A multi-scale intelligent decision-making method based on transfer learning

By adopting a multi-scale intelligent decision-making method based on transfer learning in the mechanical fault diagnosis system, combining time domain and frequency domain characteristics, the problem of existing systems dependence on label data and single features is solved, and efficient and accurate fault diagnosis is achieved.

CN118797448BActive Publication Date: 2024-11-15JIANGNAN UNIV +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411284257.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-11-15
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

The existing intelligent mechanical fault diagnosis system requires a large amount of label data for training, and only relies on time domain features, ignoring the frequency domain features, resulting in inaccurate decision-making and poor generalization capabilities.

Method used

Using a multi-scale intelligent decision-making method based on transfer learning, data augmentation and wavelet transformation are performed by obtaining mechanical fault vibration timing signals, extracting time domain and frequency domain features, building a cross-correlation loss function, training time domain and frequency domain encoder, and forming Multi-scale intelligent decision-making model and fine-tune it in the target domain using transfer learning.

Benefits of technology

It improves the efficiency and accuracy of fault detection, reduces the calculation amount and complexity of feature extraction, enhances the extraction ability of frequency domain features, solves the problem of insufficient label data, and realizes accurate fault diagnosis in actual industrial environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118797448B_ABST
    Figure CN118797448B_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of machine fault diagnosis, and discloses a multi-scale intelligent decision-making method based on transfer learning, comprising: obtaining mechanical fault vibration time series long signals in a source domain and a target domain, and constructing a source domain training set and a target domain training set; in the source domain training set, after pre-processing the time series sub-signals, extracting time domain features and frequency domain features, and mapping them into low-dimensional time domain features and low-dimensional frequency domain features; calculating the cross-correlation matrix between the features, constructing a cross-correlation loss function, to train an encoder, obtain a source domain optimized time domain signal encoder, and migrate it to a target domain, and form an initial multi-scale intelligent decision model with an initial classifier; predicting the target domain time series sub-signals, and constructing a cross entropy loss function with their true labels, and training to obtain a target multi-scale intelligent decision model; pre-processing the real-time collected mechanical fault vibration time series long signals, respectively inputting them into the target multi-scale intelligent decision model, and obtaining the corresponding fault category.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of machine fault diagnosis, and in particular to a multi-scale intelligent decision-making method based on transfer learning. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, intelligent mechanical fault diagnosis and decision systems are becoming increasingly important in modern industry. These systems use advanced sensing technology, data processing technology, and artificial intelligence algorithms to monitor and analyze the operating status of mechanical equipment in real time, predict and diagnose possible faults in advance, thereby improving production efficiency and reducing maintenance costs. Traditional mechanical fault diagnosis methods usually rely on expert experience and regular maintenance, which is not only inefficient, but may also lead to failure to detect faults in time, causing equipment downtime or even safety accidents. With the advancement of computer technology and big data technology, intelligent fault diagnosis systems have gradually become an effective way to solve these problems. By analyzing a large amount of operating data, intelligent fault diagnosis systems can extract useful fault features, establish fault diagnosis models, and achieve accurate evaluation and prediction of equipment status.

[0003] At present, decision-making systems based on deep learning are often used, especially those based on convolutional neural networks (CNNs), which have achieved great success in the fields of image processing and pattern recognition and have also been introduced into mechanical fault diagnosis. CNNs can automatically extract high-level features of input data, reducing the reliance on manual feature engineering. However, deep learning models usually require a large amount of labeled data for training, and in actual industrial environments, it is often difficult to obtain large-scale labeled data.

[0004] Although there are many deep learning decision systems in this field, most of these methods require a large amount of actual industrial labeled data for training. In order to obtain a better decision model, a large amount of operating data needs to be analyzed. At the same time, these models usually only extract time domain information features and ignore frequency domain information features, resulting in low generalization ability, poor portability, and unstable decision-making of these decision models.

[0005] In summary, the existing diagnostic system requires a large amount of labeled data for intelligent decision-making, but there is no large amount of labeled data for actual decision-making; and the features relied on for analysis by its intelligent decision-making are too single, relying only on time domain features while ignoring frequency domain features. The correlation and effectiveness of the extracted related features are insufficient, which leads to inaccurate diagnostic results, inability to carry out timely maintenance, and affecting the normal operation of mechanical equipment. Summary of the invention

[0006] To this end, the technical problem to be solved by the present invention is to overcome the problem in the prior art that intelligent decision-making requires a lot of label data and ignores frequency domain features when making decisions, resulting in inaccurate decisions.

[0007] In order to solve the above technical problems, the present invention provides a multi-scale intelligent decision-making method based on transfer learning, comprising:

[0008] Obtain the long time series signals of mechanical fault vibration in the source domain and the target domain, slice them according to the preset signal length, obtain multiple source domain time series sub-signals and target domain time series sub-signals, and construct the source domain training set and the target domain training set;

[0009] Data enhancement is performed on each time-series sub-signal in the source domain training set to obtain the corresponding enhanced time-series sub-signal, which is input into the source domain initial time-domain signal encoder to obtain the corresponding time-domain features;

[0010] Perform continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain the corresponding frequency domain signal, which is input into the source domain initial frequency domain signal encoder to obtain the corresponding frequency domain features;

[0011] The time domain features and frequency domain features of each time series sub-signal are respectively input into the initial time domain mapping transformer and the initial frequency domain mapping transformer of the same dimension to obtain the corresponding low-dimensional time domain features and low-dimensional frequency domain features;

[0012] Based on the Euclidean norm, the cross-correlation matrix between the low-dimensional time domain features and the low-dimensional frequency domain features of each time series sub-signal is calculated to construct a cross-correlation loss function;

[0013] Based on the source domain training set and the cross-correlation loss function, the source domain initial time domain signal encoder and the source domain initial frequency domain signal encoder are trained until the cross-correlation loss function converges, thereby obtaining a source domain optimized time domain signal encoder and a source domain optimized frequency domain signal encoder;

[0014] Initializing the source domain optimized time domain signal encoder as the target domain initial time domain signal encoder and the initial classifier to form an initial multi-scale intelligent decision model;

[0015] Obtain the target domain time series sub-signal and its true label in the target domain training set, input the target domain time series sub-signal into the initial multi-scale intelligent decision model, and obtain the corresponding predicted fault category label;

[0016] Based on the predicted fault category label and the true label of the target domain time series signal, a cross entropy loss function is constructed, and the initialized multi-scale intelligent decision model is trained until the cross entropy loss function converges to obtain the target multi-scale intelligent decision model;

[0017] The long time-series vibration signal of mechanical fault collected in real time is segmented according to the preset signal length to obtain multiple time-series sub-signals, which are respectively input into the target multi-scale intelligent decision-making model to obtain the corresponding fault category.

[0018] Preferably, performing data enhancement on each time series sub-signal in the source domain training set to obtain the corresponding enhanced time series sub-signal includes: adding a Gaussian signal-to-noise ratio to each time series sub-signal, performing a Mask operation, performing a Gain operation, data translation, horizontal flipping, vertical flipping, and center flipping;

[0019] The adding of Gaussian signal-to-noise ratio comprises: Timing signal Add Gaussian SNR , get the corresponding noise-added time-series signal , expressed as: ;

[0020] The Mask operation is performed, including obtaining the Timing signal Frequency domain representation of , using the Mask function Perform the Mask operation on the frequency domain representation of the time series signal to obtain the Mask signal, which is expressed as: ;

[0021] The performing of the Gain operation includes adjusting the first The vibration value of each time-series divided signal is a preset vibration value.

[0022] Preferably, the time-series signal is input into a time-domain signal encoder to obtain time-domain features, including:

[0023] The time-series sub-signals are input into the time-domain signal encoder based on ResNet, and after passing through a one-dimensional convolution block, the convolution time-series sub-signals are obtained;

[0024] The convolution time-series signal is input into the first-layer residual block to obtain the first soft threshold output feature; the first-layer residual block is a soft threshold module, including:

[0025] The convolution time series signal is input into the first convolution layer unit, and convolved through the convolution layer, batch normalization layer, and activation function layer connected in series to obtain the first output feature;

[0026] The first output feature is input into the second convolutional layer unit, and convolved through the convolutional layer, batch normalization layer, and activation function layer connected in series to obtain the second output feature;

[0027] The second output feature is input into the connection layer unit, and the threshold is calculated through the normalization layer, the adaptive averaging layer and the fully connected layer connected in series, so as to obtain the channel threshold of each channel of the second output feature;

[0028] Inputting the second output feature and the channel threshold of each channel corresponding to it into the adaptive threshold unit, scaling the second output feature, and obtaining an adaptive feature;

[0029] Perform residual connection on the adaptive features and the input features to obtain the first soft threshold output features;

[0030] Inputting the first soft threshold output feature into the second layer residual block to obtain the second soft threshold output feature, where the second layer residual block is a soft threshold module;

[0031] The second soft threshold output feature is made to pass through the third convolution block and the fourth convolution block in sequence, and the convolution outputs the time domain feature corresponding to the time series sub-signal.

[0032] Preferably, the step of inputting the second output feature and the channel threshold of each channel corresponding to the second output feature into an adaptive threshold unit, scaling the second output feature, and obtaining the adaptive feature comprises:

[0033] Compare the second output feature Input channel characteristics of channels With preset channel threshold , calculate the output channel feature corresponding to the channel , expressed as:

[0034] ;

[0035] Obtaining adaptive features based on output channel features corresponding to all channels in the second output features;

[0036] in, Indicates the channel index and preset channel threshold Expressed as ; Indicates The scaling parameter of each channel is expressed as , Indicates Neuronal characteristics in each channel; Indicates that the width is , height is The feature map is averaged. Indicates The width of the channel is , height is The feature map of and Indicates the width and height of the feature map.

[0037] Preferably, performing continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain a corresponding frequency domain signal includes:

[0038] Use the preset mother wavelet to Timing signal Perform convolution to obtain the Timing signal The corresponding frequency domain signal , expressed as:

[0039] ;

[0040] in, is the preset scale parameter, To preset translation parameters, Represents the mother wavelet function The complex conjugate of .

[0041] Preferably, the frequency domain signal is input into a frequency domain signal encoder to obtain corresponding frequency domain features, including:

[0042] The frequency domain signal is input into the frequency domain signal encoder based on ResNet, and passes through the two-dimensional convolution block to obtain the convolution frequency domain signal;

[0043] The convolution frequency domain signal is input into the first layer residual block to obtain the first attention feature; the first layer residual block is a residual network based on a multi-scale attention module, including:

[0044] After the convolution frequency domain signal passes through the convolution block, batch normalization layer, and activation function layer in series, feature grouping is performed to obtain height features, width features, and global features;

[0045] After local average pooling of the height feature and the width feature, the two outputs are concatenated to obtain the concatenated feature, and 1×1 convolution and activation are performed to obtain the activated height feature and the activated width feature;

[0046] After performing 3×3 convolution on the global features, global average pooling and activation function are performed in sequence to obtain activated global features;

[0047] Based on global features, activation height features, and activation width features, feature dimension reconstruction and feature group normalization are performed to obtain local normalized features;

[0048] The local normalized features are passed through the average pooling layer and the activation function layer in sequence, and then multiplied with the global features by the feature matrix to obtain the local features;

[0049] After the local normalized features are multiplied by the feature matrix and the activated global features are residually connected with the local features, they are sequentially reconstructed, 3×3 convolution and batch normalization to obtain the global normalized features;

[0050] After the global normalized feature is residually connected with the convolution frequency domain signal, the first attention feature is output through the activation function;

[0051] Input the first attention feature into the second layer residual block, where the second layer residual block is a residual network based on a multi-scale attention module and outputs a second attention feature;

[0052] The second attention feature is passed through the third convolution block and the fourth convolution block in sequence, and the convolution outputs the frequency domain features corresponding to the frequency domain signal.

[0053] Preferably, the calculating of the cross-correlation matrix between the low-dimensional time domain features and the low-dimensional frequency domain features of each time series sub-signal based on the Euclidean norm includes:

[0054] ;

[0055] in, Represents the dimensions of low-dimensional time domain features and low-dimensional frequency domain features, , and Represent the low-dimensional frequency domain features and their transpose and Euclidean norm respectively; and Represent the low-dimensional time domain features and their Euclidean norms respectively.

[0056] Preferably, the cross-correlation loss function is expressed as:

[0057] ;

[0058] in, represents the cross-correlation matrix between the i-th dimension low-dimensional time domain features and the i-th dimension low-dimensional frequency domain features, represents the square of the cross-correlation matrix between the i-th dimension low-dimensional time domain feature and the j-th dimension low-dimensional frequency domain feature, A constant that represents the weighting of the contribution of unrelated terms.

[0059] Preferably, the predicted fault category label and the true label based on the target domain time series signal are used to construct a cross entropy loss function, which is expressed as:

[0060] ;

[0061] in, represents the cross entropy loss, Indicates the number of target domain timing sub-signals, Indicates the total number of fault categories; Indicates The true label of the target domain time series signal, Indicates The predicted fault category labels of target domain time series signals.

[0062] Preferably, the initializing multi-scale intelligent decision model is trained until the cross entropy loss function converges to obtain the target multi-scale intelligent decision model, including:

[0063] Train the initial classifier, back-propagate and update the model parameters to obtain the target classifier;

[0064] Train the target domain initial time domain signal encoder, and perform reverse transfer to fine-tune the parameters of the target domain initial time domain signal encoder to obtain the target domain optimized time domain signal encoder;

[0065] Based on the target classifier and the target domain optimized time domain signal encoder, a target multi-scale intelligent decision-making model is formed.

[0066] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0067] The multi-scale intelligent decision-making method based on transfer learning described in the present invention constructs a source domain training set containing multiple time-series sub-signals in the source domain based on the long time-series signal of mechanical fault vibration, and uses continuous wavelet transform to obtain the corresponding frequency domain signal; by constructing a cross-correlation matrix between low-dimensional time domain features and low-dimensional frequency domain features mapped based on the time domain features of the time-series sub-signals and the frequency domain features, a cross-correlation loss function is constructed, and the source domain initial time domain signal encoder and the source domain initial frequency domain signal encoder are trained to obtain the source domain optimized time domain signal encoder and the source domain optimized frequency domain signal encoder; the present invention uses continuous wavelet transform to obtain multi-scale frequency domain features, and uses the cross-correlation matrix to make the time domain and frequency domain correlated with each other to train the encoder, so that the time domain signal encoder has the ability to extract frequency domain features, reduces the amount and complexity of time domain and frequency domain feature extraction operations, and improves fault detection efficiency and accuracy. The present invention transfers the optimized time domain signal encoder in the source domain to the time series encoder in the target domain, and forms a multi-scale intelligent decision-making model with the initial classifier; based on the predicted fault category label and the true label of the target domain time series signal, a cross entropy loss function is constructed to train the initialized multi-scale intelligent decision-making model, optimize the classifier, fine-tune the target domain time domain signal encoder, and obtain the target multi-scale intelligent decision-making model for fault diagnosis; the present invention transfers the time domain signal encoder obtained by using the source domain training set to the target domain through transfer learning, and fine-tunes it, thereby solving the problem of insufficient target domain label data and being more suitable for actual industrial environments.

[0068] The present invention sets corresponding residual blocks in the time domain signal encoder and frequency domain signal encoder based on ReaNet for feature extraction; in the time domain signal encoder, a soft threshold module is constructed, and an adaptive threshold unit is used to reduce the interference of noise on the extracted features and extract more accurate features; in the frequency domain signal encoder, a multi-scale attention mechanism is embedded, and after the multi-scale feature groups are normalized, the frequency domain features related to the time domain are extracted from different spatial directions, which improves the robustness of the frequency domain signal encoder and the accuracy of the representation of the frequency domain features. The present invention uses wavelet transform to provide rich time-frequency features, uses deep learning models to automatically extract high-level features, uses transfer learning to solve the problem of insufficient labeled data, uses attention mechanisms to enhance the relevance and stability of features, and constructs an efficient, accurate and robust target multi-scale intelligent decision model to achieve accurate diagnosis of faults and ensure stable operation and efficient production of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0070] Figure 1 It is a flowchart of the steps of the multi-scale intelligent decision-making method based on transfer learning provided by the present invention;

[0071] Figure 2 It is a model architecture diagram corresponding to the multi-scale intelligent decision-making method based on transfer learning provided by the present invention;

[0072] Figure 3 It is a residual network structure diagram based on a multi-scale attention module provided by the present invention;

[0073] Figure 4 It is a schematic diagram of the migration training process of the multi-scale intelligent decision-making method based on migration learning provided by the present invention. DETAILED DESCRIPTION

[0074] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0075] Reference Figure 1 As shown, the step flow chart of the multi-scale intelligent decision-making method based on transfer learning provided by the present invention includes the following specific steps:

[0076] S101: Acquire a long time series signal of mechanical fault vibration in a source domain and a target domain, slice it according to a preset signal length, acquire a plurality of source domain time series sub-signals and target domain time series sub-signals, and construct a source domain training set and a target domain training set;

[0077] S102: Perform data enhancement on each time series sub-signal in the source domain training set to obtain a corresponding enhanced time series sub-signal, and input the enhanced time series sub-signal into the source domain initial time domain signal encoder to obtain a corresponding time domain feature;

[0078] S103: Perform continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain a corresponding frequency domain signal, and input the signal into a source domain initial frequency domain signal encoder to obtain a corresponding frequency domain feature;

[0079] S104: inputting the time domain features and frequency domain features of each time series sub-signal into an initial time domain mapping transformer and an initial frequency domain mapping transformer of the same dimension respectively, to obtain corresponding low-dimensional time domain features and low-dimensional frequency domain features;

[0080] S105: Based on the Euclidean norm, the cross-correlation matrix between the low-dimensional time domain features and the low-dimensional frequency domain features of each time series sub-signal is calculated, and a cross-correlation loss function is constructed, including:

[0081] S105-1: Construct a cross-correlation matrix, expressed as:

[0082] ;in, Represents the dimensions of low-dimensional time domain features and low-dimensional frequency domain features, , and Represent the low-dimensional frequency domain features and their transpose and Euclidean norm respectively; and Represent the low-dimensional time domain features and their Euclidean norms respectively.

[0083] S105-2: Construct a cross-correlation loss function, expressed as:

[0084] ;in, represents the cross-correlation matrix between the i-th dimension low-dimensional time domain features and the i-th dimension low-dimensional frequency domain features, represents the square of the cross-correlation matrix between the i-th dimension low-dimensional time domain feature and the j-th dimension low-dimensional frequency domain feature, A constant that represents the weighting of the contribution of unrelated terms.

[0085] S106: Based on the source domain training set and the cross-correlation loss function, the source domain initial time domain signal encoder and the source domain initial frequency domain signal encoder are trained until the cross-correlation loss function converges, thereby obtaining a source domain optimized time domain signal encoder and a source domain optimized frequency domain signal encoder;

[0086] S107: Initializing the source domain optimized time domain signal encoder as the target domain initial time domain signal encoder and the initial classifier to form an initial multi-scale intelligent decision model;

[0087] S108: Obtain the target domain time series sub-signals and their true labels in the target domain training set, input the target domain time series sub-signals into the initial multi-scale intelligent decision model, and obtain the corresponding predicted fault category label;

[0088] S109: constructing a cross entropy loss function based on the predicted fault category label and the true label of the target domain time series signal, training the initialized multi-scale intelligent decision model until the cross entropy loss function converges, and obtaining the target multi-scale intelligent decision model;

[0089] The cross entropy loss function is expressed as:

[0090] ;

[0091] in, represents the cross entropy loss, Indicates the number of target domain timing sub-signals, Indicates the total number of fault categories; Indicates The true label of the target domain time series signal, Indicates The predicted fault category labels of target domain time series signals.

[0092] Training the initialized multi-scale intelligent decision-making model includes: training the initial classifier, back-propagating to update the model parameters, and obtaining the target classifier; training the target domain initial time domain signal encoder, back-propagating to fine-tune the parameters of the target domain initial time domain signal encoder, and obtaining the target domain optimized time domain signal encoder; based on the target classifier and the target domain optimized time domain signal encoder, forming a target multi-scale intelligent decision-making model.

[0093] S110: Slice the long time-series vibration signal of the mechanical fault collected in real time according to the preset signal length to obtain multiple time-series sub-signals, which are respectively input into the target multi-scale intelligent decision-making model to obtain the corresponding fault category.

[0094] The multi-scale intelligent decision-making method based on transfer learning described in the present invention constructs a source domain training set containing multiple time-series sub-signals in the source domain based on the long time-series signal of mechanical fault vibration, and uses continuous wavelet transform to obtain the corresponding frequency domain signal; constructs a cross-correlation loss function by constructing a cross-correlation matrix between low-dimensional time-domain features and low-dimensional frequency-domain features based on the time-domain features and frequency-domain feature mapping of the time-series sub-signals, trains the source domain initial time-domain signal encoder and the source domain initial frequency-domain signal encoder, and obtains the source domain optimized time-domain signal encoder and the source domain optimized frequency-domain signal encoder; the present invention uses continuous wavelet transform to obtain multi-scale frequency-domain features, and uses the cross-correlation matrix to make the time domain and the frequency domain correlated with each other to train the encoder, so that the time-domain signal encoder has the ability to extract frequency-domain features, reduces the amount and complexity of time-domain and frequency-domain feature extraction operations, and improves the efficiency and accuracy of fault detection. The present invention transfers the optimized time domain signal encoder in the source domain to the time series encoder in the target domain, and forms a multi-scale intelligent decision-making model with the initial classifier; based on the predicted fault category label and the true label of the target domain time series signal, a cross entropy loss function is constructed to train the initialized multi-scale intelligent decision-making model, optimize the classifier, fine-tune the target domain time domain signal encoder, and obtain the target multi-scale intelligent decision-making model for fault diagnosis; the present invention transfers the time domain signal encoder obtained by using the source domain training set to the target domain through transfer learning, and fine-tunes it, thereby solving the problem of insufficient target domain label data and being more suitable for actual industrial environments.

[0095] Specifically, an embodiment of the present invention performs data enhancement on each time series sub-signal in the source domain training set to obtain a corresponding enhanced time series sub-signal, including: adding a Gaussian signal-to-noise ratio to each time series sub-signal, performing a Mask operation, performing a Gain operation, data translation, horizontal flipping, vertical flipping, and center flipping.

[0096] ① The adding of Gaussian signal-to-noise ratio comprises: Timing signal Add Gaussian SNR , get the corresponding noise-added time-series signal , expressed as: ;

[0097] ② The Mask operation is performed, including obtaining the Timing signal Frequency domain representation of , using the Mask function Perform the Mask operation on the frequency domain representation of the time series signal to obtain the Mask signal, which is expressed as: ;

[0098] ③ The Gain operation includes adjusting the first The vibration value of each time-series divided signal is a preset vibration value.

[0099] Reference Figure 2 As shown, this is a model architecture diagram corresponding to the multi-scale intelligent decision-making method based on transfer learning provided by the present invention; specifically, it includes a time domain processing flow and a frequency domain processing flow, and after processing the time domain features and the frequency domain features, a cross-correlation matrix that converges to a unit matrix is ​​constructed.

[0100] In this embodiment, the time-series divided signal is input into a time-domain signal encoder to obtain time-domain features, including:

[0101] S201: Input the time series sub-signal into a time domain signal encoder based on ResNet, and obtain the convolution time series sub-signal through a one-dimensional convolution block;

[0102] S202: Input the convolution time-series signal into the first-layer residual block to obtain a first soft threshold output feature; the first-layer residual block is a soft threshold module, including:

[0103] The convolution time series signal is input into the first convolution layer unit, and convolved through the convolution layer, batch normalization layer, and activation function layer connected in series to obtain the first output feature;

[0104] The first output feature is input into the second convolutional layer unit, and convolved through the convolutional layer, batch normalization layer, and activation function layer connected in series to obtain the second output feature;

[0105] The second output feature is input into the connection layer unit, and the threshold is calculated through the normalization layer, the adaptive averaging layer and the fully connected layer connected in series, so as to obtain the channel threshold of each channel of the second output feature;

[0106] Inputting the second output feature and the channel threshold of each channel corresponding to it into the adaptive threshold unit, scaling the second output feature, and obtaining an adaptive feature;

[0107] Perform residual connection on the adaptive features and the input features to obtain the first soft threshold output features;

[0108] S203: Inputting the first soft threshold output feature into the second layer residual block to obtain a second soft threshold output feature, where the second layer residual block is a soft threshold module;

[0109] S204: Let the second soft threshold output feature pass through the third convolution block and the fourth convolution block in sequence, and convolute the output time domain features corresponding to the time series division signal.

[0110] The second output feature and the channel threshold of each channel corresponding to it are input into the adaptive threshold unit, and the second output feature is scaled to obtain the adaptive feature, including:

[0111] Compare the second output feature Input channel characteristics of channels With preset channel threshold , calculate the output channel feature corresponding to the channel , expressed as:

[0112] ;

[0113] Obtaining adaptive features based on output channel features corresponding to all channels in the second output features;

[0114] in, Indicates the channel index and preset channel threshold Expressed as ; Indicates The scaling parameter of each channel is expressed as , Indicates Neuronal characteristics in each channel; Indicates that the width is , height is The feature map is averaged. Indicates The width of the channel is , height is The feature map of and Indicates the width and height of the feature map.

[0115] Specifically, performing continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain the corresponding frequency domain signal includes:

[0116] Use the preset mother wavelet to Timing signal Perform convolution to obtain the Timing signal The corresponding frequency domain signal , expressed as:

[0117] ;

[0118] in, is the preset scale parameter, To preset translation parameters, Represents the mother wavelet function The complex conjugate of .

[0119] In the field of intelligent mechanical fault diagnosis, wavelet transform is widely used in signal processing due to its good time-frequency localization characteristics. Compared with Fourier transform, wavelet transform can not only provide frequency domain information, but also retain time information, so that the extracted features are richer and more accurate. Through wavelet transform, the signal can be decomposed into sub-signals at different scales, each of which contains features within a specific frequency range; this multi-scale analysis method can effectively capture various features in the fault signal, which helps to improve the accuracy of fault diagnosis. Therefore, in the embodiment of the present invention, the multi-scale frequency domain features of wavelet transform are used to improve the accuracy of feature extraction.

[0120] In this embodiment, the frequency domain signal is input into a frequency domain signal encoder to obtain corresponding frequency domain features, including:

[0121] S301: Input the frequency domain signal into a frequency domain signal encoder based on ResNet, and obtain a convolution frequency domain signal through a two-dimensional convolution block;

[0122] S302: Input the convolution frequency domain signal into the first layer residual block to obtain a first attention feature; the first layer residual block is a residual network based on a multi-scale attention module, including:

[0123] After the convolution frequency domain signal passes through the convolution block, batch normalization layer, and activation function layer in series, feature grouping is performed to obtain height features, width features, and global features;

[0124] After local average pooling of the height feature and the width feature, the two outputs are concatenated to obtain the concatenated feature, and 1×1 convolution and activation are performed to obtain the activated height feature and the activated width feature;

[0125] After performing 3×3 convolution on the global features, global average pooling and activation function are performed in sequence to obtain activated global features;

[0126] Based on global features, activation height features, and activation width features, feature dimension reconstruction and feature group normalization are performed to obtain local normalized features;

[0127] The local normalized features are passed through the average pooling layer and the activation function layer in sequence, and then multiplied with the global features by the feature matrix to obtain the local features;

[0128] After the local normalized features are multiplied by the feature matrix and the activated global features are residually connected with the local features, they are sequentially reconstructed, 3×3 convolution and batch normalization to obtain the global normalized features;

[0129] After the global normalized feature is residually connected with the convolution frequency domain signal, the first attention feature is output through the activation function;

[0130] S303: inputting the first attention feature into a second-layer residual block, where the second-layer residual block is a residual network based on a multi-scale attention module, and outputting a second attention feature;

[0131] S304: Let the second attention feature pass through the third convolution block and the fourth convolution block in sequence, and convolute the output frequency domain features corresponding to the frequency domain signal.

[0132] Reference Figure 3 As shown, it is a residual network structure diagram based on a multi-scale attention module provided by the present invention; the present invention sets corresponding residual blocks in the time domain signal encoder and the frequency domain signal encoder based on ReaNet for feature extraction; in the time domain signal encoder, a soft threshold module is constructed, and an adaptive threshold unit is used to reduce the interference of noise on the extracted features and extract more accurate features; in the frequency domain signal encoder, a multi-scale attention mechanism is embedded, and after the multi-scale feature groups are normalized, the frequency domain features related to the time domain are extracted from different spatial directions, which improves the robustness of the frequency domain signal encoder and the accuracy of the representation of the frequency domain features. The present invention provides rich time-frequency features using wavelet transform, can automatically extract high-level features using a deep learning model, solves the problem of insufficient labeled data using transfer learning, and enhances the relevance and stability of features using the attention mechanism, constructs an efficient, accurate and robust target multi-scale intelligent decision model, realizes accurate diagnosis of faults, and ensures stable operation and efficient production of equipment.

[0133] Based on the above embodiments, in an embodiment of the present invention, fault diagnosis is performed based on the multi-scale intelligent decision-making method based on transfer learning proposed in the present invention. The method is based on transfer learning and deep learning, and can extract frequency domain features through wavelet transform and establish time-frequency domain connection through correlation matrix. The specific process includes:

[0134] S401: Process the original data, including dividing the data into features of corresponding lengths, dividing 99% of the data into unlabeled data for decision model pre-training, and performing subsequent fine-tuning operations on 1% of the data, as well as a series of time series signal data enhancement processing methods required when a large proportion of data is input into the model for training.

[0135] The timing signal enhancement processing method includes:

[0136] (1) Add Gaussian SNR noise: represents the original signal, then the generated signal Expressed as ,in represents Gaussian noise;

[0137] (2) Mask operation: Mask can be used to filter certain frequency components of the signal. Assume is the frequency domain representation of the signal, is a Mask function, then the Mask operation can be expressed as: ,in is the signal after applying Mask;

[0138] (3) Gain operation: The Gain operation is used to adjust the vibration value of the signal. , by randomizing the gain factor To adjust the vibration value, ;when When the signal is weakened, the vibration value is reduced. When , the signal vibration is enhanced, i.e. the vibration value is amplified;

[0139] (4) Translation, horizontal flip, and vertical flip.

[0140] S402: After processing the original data and enhancing the data, the time domain data is converted into frequency domain data using wavelet transform. For continuous non-stationary vibration time series signals, continuous wavelet transform is used to convert one-dimensional time series vibration data into higher-dimensional frequency domain data. Compared with ordinary Fourier transform, wavelet transform can analyze frequency domain features from more scales. By convolving the signal with a scalable and translated mother wavelet, the frequency components of the signal at different scales and time positions are analyzed to obtain the frequency domain signal. , expressed as:

[0141] ;

[0142] in, is the preset scale parameter, To preset translation parameters, Represents the mother wavelet function The complex conjugate of . The scale parameter Directly affects the compression and stretching of the wavelet function. When , the mother wavelet function mainly captures the low-frequency components of the signal. When , the mother wave function mainly captures the high-frequency components of the signal; the translation parameter Controls the position of the mother wavelet function on the time series axis, that is, at which moment of the signal is analyzed on the time axis.

[0143] S403: inputting the one-dimensional time domain features and the multi-scale two-dimensional frequency domain features into the model for pre-training;

[0144] S404: extracting the obtained one-dimensional encoder feature and connecting it to the classifier head after retaining the training parameters, and performing fine-tuning operation using a small amount of labeled data;

[0145] S405: Testing the robustness of the intelligent decision-making model using labeled data that has not participated in model training;

[0146] Specifically, in the extraction of time series features, a one-dimensional ResNet architecture is adopted. Different from the ordinary residual network architecture, this example uses a novel soft threshold module to process the features by performing a soft threshold operation on each feature channel to eliminate information related to noise. The function of the soft threshold is expressed as: ; The threshold of each channel can be expressed as , here , , Represents the index of the channel and the width and height of the feature map. , It is The neuron features in the channels, It is Scaling parameters. This threshold scaling module can make the overall residual network more suitable for extracting time series signal features, and highlight the role of intelligent decision-making. For general signal extraction algorithms, it is necessary to design filters for each different time series feature to convert noise information into zero features. However, the combination of deep learning and soft thresholding can realize automatic learning of gradient descent algorithms without relying on a lot of professional knowledge.

[0147] Specifically, for multi-scale frequency domain features, this embodiment makes related improvements to the attention mechanism based on the residual network, such as Figure 3 As shown; Different from the one-dimensional time series signal features, the frequency domain features after wavelet transformation need to pay more attention to the correlation between multi-scale features, and the extracted features need to be more limited to the information related to the fault features. Therefore, this embodiment adopts a multi-scale attention module on the frequency domain features and embeds it in the residual network framework; By dividing an input feature into different feature groups, these feature groups are feature encoded along two different spatial directions of height and width, and a The convolution kernel is used to extract a global feature, and finally these frequency domain features are aggregated to obtain effective information between different scales, thereby realizing the efficient extraction of frequency domain features of vibration signals.

[0148] As for the time domain features and frequency domain features, this embodiment does not adopt a simple feature fusion method. For the subsequent migration and fine-tuning operations, this embodiment adopts a method to enhance the correlation between the time and frequency domains. Specifically, the residual networks of the above two different architectures are used as encoders, and then the time domain features and frequency domain features are mapped to the low-dimensional space of the same dimension through linear transformation, which will generate time vectors and frequency vectors. Since the time domain features and frequency domain features are obtained through the same time series signal and the number of training times in the residual network is the same, the frequency domain features and time domain features are correlated, so the cross-correlation loss is used as the loss function of the overall intelligent decision pre-training model: ;in represents a constant that weighs the contribution of irrelevant terms, Represents the dimensions of the time domain and frequency domain feature vectors, i and j represent the index of the features in the dimension, It is the relationship matrix between time domain features and frequency domain features in the corresponding training batches, which can be expressed as: ; This embodiment ensures that the time domain vector and the frequency domain vector are within a comparable range through the Euclidean norm, ensuring that the matrix converges to a unit matrix.

[0149] Because deep learning, especially convolutional neural networks (CNN), has achieved great success in the fields of image processing and pattern recognition, it has also been introduced into mechanical fault diagnosis. CNN can automatically extract high-level features of input data, reducing the reliance on manual feature engineering. However, deep learning models usually require a large amount of labeled data for training, and in actual industrial environments, it is often difficult to obtain large-scale labeled data. Therefore, transfer learning has been introduced into the field of fault diagnosis. By fine-tuning the pre-trained model in the relevant field, better diagnostic results can be achieved with limited labeled data. Transfer learning can effectively solve the overfitting problem of deep learning models on small data sets. By pre-training a deep learning model on a large-scale data set and applying it to a specific fault diagnosis task, the generalization ability and diagnostic accuracy of the model can be significantly improved. For example, a CNN model pre-trained on a large-scale vibration signal data set can be used to transfer it to a specific mechanical equipment fault diagnosis task, so that excellent performance can be achieved with less labeled data. In addition, the attention mechanism has performed well in the fields of natural language processing and computer vision, and has also been introduced into mechanical fault diagnosis. The attention mechanism enables the model to focus on more important features by assigning different weights to different parts of the input data. In fault diagnosis, the attention mechanism can dynamically adjust the weight of each feature and enhance the focus on key fault features, thereby improving the accuracy and stability of diagnosis. The present invention comprehensively utilizes wavelet transform, deep learning, transfer learning and attention mechanism to build an efficient, accurate and robust intelligent mechanical fault diagnosis system. Wavelet transform provides rich time-frequency features, deep learning models can automatically extract high-level features, transfer learning solves the problem of insufficient labeled data, and attention mechanism enhances the relevance and stability of features. Through the combination of these advanced technologies, the intelligent fault diagnosis system can play a better role in actual industrial environments and ensure the stable operation and efficient production of equipment.

[0150] Based on the above embodiments, refer to Figure 4 As shown, it is a schematic diagram of the migration training process of the multi-scale intelligent decision-making method based on migration learning provided by this embodiment; the specific process is represented by pseudo code, including:

[0151] Input: segmented fault vibration timing signal , timing signal encoder , frequency domain signal encoder , time domain mapping transformation , frequency domain mapping transform , training rounds , batch size , Migrate Encoder , transfer learning classifier .

[0152] Output: , .

[0153] ① Dataset preprocessing:

[0154] Divide the data into unlabeled data with a large proportion Pre-training, with a small percentage of labeled data Fine-tune transfer learning and test data .

[0155] The pre-training data set is processed and converted into an enhanced time series signal and a frequency domain signal after continuous wavelet transformation.

[0156] ;

[0157] ;

[0158] ②Model pre-training:

[0159] 2.1: Randomly initialize the encoder , , mapping transformation , The weight of .

[0160] 2.2: For each round e:

[0161] Each batch size B:

[0162] Get the features extracted by the encoder from the time domain and frequency domain respectively:

[0163]

[0164]

[0165] Map the extracted high-dimensional features to a space of the same dimension:

[0166]

[0167]

[0168] Build the cross-correlation matrix:

[0169]

[0170] Calculate the loss:

[0171]

[0172] Update by backpropagating , , , Parameters;

[0173] ③ Transfer learning:

[0174] 3.1: Build the same model as the pre-trained temporal encoder;

[0175] 3.2: Pre-training The encoder weights are assigned to :

[0176] , is the weight parameter of the response;

[0177] 3.3: For each round e:

[0178] Each batch size B:

[0179] Get time series features:

[0180] Get the output category:

[0181] Train the classifier using cross entropy loss and fine-tune the encoder:

[0182]

[0183] Update by backpropagating , Parameters;

[0184] ④Test: Use the accuracy rate to calculate the model Achieve classification performance.

[0185] This embodiment uses the operation of continuous wavelet transform to obtain multi-scale frequency domain features, and through the idea of ​​mutual correlation between time domain and frequency domain, it does not use ordinary feature fusion operations, but uses the correlation matrix to enable the time domain feature extractor, so that it has the ability to extract frequency domain features, which reduces the amount of calculation and complexity compared to ordinary fusion algorithms. In addition, in order to obtain different complex signals, this embodiment also adopts a combination of various data enhancement methods to increase the complexity of the input time series signal, which can improve the generalization ability of the model to a certain extent. At the same time, this embodiment does not adopt the general feature extractor and classifier simultaneous migration operation, but only allows the encoder with strong extraction ability, that is, the feature extractor, to migrate parameters, so that the classifier can make intelligent decisions on the features to be classified without interference, thereby improving the generalization ability of the intelligent decision-making model.

[0186] Since most current decision models do not have good anti-noise interference capabilities and the ability to extract multi-scale frequency domain features, this embodiment adds an adaptive threshold module to the channel of the time domain model in the above framework. Based on the framework of the Resnet residual model, the model does not rely on professional knowledge analysis, and the model's ability to reduce noise interference on the extracted feature signals can be improved through training; although frequency domain features of more scales can be obtained through wavelet transform, this also means that more noise and irrelevant information are mixed in. In order to extract the frequency domain features that the model needs to pay attention to from each scale, this embodiment adds a feature attention mechanism. After normalizing the multi-scale feature groups, frequency domain features related to the time domain are extracted from different spatial directions, thereby improving the robustness of the model.

[0187] In summary, this embodiment combines the ability of deep learning to extract features with the knowledge transfer ability of transfer learning, which can solve the problems of intelligent decision-making and classification in actual industrial problems when data labels are insufficient, ensure the normal operation of actual industrial production, and reduce losses caused by mechanical abnormalities.

[0188] Based on the above embodiments, in an embodiment of the present invention, the present invention conducts experiments on two benchmark data sets: the American Mechanical Failure Prevention Technology Association (MFPT) data set and the Paderborn University (PU) bearing data set. The MFPT data set contains vibration signal data from different mechanical equipment, including rolling bearings, gearboxes, and motors. The data set records the vibration signals of the equipment under different operating conditions, including normal conditions and fault conditions. For the MFPT data set, three benchmark conditions (270 lb load, input shaft speed of 25HZ), three outer ring faults (25, 100, 200 lb loads, respectively) and three inner ring faults (0, 150, 300 lb loads) were selected as raw data for processing.

[0189] The PU bearing data set was tested under four working conditions: N15_M07_F10, N09_M07_F10, N15_M01_F10, and N15_M07_F04. N15 represents a speed of 1500rpm, and N09 represents a speed of 900rpm; M07 and M01 represent load torques of 0.7Nm and 0.1Nm; F04 and F10 represent radial forces of 1000N and 400N. This embodiment selects 9 state categories for operation under four working conditions. The description is shown in Table 1:

[0190] Table 1 Introduction to experimental datasets

[0191] Tag Code Fault type Fault area Fault type Damage length 0 healthy - - - 1 gap Outer ring Single point damage ≤ 2 mm 2 dent Outer ring Single point damage >2 mm 3 dent Outer ring Single point damage ≤ 2 mm 4 gap Outer ring More losses ≤ 2 mm 5 dent Outer ring + inner ring Single point damage >2 mm 6 dent Outer ring + inner ring More losses >4.5 mm 7 gap Outer ring + inner ring More losses ≤ 2 mm 8 dent Inner Circle Single point damage ≤ 2 mm 9 dent Inner Circle Single point damage >4.5 mm

[0192] This embodiment uses a commonly used and intuitive evaluation indicator - accuracy as the evaluation indicator. Accuracy is one of the most commonly used evaluation indicators in classification problems. It indicates the proportion of the number of samples correctly predicted by the model to the total number of samples. It is defined as: TP means the prediction is positive and the actual one is positive, the prediction is correct; FP means the prediction is positive and the actual one is negative, the prediction is wrong; FN means the prediction is negative and the actual one is positive, the prediction is wrong; TN means the prediction is negative and the actual one is negative, the prediction is correct. The accuracy measures the degree of match between the prediction result of the model on a minibatch of data and the true label.

[0193] The experimental environment may affect the experimental results. The experimental environment based on this embodiment is shown in Table 2:

[0194] Table 2 Experimental environment

[0195] Experimental environment Specification processor Inter(R) Core(TM) i7-6850K CPU @ 3.60GHZ Graphics NVIDIA TITAN Xp 12G Memory 128G DDR4 Development Language Python 3.9.7 Development System Ubuntu 21.10 Development Framework Pytorch 1.11.0 + cuda10.2

[0196] Similarly, parameter settings may also affect the experimental results. The detailed parameter settings of this experiment are shown in Table 3:

[0197] Table 3 Experimental parameter settings

[0198] parameter Specification Data length 1024 Pre-trained mapping dimensions 128 dimensions Continuous Wavelet Transform Scaling 64 Dimensions Continuous wavelet function Mexh Time Window 200 Initial learning rate 1e-3 Learning rate decay factor 5e-4 Pre-training rounds 500 Training batch size batch_size 128 Pre-training unlabeled dataset size 2000 Labeled data size 20 Validation set size 300 Test set size 600

[0199] In order to ensure the effectiveness of the data enhancement method adopted in this embodiment, firstly, different combinations of data enhancement modes are used on the MFPT dataset to verify the effectiveness of the data enhancement method. The accuracy of different data enhancement combinations is shown in Table 4:

[0200] Table 4 Accuracy of different data augmentation combinations

[0201] Enhancement Type Enhancement Methods Average accuracy 1 Horizontal, center, vertical flip + Gaussian noise 86.2% 2 Horizontal, center, vertical flip + Gaussian SNR noise 92.37% 3 Horizontal, center, vertical flip + Gaussian noise + Gaussian SNR noise + Mask + Gain + fade 94.97% 4 Horizontal and vertical flip + Gaussian SNR noise + Mask + Gain 98.55%

[0202] It can be seen from Table 4 that the effectiveness of the signal enhancement method proposed in this embodiment, that is, first flipping the signal horizontally and vertically, and then adding Gaussian signal-to-noise ratio noise, masking and adjusting the vibration value by randomizing the gain factor is better than the combination of other data enhancement methods.

[0203] After data enhancement, this embodiment compares the classification accuracy of various abnormal signals on the PU data set. The comparative test results are shown in Table 5:

[0204] Table 5 Comparative test results

[0205] Working conditions N15_M07_F10 N09_M07_F10 N15_M01_F10 N15_M07_F04 CNN_1D (Supervised Learning) 52.01% 40% 55.84% 58.59% Resnet18_1d (Supervised Learning) 44.17% 55.01% 75.59% 53.72% Alexnet_1d (Supervised Learning) 17.86% 20.14% 28.29% 28.54% WDCNN (Supervised Learning) 41.4% 39.76% 64.81% 68.99% DCA_BIGRU (Supervised Learning) 58.33% 50.20% 57% 64.04% CNN_2D (CWT, supervised learning) 29.54% 30.73% 30.75% 35.45% Resnet18_2D (CWT, supervised learning) 33.43% 35.08% 34.33% 38.26% Mean teacher (semi-supervised learning) 76.53% 68.73% 75.71% 82.74% SimSiam (Semi-supervised learning) 90.1% 92.24% 77.09% 85% Barlow twins (semi-supervised learning) 93.86% 80.2% 85.67% 90.8% CNN_1d (FFT, transfer learning) 93.39% 84.05% 96.03% 93.19% TFPred (FFT, transfer learning) 97.58% 96.2% 97.13% 97.75% Ours (CWT, transfer learning) 98.13% 99.02% 99.12% 98.25%

[0206] Based on the above experiments, it can be seen that: under the same labeled data and the same training rounds, the accuracy of the semi-supervised method is higher than that of supervised learning, and the accuracy of the method used in the present invention in semi-supervised learning, that is, the classification performance, is the best under the four working conditions; from the frequency domain conversion method and the training model, the feature extraction method of the continuous wavelet transform plus the attention mechanism and the adaptive threshold used in the present invention can better extract abnormal features than the general Fourier transform plus the ordinary deep learning model, thereby achieving the purpose of abnormal intelligent decision-making; if only the continuous wavelet transform is used and then the frequency domain features are obtained through the deep learning model, the purpose of intelligent decision-making cannot be achieved well, but after the transfer learning used by this method, the frequency domain features can be effectively extracted, thereby achieving the purpose of improving the accuracy of the model intelligent decision-making.

[0207] The multi-scale intelligent decision-making method based on transfer learning described in the present invention constructs a source domain training set containing multiple time-series sub-signals in the source domain based on the long time-series signal of mechanical fault vibration, and uses continuous wavelet transform to obtain the corresponding frequency domain signal; constructs a cross-correlation loss function by constructing a cross-correlation matrix between low-dimensional time-domain features and low-dimensional frequency-domain features based on the time-domain features and frequency-domain feature mapping of the time-series sub-signals, trains the source domain initial time-domain signal encoder and the source domain initial frequency-domain signal encoder, and obtains the source domain optimized time-domain signal encoder and the source domain optimized frequency-domain signal encoder; the present invention uses continuous wavelet transform to obtain multi-scale frequency-domain features, and uses the cross-correlation matrix to make the time domain and the frequency domain correlated with each other to train the encoder, so that the time-domain signal encoder has the ability to extract frequency-domain features, reduces the amount and complexity of time-domain and frequency-domain feature extraction operations, and improves the efficiency and accuracy of fault detection. The present invention transfers the optimized time domain signal encoder in the source domain to the time series encoder in the target domain, and forms a multi-scale intelligent decision model with the initial classifier; based on the predicted fault category label and the real label of the target domain time series signal, a cross entropy loss function is constructed to train the initialized multi-scale intelligent decision model, optimize the classifier, fine-tune the target domain time domain signal encoder, and obtain the target multi-scale intelligent decision model for fault diagnosis; the present invention transfers the time domain signal encoder obtained by using the source domain training set to the target domain through transfer learning, and fine-tunes it, thereby solving the problem of insufficient target domain label data and being more suitable for actual industrial environments. The present invention sets corresponding residual blocks in the time domain signal encoder and the frequency domain signal encoder based on ReaNet for feature extraction; in the time domain signal encoder, a soft threshold module is constructed, and an adaptive threshold unit is used to reduce the interference of noise on the extracted features and extract more accurate features; in the frequency domain signal encoder, a multi-scale attention mechanism is embedded, and after the multi-scale feature groups are normalized, the frequency domain features related to the time domain are extracted from different spatial directions, thereby improving the robustness of the frequency domain signal encoder and improving the accuracy of the representation of the frequency domain features. The present invention provides rich time-frequency features using wavelet transform, can automatically extract high-level features using deep learning models, solves the problem of insufficient labeled data using transfer learning, and enhances the relevance and stability of features using the attention mechanism. It constructs an efficient, accurate and robust target multi-scale intelligent decision-making model to achieve accurate diagnosis of faults, ensure stable operation and efficient production of equipment, and reduce losses caused by mechanical abnormalities.

[0208] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0209] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0210] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0211] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0212] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A multi-scale intelligent decision-making method based on transfer learning, characterized in that: include: Obtain the long time series signals of mechanical fault vibration in the source domain and the target domain, slice them according to the preset signal length, obtain multiple source domain time series sub-signals and target domain time series sub-signals, and construct the source domain training set and the target domain training set; Data enhancement is performed on each time-series sub-signal in the source domain training set to obtain the corresponding enhanced time-series sub-signal, which is input into the source domain initial time-domain signal encoder to obtain the corresponding time-domain features; Perform continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain the corresponding frequency domain signal, which is input into the source domain initial frequency domain signal encoder to obtain the corresponding frequency domain features; The time domain features and frequency domain features of each time series sub-signal are respectively input into the initial time domain mapping transformer and the initial frequency domain mapping transformer of the same dimension to obtain the corresponding low-dimensional time domain features and low-dimensional frequency domain features; Based on the Euclidean norm, the cross-correlation matrix between the low-dimensional time domain features and the low-dimensional frequency domain features of each time series sub-signal is calculated to construct a cross-correlation loss function; Based on the source domain training set and the cross-correlation loss function, the source domain initial time domain signal encoder and the source domain initial frequency domain signal encoder are trained until the cross-correlation loss function converges, thereby obtaining a source domain optimized time domain signal encoder and a source domain optimized frequency domain signal encoder; Initializing the source domain optimized time domain signal encoder as the target domain initial time domain signal encoder and the initial classifier to form an initial multi-scale intelligent decision model; Obtain the target domain time series sub-signal and its true label in the target domain training set, input the target domain time series sub-signal into the initial multi-scale intelligent decision model, and obtain the corresponding predicted fault category label; Based on the predicted fault category label and the true label of the target domain time series signal, a cross entropy loss function is constructed, and the initialized multi-scale intelligent decision model is trained until the cross entropy loss function converges to obtain the target multi-scale intelligent decision model; The long time-series vibration signal of mechanical fault collected in real time is segmented according to the preset signal length to obtain multiple time-series sub-signals, which are respectively input into the target multi-scale intelligent decision-making model to obtain the corresponding fault category.

2. The multi-scale intelligent decision-making method based on transfer learning according to claim 1 is characterized in that: The method performs data enhancement on each time series sub-signal in the source domain training set to obtain the corresponding enhanced time series sub-signal, including: adding a Gaussian signal-to-noise ratio to each time series sub-signal, performing a Mask operation, performing a Gain operation, data translation, horizontal flipping, vertical flipping, and center flipping; The adding of Gaussian signal-to-noise ratio comprises: Timing signal Add Gaussian SNR , get the corresponding noise-added time-series signal , expressed as: ; The Mask operation is performed, including obtaining the Timing signal Frequency domain representation of , using the Mask function Perform the Mask operation on the frequency domain representation of the time series signal to obtain the Mask signal, which is expressed as: ; The performing of the Gain operation includes adjusting the first The vibration value of each time-series divided signal is a preset vibration value.

3. The multi-scale intelligent decision-making method based on transfer learning according to claim 1 is characterized in that: The time series signal is input into the time domain signal encoder to obtain the time domain features, including: The time-series sub-signals are input into the time-domain signal encoder based on ResNet, and after passing through a one-dimensional convolution block, the convolution time-series sub-signals are obtained; The convolution time-series signal is input into the first-layer residual block to obtain the first soft threshold output feature; the first-layer residual block is a soft threshold module, including: The convolution time series signal is input into the first convolution layer unit, and convolved through the convolution layer, batch normalization layer, and activation function layer connected in series to obtain the first output feature; The first output feature is input into the second convolutional layer unit, and convolved through the convolutional layer, batch normalization layer, and activation function layer connected in series to obtain the second output feature; The second output feature is input into the connection layer unit, and the threshold is calculated through the normalization layer, the adaptive averaging layer and the fully connected layer connected in series, so as to obtain the channel threshold of each channel of the second output feature; Inputting the second output feature and the channel threshold of each channel corresponding to it into the adaptive threshold unit, scaling the second output feature, and obtaining an adaptive feature; Perform residual connection on the adaptive features and the input features to obtain the first soft threshold output features; Inputting the first soft threshold output feature into the second layer residual block to obtain the second soft threshold output feature, where the second layer residual block is a soft threshold module; The second soft threshold output feature is made to pass through the third convolution block and the fourth convolution block in sequence, and the convolution outputs the time domain feature corresponding to the time series sub-signal.

4. The multi-scale intelligent decision-making method based on transfer learning according to claim 3 is characterized in that: The step of inputting the second output feature and the channel threshold of each channel corresponding to the second output feature into the adaptive threshold unit, scaling the second output feature, and obtaining the adaptive feature includes: Compare the second output feature Input channel characteristics of channels With preset channel threshold , calculate the output channel feature corresponding to the channel , expressed as: ; Obtaining adaptive features based on output channel features corresponding to all channels in the second output features; in, Indicates the channel index and preset channel threshold Expressed as ; Indicates The scaling parameter of each channel is expressed as , Indicates Neuronal characteristics in each channel; Indicates that the width is , height is The feature map is averaged. Indicates The width of the channel is , height is The feature map of and Indicates the width and height of the feature map.

5. The multi-scale intelligent decision-making method based on transfer learning according to claim 1, characterized in that: The method performs continuous wavelet transform on each time-series sub-signal in the source domain training set to obtain the corresponding frequency domain signal, including: Use the preset mother wavelet to Timing signal Perform convolution to obtain the Timing signal The corresponding frequency domain signal , expressed as: ; in, is the preset scale parameter, To preset translation parameters, Represents the mother wavelet function The complex conjugate of .

6. The multi-scale intelligent decision-making method based on transfer learning according to claim 1, characterized in that: The frequency domain signal is input into the frequency domain signal encoder to obtain the corresponding frequency domain features, including: The frequency domain signal is input into the frequency domain signal encoder based on ResNet, and passes through the two-dimensional convolution block to obtain the convolution frequency domain signal; The convolution frequency domain signal is input into the first layer residual block to obtain the first attention feature; the first layer residual block is a residual network based on a multi-scale attention module, including: After the convolution frequency domain signal passes through the convolution block, batch normalization layer, and activation function layer in series, feature grouping is performed to obtain height features, width features, and global features; After local average pooling of the height feature and the width feature, the two outputs are concatenated to obtain the concatenated feature, and 1×1 convolution and activation are performed to obtain the activated height feature and the activated width feature; After performing 3×3 convolution on the global features, global average pooling and activation function are performed in sequence to obtain activated global features; Based on global features, activation height features, and activation width features, feature dimension reconstruction and feature group normalization are performed to obtain local normalized features; The local normalized features are passed through the average pooling layer and the activation function layer in sequence, and then multiplied with the global features by the feature matrix to obtain the local features; After the local normalized features are multiplied by the feature matrix and the activated global features are residually connected with the local features, they are sequentially reconstructed, 3×3 convolution and batch normalization to obtain the global normalized features; After the global normalized feature is residually connected with the convolution frequency domain signal, the first attention feature is output through the activation function; Input the first attention feature into the second layer residual block, where the second layer residual block is a residual network based on a multi-scale attention module and outputs a second attention feature; The second attention feature is passed through the third convolution block and the fourth convolution block in sequence, and the convolution outputs the frequency domain features corresponding to the frequency domain signal.

7. The multi-scale intelligent decision-making method based on transfer learning according to claim 1, characterized in that: The method of calculating the cross-correlation matrix between the low-dimensional time domain features and the low-dimensional frequency domain features of each time series sub-signal based on the Euclidean norm includes: ; in, Represents the dimensions of low-dimensional time domain features and low-dimensional frequency domain features, , and Represent the low-dimensional frequency domain features and their transpose and Euclidean norm respectively; and Represent the low-dimensional time domain features and their Euclidean norms respectively.

8. The multi-scale intelligent decision-making method based on transfer learning according to claim 7 is characterized in that: The cross-correlation loss function is expressed as: ; in, represents the cross-correlation matrix between the i-th dimension low-dimensional time domain features and the i-th dimension low-dimensional frequency domain features, represents the square of the cross-correlation matrix between the i-th dimension low-dimensional time domain feature and the j-th dimension low-dimensional frequency domain feature, A constant that represents the weighting of the contribution of unrelated terms.

9. The multi-scale intelligent decision-making method based on transfer learning according to claim 1, characterized in that: The predicted fault category label and the true label based on the target domain time series signal are used to construct a cross entropy loss function, which is expressed as: ; in, represents the cross entropy loss, Indicates the number of target domain timing sub-signals, Indicates the total number of fault categories; Indicates The true label of the target domain time series signal, Indicates The predicted fault category labels of target domain time series signals.

10. The multi-scale intelligent decision-making method based on transfer learning according to claim 1, characterized in that: The initializing multi-scale intelligent decision model is trained until the cross entropy loss function converges to obtain the target multi-scale intelligent decision model, including: Train the initial classifier, back-propagate and update the model parameters to obtain the target classifier; Train the target domain initial time domain signal encoder, and perform reverse transfer to fine-tune the parameters of the target domain initial time domain signal encoder to obtain the target domain optimized time domain signal encoder; Based on the target classifier and the target domain optimized time domain signal encoder, a target multi-scale intelligent decision-making model is formed.