Scroll compressor hidden defect detection method based on sound and vibration complementary enhancement fusion

Abstract

The invention discloses a method for detecting hidden defects of a scroll compressor based on sound and vibration complementary enhancement fusion, and belongs to the technical field of mechanical equipment fault diagnosis. Aiming at the problems that the existing vibration signal detection method is low in signal-to-noise ratio and the characteristics are easy to submerge in a complex industrial environment, the invention provides a sound-vibration dual-mode fusion detection mechanism. The method comprises the following steps: firstly, synchronously collecting sound and multi-axis vibration signals when a compressor runs, decomposing the signals into multi-band modal components through adaptive variational mode decomposition, and effectively inhibiting various noise interferences by combining an automatic modal number selection and wavelet threshold denoising method; secondly, constructing a double-branch convolutional neural network, respectively learning deep features of sound and vibration signals, and fusing the deep features in a high-level feature space; and finally, accurate identification and classification of the hidden defects are realized through the classification network. The method gets rid of dependence on artificial feature engineering, can still maintain high detection precision and robustness under variable working conditions, and improves the engineering applicability of compressor hidden defect detection.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis technology, specifically relating to a method for detecting latent defects in scroll compressors based on acoustic-vibration complementary enhancement fusion. Background Technology

[0002] The thermal management system is the core component ensuring the safe operation of new energy vehicles, and the scroll compressor, as the "heart" of this system, directly affects the safety of the entire vehicle. Therefore, it is crucial to conduct precise defect detection before the vehicle leaves the factory.

[0003] Traditional testing relies mainly on human experience, making subjective judgments by looking and listening, which has problems such as large errors and poor stability.

[0004] With technological advancements, detection methods have gradually become more scientific and systematic, incorporating signal processing and sensor technologies to achieve vibration signal-based analysis and detection, thus improving controllability. However, this method still requires complex data preprocessing and expert experience to extract features, limiting its detection efficiency and accuracy, and making end-to-end automated detection difficult.

[0005] In recent years, artificial intelligence technologies such as machine learning have revolutionized this field, enabling automatic defect identification and classification, reducing reliance on human experience, and simplifying preprocessing while improving detection speed and accuracy. Currently, mainstream methods are based on vibration signal analysis. However, the complexities of actual production environments and severe background interference lead to low signal-to-noise ratios in vibration signals and easily obscured features, thus limiting the robustness of the models.

[0006] To overcome the limitations of single signals, those skilled in the art have begun to explore multimodal detection methods that fuse sound and vibration signals, aiming to enhance the perception of subtle defects by leveraging complementary information. However, most existing fusion methods still rely on manually designed features, making it difficult to accurately capture the essential information of defects. Meanwhile, the complex noise and variable operating conditions (such as changes in rotational speed and load) in industrial environments severely contaminate sound signals and cause variations in signal distribution, resulting in persistent problems such as poor feature stability, weak model adaptability across different scenarios, and high false positive rates. Summary of the Invention

[0007] In view of the above-mentioned problems in the prior art, the purpose of this invention is to provide a method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion.

[0008] This invention provides the following technical solution: a method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion. This method employs automatic modal number selection and intelligent wavelet type selection. First, the time-domain sound and vibration signals are decomposed into modal components of different frequency bands. Noise reduction is then applied to different frequency bands to obtain richer information. A convolutional neural network is then used to extract more obvious defect features from the sound and vibration signals. Combining the features extracted from the vibration and sound signals, the method distinguishes between normal and defective machines and identifies the defect type, thereby improving the accuracy of the diagnostic model. The specific steps are as follows: S1. A multimodal data acquisition system for defect identification of scroll compressors is constructed. The system includes a triaxial accelerometer and a microphone. The two synchronously acquire the casing vibration signal and radiated sound signal of the compressor under actual production conditions under a unified time reference, thereby forming time-aligned vibration-sound raw data, which provides a basis for subsequent cross-modal collaborative analysis.

[0009] S2. The collected vibration and sound raw signals are preprocessed. Through operation status discrimination and signal stability analysis, data segments of the compressor stable operation stage are automatically extracted, and vibration-sound raw datasets are constructed according to different defect conditions to avoid the uncertainty caused by manual interception.

[0010] S3. For the preprocessed vibration and sound signals, variational mode decomposition is performed in the range of 2 to 15 decomposition modes. The correlation coefficient and signal-to-noise ratio are calculated for the reconstructed signals under each candidate mode number, and a comprehensive decision criterion based on multiple evaluation indicators is constructed to realize the automatic optimization selection of the decomposition mode number K, so that different types of signals can obtain the optimal time-frequency representation.

[0011] S4. Divide each IMF component from the original signal into three groups according to its frequency band: low frequency, mid frequency, and high frequency. Use wavelet thresholding to denoise the signals in different groups using different wavelets. By calculating the kurtosis, spectral entropy, and energy retention rate after denoising, select the appropriate optimal wavelet and finally obtain the corresponding parameter configuration table.

[0012] For different inputs, the optimal wavelet is selected by calculating kurtosis, spectral entropy and energy retention rate, and noise reduction is performed on sound and vibration signals respectively.

[0013] S5. Construct a dual-branch convolutional neural network model with significantly enhanced structure. By dividing the intrinsic mode functions obtained from variational mode decomposition into low, medium and high frequency subsets according to frequency, feature extraction is performed using a multi-scale one-dimensional convolutional network with channel attention. Vibration and sound features are then fused in a high-level feature space to achieve high-precision identification of mechanical equipment faults under complex working conditions.

[0014] Furthermore, the specific process of step S3 is as follows: S31. Traverse the preset data directory, including normal state and at least one abnormal state, read the vibration acceleration signals of the X, Y and Z axes from each Excel file, and set the VMD basic parameters (bandwidth constraint parameters, noise tolerance, initialization method, convergence tolerance).

[0015] S32. Set the search range for the mode number K to 2 to 15, covering various cases from simple decomposition to fine decomposition. This range is determined based on the following considerations: K=2-3: suitable for cases with simple signal components; K=4-8: suitable for most rotating machinery vibration signals; K=9-15: suitable for fine decomposition under complex defect conditions.

[0016] S33. The selection of the optimal K value is based on two key quality indicators: the reconstructed correlation coefficient. With signal-to-noise ratio For each K value, the quality index is obtained by minimum-maximum normalization. and The score is calculated using an experience-weighted comprehensive scoring formula: , Choose the value with the highest comprehensive score from K=2-15 as the optimal number of modes for the signal.

[0017] S34. After determining the optimal K, perform variational mode decomposition, sort the decomposed components in order of frequency band from low to high, and finally store the data.

[0018] Furthermore, the specific process of step S4 is as follows: S41. By analyzing the physical characteristics and frequency features of vibration and sound signals, a differentiated parameter system is established. Each intrinsic mode function (IMF) component obtained by variational mode decomposition is divided into three groups according to its center frequency from low to high: IMF0-1 is the low-frequency band, IMF2-4 is the mid-frequency band, and IMF5-7 is the high-frequency band.

[0019] S42. Based on the characteristics of vibration signals and sound signals, establish wavelet candidate pools for vibration signals and sound signals respectively.

[0020] S43. Apply corresponding threshold scaling factors to different frequency bands to avoid erasing defect impact characteristics.

[0021] S44. Select candidate wavelets one by one from the corresponding wavelet candidate pool for experimental denoising. Calculate the kurtosis, spectral entropy, and energy retention rate of the signals before and after denoising to determine the optimal wavelet for the IMF in this frequency band.

[0022] S45. After selecting the optimal wavelet, perform wavelet decomposition on each modal component, and then store the data.

[0023] Furthermore, the specific process of step S5 is as follows: S51. The multi-axis signal obtained from VMD decomposition is reconstructed at the channel level with the IMF dimension to form a multi-channel temporal input tensor of "axis × IMF". As before, according to the frequency distribution characteristics of the IMF, it is divided into three frequency band subsets: low frequency, mid frequency and high frequency, to provide a structural basis for subsequent differential feature extraction.

[0024] S52. For each IMF frequency band subset, construct an independent convolutional subnetwork. In each subnetwork, adopt a multi-scale one-dimensional convolutional structure, extract local and long-term temporal features in parallel through different convolutional kernel sizes, and perform channel splicing and pooling processing on the multi-scale convolution results to achieve feature fusion across time scales.

[0025] S53. An IMF channel attention module is introduced after the multi-scale convolution output. Through global temporal statistics and nonlinear mapping, the importance weights of different IMF channels in fault identification are automatically learned, and the feature channels are adaptively recalibrated to suppress redundant IMFs and enhance key fault modes.

[0026] S54. Low-frequency, mid-frequency, and high-frequency IMF subsets are learned using independent network branches. Each frequency band branch is spliced ​​and fused in a high-dimensional feature space to form a complete single-modal deep feature representation, preserving the structural independence and complementarity of features in different frequency bands.

[0027] S55. Construct deep feature extraction networks with consistent structures for vibration signals and sound signals respectively, and splice and fuse the features of the two modes in a high-level feature space. A multimodal collaborative decision-making output is achieved through a fully connected fusion network.

[0028] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows: 1) The method of this invention achieves adaptive automatic selection of the number of VMD modes through a comprehensive evaluation based on signal correlation, complexity and reconstruction quality, avoiding under-decomposition or over-decomposition problems caused by manual experience setting, thereby improving the accuracy and stability of signal decomposition and reconstruction. 2) The method of this invention utilizes the frequency band characteristics of each IMF component after VMD decomposition to construct an adaptive evaluation mechanism based on statistical features, realizing the joint automatic optimization selection of wavelet function type and threshold parameter. Without the need for manual experience parameter tuning, it effectively suppresses noise interference and retains fault impact and structural information to the maximum extent, thereby significantly improving the denoising effect of complex vibration and sound signals and the stability and accuracy of subsequent fault identification. 3) In the method of the present invention, VMD and wavelet thresholding are organically combined to give full play to their respective advantages. VMD effectively decomposes the signal frequency band, and wavelet thresholding is used for targeted noise reduction. 4) The method of this invention splices and fuses signals with complementary physical mechanisms of vibration and sound in a high-level feature space and sends them into a fusion classification network for decision-making. It effectively utilizes the complementarity of vibration signals being sensitive to structural state and sound signals being responsive to overall operating state, thereby improving the ability to identify early and weak defects under complex working conditions and the generalization performance of the model. 5) In the method of this invention, the IMF signals after VMD decomposition are grouped and modeled by multi-scale convolution + channel attention. The model can more accurately capture the characteristics of compressor fault signals in different frequency bands and physical modes, thereby improving the accuracy and robustness of fault classification. Attached Figure Description

[0029] Figure 1 : Overall flowchart of the method of the present invention; Figure 2 In this invention, a comparison diagram of normal signals and defective signals is shown. Figure 3 In this invention, a comparison diagram of intermediate frequency components before and after wavelet threshold denoising is shown. Figure 4 In this invention, a comparison diagram of high-frequency components before and after threshold noise reduction is shown. Figure 5 The dual-branch CNN network structure diagram in this invention; Figure 6 The confusion matrix result diagram for defect identification in this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0032] The overall flowchart of the present invention is as follows: Figure 1 As shown.

[0033] S1. The deployment of the data acquisition system and the data acquisition work are as follows: A multimodal data acquisition system for defect identification in scroll compressors was constructed. The system includes a scroll compressor, a comprehensive performance test bench for the scroll compressor, a triaxial vibration sensor, a test microphone, a microphone preamplifier, a data acquisition board, and a host computer. The compressor is a scroll compressor, the sensor is a triaxial vibration sensor and a microphone, the data acquisition board is an NI board, and the host computer is a laptop computer configured with an i7-12700H CPU. The software uses a dedicated development tool to create an operation panel for the data acquisition device. This software can control the acquisition in real time and save the data to an Excel file.

[0034] The system uses an accelerometer and a microphone to synchronously collect the compressor's casing vibration and radiated acoustic signals under a unified time reference, thus forming time-aligned raw vibration-sound data. This provides a foundation for subsequent cross-modal collaborative analysis. The collected data is categorized into five types: Normal; Metal foreign object contamination (J); Anti-rotation missing part (L); Missing wear-resistant pad (N); Insufficient lubrication of the moving plate (W).

[0035] S2. Data preprocessing, as detailed below: A program was written using Python 3 to first establish a mapping relationship between data categories and labels. Vibration signal data was divided into normal states and four abnormal states, with the following mapping relationships: Normal state (Normal) mapped to category label 0, J-type abnormal state to category label 1, L-type abnormal state to category label 2, N-type abnormal state to category label 3, and W-type abnormal state to category label 4. Data for each category was stored in different subdirectories, and the data category was automatically identified by the file path. A sliding window was then used to segment each Excel data file. The specific steps were as follows: Data from columns 2, 3, and 4 were extracted, corresponding to vibration signals along the X, Y, and Z axes, respectively; the window length L was set to 3000 data points, the sliding step size S to 1500 data points, and the window overlap rate to 50% to ensure sufficient data coverage; the data within each sliding window was standardized to eliminate the influence of dimensions and improve the stability of model training. The standardization formula is as follows: , Where: μ is the mean of the data within the window, σ is the standard deviation of the data within the window, and ε is a local constant (taking...). This prevents division by zero errors. The processed data is saved as an NPY format file with a specific dictionary structure. A comparison chart of the processed normal and defective signals is shown below. Figure 2As shown in the figure, the horizontal axis represents only the number of sample points and has no specific physical meaning, while the vertical axis represents the amplitude of the vibration. The figure shows that although the defect signal differs from the normal signal, further processing is needed to extract more obvious defect features and improve the accuracy of the final identification.

[0036] S3. Determine the optimal number of decomposition modes K for VMD, and perform variational mode decomposition as follows: The system iterates through five preset data directories, including normal and four abnormal states (J, L, N, W), and reads vibration acceleration signals along the X, Y, and Z axes from each Excel file. It then sets the VMD basic parameters, including a bandwidth constraint α=2000, noise tolerance τ=0, no DC component, initialization mode of 1, and convergence tolerance of [value missing]. The search range for the modal number K is set to 2 to 15, covering various cases from simple decomposition to fine decomposition. This range is determined based on the following considerations: K=2-3: suitable for cases with simple signal components; K=4-8: suitable for most rotating machinery vibration signals; K=9-15: suitable for fine decomposition under complex defect conditions. For each candidate K value, the following calculation procedure is performed: VMD decomposition is performed, calculating each modal component. Correlation coefficient with the original signal x Calculate the sample entropy for each modal component. Characterize signal complexity and define modal effectiveness metrics: Set threshold Only retain The modal components participate in the reconstruction; the signal is reconstructed using the effective modal components. And calculate two key quality indicators: , Reconstructing the correlation coefficient This metric measures the waveform fidelity between the reconstructed signal and the original signal: , Signal-to-noise ratio This indicator assesses the accuracy level of signal reconstruction.

[0037] Finally, a comprehensive evaluation of multiple indicators is conducted to determine the optimal K value. A weighted comprehensive evaluation model is then established to optimize the K value. The quality indicators corresponding to each K value are subjected to minimum-maximum normalization. , , The score is calculated using an experience-weighted comprehensive scoring formula: , The weighting coefficients were determined based on extensive experiments, highlighting the importance of waveform fidelity. The value with the highest overall score from K=2-15 was selected as the optimal number of modes for the signal. After determining the optimal K value, variational mode decomposition was performed. Based on previous optimization results, the optimal decomposition parameter was set: the number of modes. Bandwidth constraint parameters Noise tolerance DC component constraint Initialization method Convergence tolerance For each axial signal x(t), VMD decomposes it into K eigenmode functions: , The constraints are satisfied: , The uniaxial vibration signal was extracted, and the VMD algorithm was executed to obtain eight modal components and a center frequency. The modal components were arranged in ascending order of center frequency, and the matrix dimensions were transposed to obtain an IMF matrix of shape (K,L). The modal components were initialized in the frequency domain to improve convergence efficiency. The ADMM algorithm was used to solve the constrained optimization problem: updating the modes and updating the center frequency. After modal component post-processing, data quality control, and verification, the eight decomposed components were sorted in ascending order of frequency band (IMF0-7) and finally the data was stored.

[0038] S4. Denoising is performed on the components after variational mode decomposition using the wavelet thresholding method, as follows: First, the eight Intrinsic Mode Functions (IMFs) obtained from variational mode decomposition of each original signal are divided into three groups according to their center frequencies from low to high: IMF0-1 for the low-frequency band, IMF2-4 for the mid-frequency band, and IMF5-7 for the high-frequency band. This grouping allows the IMF components of different frequency bands to be processed according to their physical meanings. The low-frequency IMFs mainly reflect the overall trend of the signal, the mid-frequency IMFs mainly contain modulation features related to defects, and the high-frequency IMFs mainly reflect transient impacts and random noise components. Considering the differences between vibration signals and sound signals in their physical generation mechanisms and spectral characteristics, this embodiment constructs wavelet candidate pools for different signal types. For vibration signals: For high-frequency IMF, db2 and Haar wavelets are selected because high-frequency vibrations have physical characteristics such as short-term impacts, spikes, non-stationarity, and are often accompanied by noise. db2 and Haar wavelets have short support, are sensitive to protrusions, and can effectively reduce noise. For mid-frequency IMF, sym4 and coif3 wavelets are selected because mid-frequency IMF contains the main defect modulation information, and its symmetry and phase are crucial. The sym4 wavelet has approximately linear phase and maintains morphological fidelity, while the coif3 wavelet has a high vanishing moment number and good structural representation. For low-frequency IMF, db6 wavelets are selected because low-frequency IMF has long periods, is stable, and basically does not contain impacts. db6 wavelets have the advantages of long support, smoothness, and strong trend expression capabilities. Noise reduction for low-frequency IMFs will almost never inadvertently damage defect features. For audio signals: For high-frequency IMF, db4 and sym4 wavelets are selected because they are smoother, reducing harshness and offering better symmetry, thus avoiding timbre distortion. For mid-frequency IMF, sym6 and coif3 wavelets are selected because the mid-frequency band contains formants, requiring higher vanishing moments and better reconstructed phase. For low-frequency IMF, db8 wavelet is selected because low-frequency audio signals are continuous, stable, and have long-term correlation characteristics; the db8 wavelet is very smooth, minimizing low-frequency ringing and ensuring a natural sound. Finally, for threshold scaling factors, a strong noise reduction factor of 1.0 is used for low frequencies, a compromise noise reduction factor of 0.5 is used for mid-frequency frequencies, and a conservative noise reduction factor of 0.2 is used for high frequencies to avoid erasing defect impacts. Based on this, a parameter configuration table for signal type adaptation is established: Table 1. Parameter Configuration Table for Signal Type Adaptation

[0039] A set of values ​​before and after noise reduction is shown in the table below: Table 2 A set of values ​​before and after noise reduction .

[0040] For each frequency band's IMF group, all IMF components within the group are concatenated to form a representative signal. Candidate wavelets are then selected one by one from the corresponding wavelet candidate pool for experimental denoising. Kujicic, spectral entropy, and energy retention rate are calculated for the signals before and after denoising, and a comprehensive evaluation function is constructed. , in, and These represent the kurtosis values ​​of the signals before and after noise reduction, respectively. and These represent the spectral entropy values ​​of the signal before and after noise reduction, respectively. This represents the energy retention rate. The candidate wavelet with the largest comprehensive evaluation function value is determined as the optimal wavelet for the IMF in this frequency band. After selecting the optimal wavelet, wavelet decomposition is performed on each modal component to obtain approximation coefficients and detail coefficients. The number of decomposition levels is determined by the signal length and wavelet basis functions, and the maximum number of levels does not exceed the preset max_level_cap. The noise standard deviation is estimated using the following formula: , in These are the highest frequency wavelet detail coefficients. The threshold is calculated based on the selected thresholding method. Bayesian threshold: , General threshold: , Where cD is the detail coefficient and n is the signal length.

[0041] A soft thresholding function is used to process the detail coefficients: Soft threshold: , Wavelet reconstruction is performed on the thresholded coefficients to obtain the denoised modal components, and the data is finally stored. The denoising effect on different IMFs is compared; for the mid-frequency IMF, some peaks are removed, contrast is improved, and other subtle features are revealed, such as... Figure 3 As shown; for high-frequency IMF, some peaks were also removed, making subtle features more visible and improving feature contrast, such as... Figure 4 As shown.

[0042] S5. Construct a two-branch CNN network structure to extract defect features from vibration and sound signals respectively, fuse them in the middle stage, and then perform classification. The neural network structure diagram is shown below. Figure 5 As shown, the specific construction steps are as follows: The IMF components obtained from VMD decomposition are grouped as before: the first and second IMFs are designated as low-frequency components, the third to fifth as mid-frequency components, and the sixth to eighth as high-frequency components. The IMF components within each frequency band are concatenated along the signal axis to construct a multi-channel temporal input tensor, where the input signal duration is preferably 3000 sampling points. In the feature extraction stage, independent convolutional feature extraction sub-networks are constructed for the low-frequency, mid-frequency, and high-frequency IMF bands. Each sub-network employs a multi-scale one-dimensional convolutional structure; specifically, the kernel sizes are set to 3, 7, and 15, and the convolution process can be represented as follows: , in, This represents a one-dimensional convolution operation. The kernel size is The convolution weights.

[0043] The outputs of convolutions at different scales are concatenated along the channel dimension to obtain a multi-scale feature representation: , Feature information at different time scales is extracted using parallel convolution. The optimal number of output channels for multi-scale convolution is 16 or 32, and batch normalization and ReLU nonlinear activation functions are introduced after each convolutional layer. Simultaneously, max pooling with a stride of 2 is used to downsample the features. To enhance the discriminative ability of different IMF components in fault identification, a feature weighting mechanism based on channel attention is introduced in each frequency band sub-network. First, global average pooling is performed on the features along the time dimension: ,

[0044] Specifically, the channel attention module adopts a fully connected structure with a compression ratio of 8, maps the global temporal mean of each IMF channel to obtain weight coefficients between 0 and 1, and multiplies these weight coefficients with the original features channel by channel to achieve adaptive enhancement of key modal features. After feature extraction from the low-frequency, mid-frequency, and high-frequency sub-networks, the deep features corresponding to the three frequency bands are concatenated to form a high-dimensional feature representation of a single modality, and feature mapping is performed through a fully connected layer with a dimension of 128. For vibration signals and sound signals, feature extraction networks with the same structure are used to process them to obtain feature vectors for the two modalities. Further, the vibration modal features and sound modal features are concatenated and fused, and then input into the fusion classification network for decision-making. The fusion network preferably includes a fully connected layer with an output dimension of 128, a batch normalization layer, a ReLU activation layer, and a random deactivation layer with a deactivation probability of 0.3, and finally achieves classification of multi-class fault states through the output layer. During model training, the cross-entropy loss function is preferably used, and a label smoothing strategy is introduced. The loss function is defined as: , in, .

[0045] The label smoothing coefficient is preferably 0.05; the Adam optimization algorithm is used for parameter updates, and the learning rate is preferably... And combined with the weight decay coefficient is Regularization strategies are employed to prevent overfitting. Simultaneously, the learning rate is adaptively adjusted based on the changes in the validation set loss, and training is terminated early when the validation loss no longer decreases after several consecutive training epochs. Finally, the model is saved. Training results show that the accuracy on the validation set is above 98%, and the confusion matrix of the final model is as follows. Figure 6 The results show that the model has high classification and recognition accuracy.

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion, characterized in that, Includes the following steps: S1. Synchronously acquire the casing vibration signal and radiated sound signal of the scroll compressor under actual production conditions to form time-aligned raw vibration-sound data; S2. Preprocess the raw data to construct a dataset containing normal and abnormal states; S3. Perform variational mode decomposition on the preprocessed vibration and sound signals respectively, automatically determine the optimal number of decomposed modes K through a comprehensive decision criterion of multiple evaluation indicators, and obtain the IMF components of each intrinsic mode function; S4. Divide the decomposed intrinsic mode function (IMF) components into low-frequency, mid-frequency, and high-frequency subsets according to the set frequency band range, and perform wavelet threshold noise reduction on the signal characteristics of each frequency band using an adaptively selected wavelet function. S5. Construct a dual-branch convolutional neural network model to extract and fuse features of each frequency band component of the denoised vibration signal and sound signal, and output the defect identification result.

2. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 1, characterized in that, The abnormal state types in S1 include at least metal foreign object contamination, anti-rotation missing parts, missing wear-resistant pads, and insufficient lubrication of the moving disc.

3. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 1, characterized in that, The preprocessing process in S2 is as follows: The original signal is segmented by a sliding window to establish a mapping relationship between data categories and labels. The original signal is segmented by overlapping sliding windows, and the data within the window is standardized.

4. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 1, characterized in that, The process of automatically selecting the number of decomposition modes and performing variational mode decomposition in S3 is as follows: S31. Traverse the original dataset containing various state signals, read multi-axial vibration and sound signals, and set the basic parameters of variational mode decomposition; S32. Set the search range for the number of decomposed modes to adapt to different signal components; S33. Perform variational mode decomposition on each candidate mode number within the search range, calculate the correlation coefficient between the reconstructed signal and the original signal and the signal-to-noise ratio of the reconstructed signal based on the decomposition results, and calculate the weighted comprehensive score based on the correlation coefficient and the signal-to-noise ratio, and select the candidate value with the highest comprehensive score as the optimal decomposition mode number; S34. Based on the determined optimal number of modes, perform variational mode decomposition on the vibration and sound signals respectively, and sort the decomposed modal components according to the center frequency from low to high to obtain a structured time-frequency characterization.

5. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion according to claim 1, characterized in that, The specific process of S4 is as follows: S41. The modal components obtained from variational mode decomposition are divided into low-frequency, mid-frequency, and high-frequency bands according to a preset frequency range by analyzing the physical characteristics and frequency features of the vibration and sound signals. S42. Based on the characteristics of vibration signals and sound signals, establish wavelet candidate pools for vibration signals and sound signals respectively; S43. Apply corresponding threshold scaling factors to different frequency bands to avoid erasing defect impact characteristics; S44. In the wavelet candidate pool of the corresponding frequency band, select candidate wavelets one by one for experimental noise reduction processing. By calculating the kurtosis, spectral entropy and energy retention rate of the signal before and after noise reduction, determine the optimal wavelet type for the corresponding frequency band. S45. Based on the determined optimal wavelet type, perform wavelet thresholding noise reduction on each modal component.

6. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 5, characterized in that, In S41, the low-frequency band, the mid-frequency band, and the high-frequency band each contain 2-3 modal components.

7. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 5, characterized in that, In step S42, the wavelet candidate pool for vibration signal and the wavelet candidate pool for sound signal each contain 2-3 wavelet types in the low-frequency, mid-frequency, and high-frequency bands.

8. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion according to claim 5, characterized in that, In step S44, the kurtosis, spectral entropy, and energy retention rate are used to determine the optimal wavelet type through a weighted comprehensive scoring method.

9. The method for detecting latent defects in a scroll compressor based on acoustic-vibration complementary enhancement fusion as described in claim 1, characterized in that, The specific process of S5 is as follows: S51. The intrinsic mode functions of the multi-axis signal obtained by VMD decomposition are reconstructed in the channel dimension to form a multi-channel time-series input tensor. Based on the frequency characteristics of the intrinsic mode functions, the IMF channels in the input tensor are divided into three frequency band subsets: low frequency, medium frequency, and high frequency, according to the set frequency band range. S52. For each frequency band subset, construct an independent convolutional subnetwork. In each subnetwork, adopt a multi-scale one-dimensional convolutional structure, extract features in parallel through convolutional kernels of different sizes, and perform channel splicing and pooling processing on the multi-scale convolution results. S53. An IMF channel attention module is introduced after the multi-scale convolution output. Through global temporal statistics and nonlinear mapping, the importance weights of different IMF channels in fault detection are learned, and the feature channels are adaptively recalibrated. S54. The features extracted from the low-frequency, mid-frequency, and high-frequency IMF subsets through network branches are spliced ​​and fused in a high-dimensional feature space to form a single-modality deep feature representation. S55. Deep features are extracted from vibration and sound signals through processes S51 to S54 respectively, and the deep features of the two modes are spliced ​​and fused in the high-level feature space. Then, multi-class defect identification is achieved through a fully connected fusion network.

Images