realize fault severity classification, equipment health index (HI) assessment, and remaining useful life (RUL) prediction, providing comprehensive decision support for enterprises' predictive maintenance strategies. Attached Figure Description
[0022] Figure 1 This is a basic structural diagram of the Transformer model of this invention. Detailed Implementation
[0023] The present invention will be further described in detail below with reference to the accompanying drawings.
[0024] A method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer includes the following steps: Step 1: Synchronously collect vibration and sound signals from rotating machinery and perform preprocessing.
[0025] Step 1.1: Install an accelerometer at the mechanical housing or bearing seat to collect vibration signals; Step 1.2: Deploy condenser or MEMS microphones to collect sound signals; Step 1.3: Synchronously sample the vibration signal a(t) and the sound signal s(t) according to a unified clock. The sampling rate is set to 20kHz-64kHz, which can be adjusted according to the equipment speed.
[0026] Step 1.4: Perform bandpass filtering (e.g., 0.5Hz-10kHz) on the sampled signal to remove power frequency noise and background interference; Step 1.5: Divide the signal into segments of length N and construct a sample sequence using a sliding window.
[0027] Step 2: Perform variational mode decomposition on the preprocessed signal to obtain the intrinsic mode function components.
[0028]
[0029] in, Represents the k-th IMF The instantaneous amplitude, represent The phase function has the property of being non-decreasing. The instantaneous frequency is The expression is shown in (2).
[0030]
[0031] because and The rate of change relative to the phase function It is relatively slow, therefore, within the time interval [t - δ, t + δ], It can be considered as a harmonic signal, and its instantaneous frequency is .
[0032] Step 3: Perform continuous wavelet transform on each intrinsic mode function component to obtain multi-scale time-frequency features.
[0033] The Morlet wavelet is selected as the mother wavelet to construct a continuous wavelet basis. Then, wavelet transforms are performed on the center frequencies of each mode extracted in step 2. The main process is shown in equation (3). (3) Then, the wavelet coefficients of all K IMFs are spliced together according to the scale dimension or the IMF dimension. The wavelet coefficients present a structural signal in which the amplitude changes with the oscillation and the scale changes with the fault characteristics. It can be directly used as a sequence token without the need to construct a time-frequency diagram, thereby reducing the amount of computation and improving real-time performance.
[0034] Step 4: Input the multi-scale time-frequency features into the Transformer network for acoustic-vibration fusion and deep feature encoding to obtain the fused representation vector.
[0035] In the core model, a Transformer is used as the base model, employing its encoder-decoder structure. The modal waveforms extracted in steps 2 and 3, along with the wavelet transform results of the center frequency wavelets of each modality, are input into the Transformer. The model structure is as follows: Figure 1 As shown.
[0036] Each of the two modal signals has a Class Token added to it, which is combined with the position code and then input into the encoder and decoder structure of the VMD-CWT-Transformer respectively:
[0037] Step 5: Based on the task requirements, input the fused representation vector into the multilayer perceptron classification head or regression head, and output the fault diagnosis or health assessment results.
[0038] Finally, the fused representation obtained by the decoder is fed into the MLP classification head or regression head. The decoding end or the last layer encoding output of the VMD-CWT-Transformer network will yield a global fused representation vector (FFF) containing the mechanical operating state pattern. To adapt to different industrial scenarios, this invention designs a switchable multi-task output structure, using the MLP classification head or MLP regression head for prediction output, enabling the system to simultaneously possess capabilities such as fault identification, fault severity assessment, health index calculation, and remaining life prediction. Specifically, it includes the following: I. Acquisition of Fusion Characterization The unified dimensional embedding is obtained from the output of the fusion layer through splicing, projection, or compression: This vector, as a high-dimensional vector representation of the fault health state, contains cross-modal correlation features, oscillation energy features, and multi-scale wavelet coefficient structure information, and can be directly used for downstream prediction.
[0039] II. MLP Classification Header (Discrete Output) for Fault Diagnosis Suitable for tasks such as fault category identification, fault location determination, working status identification, and multi-type anomaly detection. Construct a multilayer perceptron consisting of several fully connected layers and activation functions. The output is a probability distribution vector, with dimension C being the number of categories. Loss functions can be selected from: Cross Entropy, Focal Loss to handle imbalanced samples, and Sigmoid + BCELoss for multi-label tasks. III. MLP Regression Header for Health Assessment (Continuous Value Output) Applicable to scenarios such as Remaining Life Prediction (RUL), Failure Level Estimation, Degradation Trend Scoring, and Health Index (HI) Output: The output is a continuous real number or an interval value, for example: 0–100% — Remaining Life Percentage; μm / mm — Crack Size Estimation. Loss functions can be: MSE, MAE, SmoothL1. The regression head can be extended into a trend model to achieve online health monitoring and dynamic remaining life prediction.
[0040] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.
Claims
1. A method for fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer acoustic-vibration fusion, characterized in that, Includes the following steps: Step 1: Synchronously acquire vibration and sound signals from rotating machinery and perform preprocessing; Step 2: Perform variational mode decomposition on the preprocessed signal to obtain the intrinsic mode function components; Step 3: Perform continuous wavelet transform on each intrinsic mode function component to obtain multi-scale time-frequency features; Step 4: Input the multi-scale time-frequency features into the Transformer network for acoustic-vibration fusion and deep feature encoding to obtain the fused representation vector; Step 5: Based on the task requirements, input the fused representation vector into the multilayer perceptron classification head or regression head, and output the fault diagnosis or health assessment results.
2. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, The synchronous acquisition includes the following steps: Step 1.1: Install an accelerometer at the mechanical housing or bearing seat to collect vibration signals; Step 1.2: Deploy condenser or MEMS microphones to collect sound signals; Step 1.3: Synchronously sample the vibration signal a(t) and the sound signal s(t) according to a unified clock.
3. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, The preprocessing in step 1 includes bandpass filtering and signal segmentation.
4. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, In step 2, the number K of intrinsic mode function components in variational mode decomposition is adaptively determined based on the signal characteristics.
5. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, In step 3, the continuous wavelet transform uses the Morlet wavelet as the mother wavelet.
6. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, In step 4, the Transformer network includes an encoder and a decoder structure, and vibration and sound features are respectively input into the encoder and decoder for fusion.
7. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, In step 5, the multilayer perceptron classification head is used to output the probability distribution of fault categories, and the loss function is cross-entropy or Focal Loss; the multilayer perceptron regression head is used to output the health index, fault degree estimate or remaining life prediction value, and the loss function is mean squared error or mean absolute error.
8. The method for fusion-based acoustic and vibration fault diagnosis of industrial rotating machinery based on VMD-CWT-Transformer according to claim 1, characterized in that, Step 5 of the Transformer network also includes the use of data augmentation strategies and regularization mechanisms during the training phase to improve the robustness of the model under imbalanced samples and changing operating conditions.