Compressor defect diagnosis method based on multi-information fusion based on improved autoencoder
By improving the multi-information fusion method of the autoencoder and utilizing variational mode decomposition and supervised learning, the complementary features of the compressor vibration signal are extracted, which solves the accuracy and efficiency problems of defect diagnosis in fully enclosed compressors and achieves efficient defect detection with a small number of labeled samples.
Patent Information
- Application Number
- CN202310596890.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2025-09-23
- Estimated Expiration
- 2043-05-25
AI Technical Summary
Existing technologies make it difficult to accurately diagnose hidden defects during the compressor production process, especially in fully enclosed structures. The instability and nonlinearity of the shell vibration signal make manual feature extraction difficult, and supervised learning requires a large number of labeled samples.
An improved autoencoder is used for multi-information fusion. The complementary time domain and time-frequency domain features of the vibration signal are extracted through variational mode decomposition. The autoencoder is trained with unlabeled data and combined with supervised learning to construct a compressor defect diagnosis model.
With a small number of labeled samples, the compressor defect detection rate is improved, the noise impact is reduced, and fast and accurate defect diagnosis is achieved.
Smart Images
Figure CN116662735B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fully enclosed compressor defect diagnosis, and in particular relates to a multi-information fusion compressor defect diagnosis method based on an improved autoencoder. Background Art
[0002] The compressor is the most complex and challenging core component in a refrigeration system. Failures can impact the system's service life and comfort, and even damage it. Most compressor failures during operation stem from hidden defects that are not readily apparent during production. Therefore, timely and accurate diagnosis of compressor manufacturing defects is crucial during production.
[0003] Internal manufacturing defects in compressors cannot be directly detected because assembled compressors are fully enclosed structures. Currently, the vibration signal from the casing is commonly used for diagnosis. Due to the large number of components and the presence of noise interference, the casing vibration signal exhibits instability and nonlinearity, making it difficult to manually extract effective features to distinguish fault types. With the development of deep learning theory and methods, it is possible to automatically extract defect-related features for compressor defect classification. Supervised learning requires a large number of labeled samples for training, but in practice, collecting a large number of labeled samples requires considerable expertise and manpower. Therefore, it is necessary to train diagnostic models using a small number of labeled samples. Summary of the Invention
[0004] In view of the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide a multi-information fusion compressor defect diagnosis method based on an improved autoencoder, which can extract and fuse the time domain and time-frequency domain features of different sensors for compressor defect detection with a small number of labeled samples, thereby improving the detection rate of compressor defects.
[0005] The present invention provides the following technical solutions:
[0006] The multi-information fusion compressor defect diagnosis method based on the improved autoencoder is characterized by: decomposing and decoupling the vibration signal by using the optimized variational mode decomposition method, training the improved autoencoder with an unlabeled data set, and fine-tuning the classifier with a labeled data set to achieve compressor defect diagnosis. Specifically, the method includes the following steps:
[0007] S1: Build a vibration data acquisition system to collect and save the shell vibration signals of compressors with different defect types.
[0008] S2: Use the maximum sum of kurtosis factors as the appropriate value to find the optimal number of decompositions K for each sensor, and use variational mode decomposition to decompose each sample to obtain K intrinsic mode components IMFs of different frequency scales.
[0009] Specifically, during compressor operation, multiple vibration sources are generated, transmitting vibration signals to the housing through different transmission paths. Simultaneously, these signals are affected by the noise of the production environment, causing the housing vibration signal to exhibit nonlinear and non-stationary characteristics. Due to the varying number of vibration sources in different directions, a varying number of subsequences at different frequency scales are generated, allowing for the extraction of more complementary time-frequency domain features to mitigate the impact of noise and signal aliasing on test results.
[0010] S3: K intrinsic mode components (IMFs) of different frequency scales represent the vibration signals of the shell represented by vibration sources of different frequencies. The characteristics represented by each subsequence are different. Therefore, the subsequences of different frequency scales are spliced with the original vibration signal as the input of the model to extract complementary time domain and time-frequency domain features.
[0011] S4: In order to make full use of the collected data samples for training and validating the proposed model, the input data is divided into unlabeled training set, labeled training set and test set.
[0012] S5: Construct three parallel improved autoencoder models to extract and expand the features of input data in three directions, and use unsupervised learning methods to obtain the optimal parameters of the improved autoencoder model.
[0013] S6: Freeze the optimal parameters of the encoder in the improved autoencoder, use supervised learning methods to train the classifier, and obtain the model with the optimal classifier parameters for compressor defect diagnosis.
[0014] Specifically, three parallel improved encoders fuse the input data into the dataset, which can respectively extract and compress the time domain and time-frequency domain features that are complementary in three directions, and three parallel decoders expand the extracted features; the improved convolution and pooling layers in the autoencoder can capture local features and reduce redundant features; the Transformer not only models the long-distance relationship of time series signals and the relationship between subsequences at different frequency scales, but also assigns weights to each feature.
[0015] S7: Evaluate and validate the diagnostic performance and robustness of the model.
[0016] Furthermore, the specific process of step S2 is as follows:
[0017] S21: Set the value range of K [k min ,k max ], the vibration signals in different directions are traversed to obtain K IMFs respectively;
[0018] S22: Calculate the kurtosis factors of K IMFs in different directions and add them together to obtain [KI kmin ,...,KIkmax ];
[0019] S23: Obtain the K value corresponding to the maximum sum of the kurtosis factors as the number of decompositions of the vibration signal in this direction.
[0020] Furthermore, the specific process of step S5 is as follows:
[0021] S50: Construct a convolutional Transformer autoencoder model to learn the distribution of unlabeled data samples;
[0022] S51: Input the data of the three directions in the unlabeled training set into three parallel encoders respectively, and extract and compress the time domain and time-frequency domain features of the input data;
[0023] S52: The time domain and time-frequency domain features extracted by the three parallel encoders are used as inputs of three parallel decoders, and the decoders expand the feature dimensions;
[0024] S53: The improved autoencoder is trained, and the mean square error between the decoder output and the original vibration signal is used as the loss function, so as to obtain the optimal parameters to minimize the loss function.
[0025] Furthermore, in step S6, the classifier is supervised and trained using the labeled training set to obtain the optimal classifier parameters, and the time domain and time-frequency domain features of multiple sensors are fused and extracted. The specific process is as follows:
[0026] S61: In the feature set of the model, the output features of the three parallel encoders are fused as the input of the classifier to achieve the fusion of multi-sensor information;
[0027] S62: The model is trained using the stochastic gradient descent algorithm with 20 training iterations. The mean square error between the classifier output and the label is used as the loss function to optimize the model parameters.
[0028] Furthermore, in step S7, different levels of signal-to-noise ratios are added to the test set to evaluate and verify the diagnostic performance and robustness of the model.
[0029] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:
[0030] 1) In the present invention, the vibration signal is decomposed through variational modal decomposition to obtain subsequences of different frequency scales. The subsequences with complementary features are spliced with the original vibration signal to provide the model with more complementary time domain and time-frequency domain features. The improved autoencoder uses convolution and Transformer to extract effective features for classification. The convolution and pooling layers have the ability to extract local features and remove redundant features. The Transformer can realize long-distance relationship connections in time series, realizing the capture and identification of local and global features. The classifier can extract and fuse the time domain and time-frequency domain features of different sensors, thereby improving the detection rate of compressor defects and meeting the actual needs of the production line for fast and accurate detection.
[0031] 2) In the present invention, a semi-supervised learning method is used to diagnose compressor defects. With a small number of labeled samples, the detection rate of compressor defects is improved;
[0032] 3) In the present invention, the original vibration is decomposed and decoupled by using optimized variational modes, thereby reducing the impact of noise in the actual production environment on effective feature extraction;
[0033] 4) In the present invention, the natural modes and the original vibration signals are fused in the data set as the input of the model, which can extract more time domain and time-frequency domain features related to the defects.
[0034] 5) In the present invention, the time-frequency domain features and different sensor features are fused in the data set and feature set respectively, and convolution pooling and Transformer are used to extract and fuse the local features and global features of the time domain and time-frequency domain features in different directions, and the weights of the features are automatically allocated, thereby increasing the expression capabilities of the time domain and time-frequency domain features. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A flow chart of a method for diagnosing compressor defects according to the present invention;
[0036] Figure 2 Schematic diagram of the compressor defect diagnosis model of the present invention;
[0037] Figure 3 Schematic diagram of the confusion matrix of the diagnostic model when the number of labeled samples is 270 in an embodiment of the present invention. DETAILED DESCRIPTION
[0038] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0039] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.
[0040] See also Figure 1-3 , a multi-information fusion compressor defect diagnosis method based on an improved autoencoder specifically includes the following steps:
[0041] S10: Build a data acquisition system, use a three-axis acceleration sensor to collect and save the vibration signal of the compressor housing;
[0042] S11: Connect the compressor, triaxial sensor, data acquisition card and computer in the data acquisition system;
[0043] S12: The triaxial acceleration sensor is installed in the center of the upper housing, with the X direction parallel to the piston's movement direction, the Z direction perpendicular to the compressor placement plane, and the Y direction perpendicular to the X and Z directions.
[0044] S13: Set the frequency of the three-axis acceleration sensor to 25.6kHz, collect the shell vibration signals of compressors with different manufacturing defect types, and divide the vibration signals into 0.1s time intervals to obtain each sample data, that is, the dimension of the sample data is (2560, 1).
[0045] S20: Using the sum of the kurtosis factors as the appropriate value to find the number of decompositions K of variational mode decomposition, use variational mode decomposition to decompose vibration signals in different directions, and obtain K intrinsic mode components IMFs of different frequency scales;
[0046] S21: Set the value range of K [k min ,k max ], the vibration signals in different directions are traversed to obtain K IMFs respectively;
[0047] S22: Calculate the kurtosis factors of K IMFs in different directions and add them together to obtain [KI kmin ,...,KI kmax ];
[0048] S23: Get the K value corresponding to the maximum sum of kurtosis factors as the number of decompositions of the vibration signal in that direction. The number of decompositions in different directions is different, and the number of decompositions is shown in Table 1:
[0049] Table 1
[0050] Different directions X direction Y direction Z direction K 3 8 6
[0051] S30: The dimensions of the subsequences of different frequency scales of different sensors are (2560, 3), (2560, 8), and (2560, 6), respectively. The IMFs are fused with the original vibration signal as the input data of the model. The input dimensions of the vibration signals in the three directions are (2560, 4), (2560, 9), and (2560, 7), respectively.
[0052] S40: In order to make full use of the collected data samples for training and validating the proposed model, the sample data is divided;
[0053] S41: Divide the data samples into an unlabeled training dataset, a labeled training set, and a test set, which are used for unsupervised training, supervised training, and evaluation of model performance, respectively;
[0054] S42: The number of unlabeled training sets is 3000, the number of labeled training sets gradually increases from 90 to 1080 (the number of samples for each fault type is 15 to 180), and the number of test samples is 1200.
[0055] S50: Construct a convolutional Transformer autoencoder model to learn the distribution of data for unlabeled data samples; the model diagram is as follows Figure 2 As shown;
[0056] S51: Data from the three directions of the unlabeled training set are fed into three parallel encoders, which extract and compress the time and frequency domain features of the input data. The encoders consist of two convolutional pooling layers and a Transformer module. The convolutional layers have kernels of 64x16 and 32x8, with strides of 4 and 2, respectively. The pooling layers use max pooling, with a stride of 4. The Transformer module consists of two basic blocks, four multi-head attention mechanisms, and 32-dimensional queries and keys. The output dimension of each parallel encoder is (None, 320, 1).
[0057] S52: The time domain and time-frequency domain features extracted by the three parallel encoders are used as the input of three parallel decoders. The decoders expand the feature dimension, and their output dimension is (None, 2560, 1). The decoder consists of five deconvolution layers. The convolution kernels and steps of the first three layers are 32X8 and 4 respectively; the convolution kernels of the last two layers are 16X8 and 1X4 respectively, and the steps are 2 and 1 respectively.
[0058] S53: The improved autoencoder is trained using the Adam algorithm with 100 training iterations. The mean square error between the decoder output and the original vibration signal is used as the loss function. The goal is to obtain the optimal parameters to minimize the loss function.
[0059] S60: Use the labeled training set to supervise the classifier training to obtain the optimal classifier parameters, and fuse and extract the time domain and time-frequency domain features of multiple sensors.
[0060] S61: The output features of the three parallel encoders are fused in the feature set of the model as the input of the classifier, thereby realizing the fusion of multi-sensor information.
[0061] S62: The classifier consists of a convolutional pooling layer, Transformer, two fully connected layers and a Softmax layer, such as Figure 2 As shown in the figure, the number of kernels and stride of the convolutional layer is 32x8 and 1, respectively. The stride of the pooling layer is 3. The Transformer parameters are the same as those of the encoder. The number of neurons in the fully connected layer is 512 and 128, respectively. The Dropout value is 0.3. The parameter of the SoftMax layer is 6, which is the total number of hidden defect types, including five hidden defect types and the healthy type.
[0062] S63: The model is trained using the stochastic gradient descent algorithm with 20 training iterations. The mean square error between the classifier output and the label is used as the loss function to optimize the model parameters.
[0063] S70: Obtain the optimal defect classification model parameters and use the test set to evaluate and verify the model's performance. 1,200 test samples are used to evaluate the model's diagnostic performance.
[0064] In this example, the number of unlabeled training set samples is 3000, and the number of labeled training sets is increasing. The diagnostic accuracy of different defect types in the test set is shown in Table 2:
[0065] Table 2
[0066] Number of labeled samples 90 180 270 540 1080 Diagnostic accuracy (%) 92.8 96.5 98.3 98.6 98.9
[0067] In this example, in order to verify the application effect of the model in real scenarios, different signal-to-noise ratios are added to each sample in the test set to simulate real environmental noise. The test accuracy is shown in Table 3.
[0068] Table 3
[0069]
[0070] 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 in the scope of protection of the present invention.
Claims
1. A multi-information fusion compressor defect diagnosis method based on an improved autoencoder, characterized by: The vibration signal is decomposed and decoupled using the optimized variational mode decomposition method. The autoencoder is improved by training with an unlabeled dataset. The classifier is adjusted with a labeled dataset to achieve compressor defect diagnosis. The specific steps include: S1: Build a vibration data acquisition system to collect and save the shell vibration signals of compressors with different defect types; S2: Use the maximum sum of kurtosis factors as the appropriate value to find the optimal number of decompositions K for each sensor, and use variational mode decomposition to decompose each sample to obtain K intrinsic mode components (IMFs) of different frequency scales; S3: concatenate the natural mode of each sample with the original vibration signal as input data; S4: Divide the input data into unlabeled training set, labeled training set and test set; S5: Build three parallel improved autoencoder models to extract and expand the features of the input data in three directions, and use unsupervised learning methods to obtain the optimal parameters of the improved autoencoder models; S6: Freeze the optimal parameters of the encoder in the improved autoencoder and use supervised learning methods to train the classifier. The model with the optimal classifier parameters is used for compressor defect diagnosis. S7: Evaluate and validate the diagnostic performance and robustness of the model; The improved autoencoder consists of two convolutional pooling layers and a Transformer module. The number of kernels in the convolutional layer is 64X16 and 32X8, and the number of steps is 4 and 2 respectively; the pooling layer adopts maximum pooling, and its number of steps is 4; the Transformer consists of 2 basic blocks, 4 multi-head attention mechanisms, and 32-dimensional queries and keys; the output dimension of each parallel encoder is (None, 320, 1).
2. The multi-information fusion compressor defect diagnosis method based on the improved autoencoder according to claim 1 is characterized in that The specific process of step S2 is as follows: S21: Set the value range of K , the vibration signals in different directions are traversed to obtain K IMFs respectively; S22: Calculate the kurtosis factors of K IMFs in different directions and add them together to obtain ; S23: Obtain the K value corresponding to the maximum sum of the kurtosis factors as the number of decompositions of the vibration signal in this direction.
3. The multi-information fusion compressor defect diagnosis method based on the improved autoencoder according to claim 1 is characterized in that In step S5, a convolutional Transformer autoencoder model is constructed to learn the distribution of data of unlabeled data samples. The specific process is as follows: S51: Input the data of the three directions in the unlabeled training set into three parallel encoders respectively, and extract and compress the time domain and time-frequency domain features of the input data; S52: The time domain and time-frequency domain features extracted by the three parallel encoders are used as inputs of three parallel decoders, and the decoders expand the feature dimensions; S53: The improved autoencoder is trained, and the mean square error between the decoder output and the original vibration signal is used as the loss function, so as to obtain the optimal parameters to minimize the loss function.
4. The multi-information fusion compressor defect diagnosis method based on the improved autoencoder according to claim 1 is characterized in that In step S6, the classifier is supervised and trained using the labeled training set to obtain the optimal classifier parameters, and the time domain and time-frequency domain features of multiple sensors are fused and extracted. The specific process is as follows: S61: In the feature set of the model, the output features of the three parallel encoders are fused as the input of the classifier to achieve the fusion of multi-sensor information; S62: The model is trained using the stochastic gradient descent algorithm with 20 training iterations. The mean square error between the classifier output and the label is used as the loss function to optimize the model parameters.
5. The multi-information fusion compressor defect diagnosis method based on the improved autoencoder according to claim 1 is characterized in that In step S7, different levels of signal-to-noise ratios are added to the test set to evaluate and verify the diagnostic performance and robustness of the model.