Fault prediction method for cold and hot rolling host

Abstract

The invention provides a cold and hot rolling host fault prediction method based on multi-source field transfer learning and Transformer architecture aiming at the small sample problem of fault sample data class imbalance in cold and hot rolling host fault prediction, and introduces a weight adaptive updating strategy in the method to improve a multi-source field transfer learning strategy. According to the method, representation of unlabeled target domain data features is improved by learning source domain data, a sparse self-attention mechanism capable of learning a sparse graph structure is introduced, a standard Transform framework is improved in combination with an SVM classifier, target domain time sequence feature representation is enhanced, then the small sample problem of class imbalance is effectively solved, and the target domain time sequence feature representation is improved. And accurate and effective prediction of the faults of the cold and hot rolling hosts is realized.

Description

Technical Field

[0001] The present invention relates to a cold and hot rolling main machine fault prediction technology in industrial equipment, and in particular to a cold and hot rolling main machine fault prediction method based on multi-source domain transfer learning and Transformer architecture. Background Art

[0002] The cold and hot rolling main machines are the power sources of the cold and hot rolling mills respectively. Their rotation generates rolling torque to drive the rollers to roll the rolled products. Whether the main machine is stable and reliable directly affects the rolling accuracy and production efficiency of the rolling mill. The traditional methods for cold and hot rolling main machine fault prediction mainly include physical model method, statistical analysis method and expert system method. These methods can identify and predict motor faults to a certain extent, but there are problems such as it is difficult to accurately establish physical models, it is difficult to capture the dynamic process of fault development through statistical analysis of historical data, and the expert system relies on the human experience of experts and is difficult to adapt to new fault modes.

[0003] In recent years, with the development of artificial intelligence and big data technology, motor fault prediction methods based on deep learning have received widespread attention, among which Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) are mainstream algorithms. Compared with traditional methods, these algorithms classify motor faults by learning the historical operation data of rolling mill motors, which improves the intelligence and efficiency of fault prediction. Since there are more normal samples and fewer fault samples in actual operation, the operation monitoring data of the main machine is class imbalanced data or even unlabeled data. If the class imbalanced data is directly used for classification, a biased classifier will be obtained (for example, there are 95 normal samples with a classification label of "1" and 5 fault samples with a classification label of "0". If the classifier always outputs the class label "1", the accuracy will reach 95%). The cost of the classification error rate of fault samples is much higher than that of normal samples, and the biased prediction results cannot provide accurate guidance for decision makers and cannot achieve the purpose of predictive maintenance. However, LSTM and CNN have some limitations when processing class imbalanced data and unlabeled data. Therefore, an effective method is needed to solve these problems and improve the accuracy and reliability of fault prediction. Summary of the invention

[0004] Aiming at the problem of imbalanced fault sample data classes and even unlabeled fault sample data in cold and hot rolling main machine fault prediction, the present invention provides a cold and hot rolling main machine fault prediction method to achieve accurate and effective prediction of cold and hot rolling main machine faults.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting failure of a hot and cold rolling main machine comprises the following steps:

[0007] Step 1: Take the historical fault data of the host in the same production line as the target host and the historical fault data of the host in other production lines with the same production process as the source domain, and the real-time operation data of the target host as the target domain; collect the vibration data, current data and temperature data of the hosts in the source domain and the target domain; label the vibration data, current data and temperature data of the host in the source domain to obtain the source domain data set, and the vibration data, current data and temperature data of the host in the target domain constitute the target domain data set;

[0008] Step 2: Convert the vibration data from the time domain to the frequency domain to obtain vibration frequency domain data, and perform data preprocessing on the vibration frequency domain data, current data, and temperature data, including removing noise and outliers, filling missing values, and standardizing data; obtain the preprocessed source domain data set and target domain data set;

[0009] Step 3: Use the MMD distance with adaptive weight assignment to optimize the multi-source domain transfer learning model on the preprocessed source domain data set, integrate the cross entropy loss on the source domain and the MMD distance regularization term to obtain the transfer learning function, use the transfer learning function to perform transfer learning on the target domain data, and obtain the feature representation of the target domain data after transfer learning;

[0010] Step 4: Build a fault prediction model integrating Transformer and SVM, and use the preprocessed target domain data set and the feature representation of the target domain data after transfer learning to train the fault prediction model integrating Transformer and SVM;

[0011] Step 5: Input the real-time operation data of the new target host into the trained fault prediction model integrating Transformer and SVM to obtain the classification result of host fault prediction.

[0012] To optimize the above technical solutions, the specific measures taken also include:

[0013] Furthermore, in step 1, the source domain dataset is represented as ,in Indicates A sample pair consisting of source domain data and its label, is the number of source domain data; Indicates The source domain data is a vector composed of vibration data, current data and temperature data. yes Corresponding labels, the target domain dataset is represented as ,in, Indicates the amount of target domain data, For the The target domain data is a vector composed of vibration data, current data and temperature data, and has no corresponding labels.

[0014] Furthermore, in step 3, the MMD distance with adaptive weight assignment is used on the preprocessed source domain data set to optimize the multi-source domain transfer learning model, and the cross entropy loss and MMD distance regularization term on the source domain are integrated to obtain the specific process of the transfer learning function:

[0015] The calculation formula of the MMD distance with adaptive weight assignment is:

[0016] ;

[0017] In the formula, The MMD distance for assigning adaptive weights, represents the source domain; represents the target domain, is the number of source domain data, is the source domain dataset, and represents the i-th source domain data and the j-th source domain data in the source domain dataset, is the Gaussian kernel function, indicating and The distance between is the number of target domain data, is the target domain dataset, For the Target domain data, is the jth target domain data, is the Gaussian kernel function, indicating and The distance between is the tth iteration Target domain data The weight of

[0018] The expression of the transfer learning function is as follows:

[0019] ;

[0020] The constraints are:

[0021] ;

[0022] In the formula, is the transfer learning function, is the weight of the i-th source domain data, which is used to measure the importance of source domain data in the transfer learning process. is the cross entropy loss function on the i-th source domain data, is the SVM classifier, is the high-dimensional representation of the source domain data obtained by the feature extractor, It is The feature matrix of source domain data, is the parameter mapped by the SVM kernel function, It is The label vector of the source domain data, is the weight of the MMD distance regularization term;

[0023] In step 3, the process of using the transfer learning function to perform transfer learning on the target domain data to obtain the feature representation of the target domain data after transfer learning is expressed by the formula:

[0024] ;

[0025] In the formula, is the feature representation of the target domain data after transfer learning, is the target domain dataset, express The ath feature of Indicates the number of target domain features.

[0026] Further, in step 4, the fault prediction model integrating Transformer and SVM includes an input layer, an embedding layer, a position encoding layer, an encoder and an SVM classifier;

[0027] The encoder is improved based on the sparse self-attention mechanism of the learnable sparse graph structure, and is specifically composed of multiple identical layers stacked, each layer consisting of a multi-head sparse self-attention sublayer and a feedforward fully connected network sublayer, the multi-head sparse self-attention sublayer and the feedforward fully connected network sublayer are connected through residual connection and normalization operation, and the multi-head attention layer is used to obtain a weighted value matrix , the feed-forward fully connected sublayer is linearly transformed through the fully connected layer.

[0028] Furthermore, in step 4,

[0029] The method of training the fault prediction model integrating Transformer and SVM by using the feature representation of the preprocessed target domain data set and the target domain data after transfer learning specifically includes:

[0030] Step 4.1, initialize Transformer hyperparameters and SVM classifier parameters;

[0031] Step 4.2: Preprocess the target domain dataset Before Continuous sampling is performed to obtain the sampled target domain data sequence , the input layer receives the sampled target domain data sequence And the feature representation of the target domain data after transfer learning , concatenate and fuse the two to form the input tensor :

[0032] ;

[0033] In the formula, Express and Perform splicing and fusion;

[0034] Step 4.3: Input tensor Input the embedding layer and get the output of the embedding layer for:

[0035] ;

[0036] in, is the weight matrix of the embedding layer;

[0037] Step 4.4: Output of the Embedding Layer Enter the position encoding layer to add position information to each input vector. The formula is as follows:

[0038] ;

[0039] in, is the position index, is the dimension index, It means that the sine function is used to give the Input the added location information, is the output of the embedding layer The dimension of the position encoding layer is:

[0040]

[0041] in, is the input to the position encoding layer The new position information after adding the position information is used as the output of the position encoding layer. The input of the position encoding layer Original location information; is the input to the position encoding layer Added location information;

[0042] Step 4.5: In the encoder, the output of the position encoding layer Convert to nodes and introduce a learnable sparse adjacency matrix , a sparse adjacency matrix Elements Representation Node and nodes The connection strength between them is the sparse adjacency matrix With the query matrix and key matrix Multiply the dot product results of to obtain a sparse attention score matrix. The score matrix of an attention head is expressed as :

[0043] ;

[0044] in, Represents element-wise multiplication; is the dimension of the key; represents the softmax function, which is used to normalize the attention score into a probability distribution;

[0045] For each attention head, the attention score matrix Sum Matrix Multiply and linearly transform to get its attention output:

[0046] ;

[0047] In the formula, is the attention output of the i-th attention head;

[0048] Concatenate the outputs of all attention heads to get the final attention output :

[0049] ;

[0050] Where E is the identity matrix, is the output weight matrix, Indicates splicing;

[0051] A two-layer feedforward neural network is used to obtain the output of the encoder as the time-enhanced feature representation H:

[0052] ;

[0053] in, is the weight matrix of a two-layer feedforward neural network, is the bias term,

[0054] Step 4.6: Convert the time-enhanced feature representation H into classification features as the input of the SVM classifier to train the SVM classifier with the goal of minimizing the value of the loss function. Backpropagate the loss function value and update the Transformer parameters until the loss converges.

[0055] Furthermore, in step 4.6, the loss function is specifically:

[0056] ;

[0057] Where L represents the total loss, , and is the balance coefficient used to balance the weights between sparsity regularization, self-supervision loss, and SVM classification loss. is the sparsity regularization term, is the self-supervisory loss term, and is a positive sample pair, yes The positive sample of , , , Respectively time, Moment and The target host operation sample data collected at all times; Express , Take the cosine distance to find the similarity between the two; , They are the weight vector and bias term of the SVM classifier respectively; is the feature representation after time enhancement No. elements; yes The target host is running sample data at this moment The true label of is the regularization parameter of the SVM classifier, and N is the number of samples of the target host running sample data processed by the SVM classifier.

[0058] The beneficial effects of the present invention are:

[0059] (1) Introducing an adaptive weight assignment method to calculate the MMD distance can improve the representation of target domain features by multi-source domain transfer learning, and can be applicable to situations where the cold / hot rolling main machine monitoring data has few or no labels; (2) By introducing a sparse self-attention mechanism with a learnable sparse graph structure, the feature representation of time series data can be enhanced, and at the same time, the computational cost can be greatly reduced, thereby significantly improving the accuracy and efficiency of cold / hot rolling main machine fault prediction; (3) The data collected by the present invention is relatively rich, and the fault result classification includes no fault, bearing wear fault, bearing damage fault, bearing overheating fault, motor rotor fault, and motor stator fault, which is relatively precise; (4) It can be widely used in the field of fault prediction of other similar industrial equipment, and has broad application prospects and important economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of the cold and hot rolling main machine failure prediction method proposed by the present invention.

[0061] Figure 2 This is a schematic diagram of the implementation principle of the cold and hot rolling main machine fault prediction method proposed in the present invention.

[0062] Figure 3 This is the characteristic diagram of vibration data after verification sample preprocessing.

[0063] Figure 4 This is the characteristic diagram of the current data after verification sample preprocessing.

[0064] Figure 5 This is the temperature data characteristic diagram after verification sample preprocessing. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0066] Embodiment 1

[0067] The present invention proposes a cold and hot rolling main machine fault prediction method based on multi-source domain transfer learning and Transformer architecture. The overall process of the method is as follows: Figure 1 As shown, the implementation principle of this method is as follows Figure 2 As shown, the following steps are included:

[0068] Step 1: In the hot strip rolling process, the historical fault data of the host in the same production line as the target host and the historical fault data of the host in other production lines with the same production process are used as the source domain, and the real-time operation data of the target host is used as the target domain; for the hosts in the source and target domains, acceleration sensors are installed on their bearing seats to collect their vibration data, current sensors are installed on their power lines to collect their current data, and temperature sensors are installed on their bearings to collect their temperature data. The vibration data, current data, and temperature data of the host in the source domain are labeled to obtain the source domain data set, and the vibration data, current data, and temperature data of the host in the target domain constitute the target domain data set;

[0069] The source domain dataset is represented as ,in Indicates A sample pair consisting of source domain data and its label, is the number of source domain data; Indicates The source domain data is a vector composed of vibration data, current data and temperature data. yes The corresponding labels, the target domain dataset is represented as ,in, Indicates the amount of target domain data, For the The target domain data is a vector composed of vibration data, current data and temperature data, and has no corresponding labels.

[0070] Step 2: Convert the vibration data from the time domain to the frequency domain through Fourier transform to obtain vibration frequency domain data, and perform data preprocessing on the vibration frequency domain data, current data, and temperature data, including removing noise and outliers, filling missing values, and standardizing data; obtain the preprocessed source domain data set and target domain data set.

[0071] For vibration frequency domain data, current data, and temperature data, a low-pass filter is used to remove high-frequency noise in the data, a threshold is set to remove outliers, the mean interpolation method is used to fill in missing data, and the maximum and minimum method is used for standardization.

[0072] Specifically, the Fourier transform of the time domain signal into the frequency domain signal is:

[0073]

[0074] in, is the converted vibration frequency domain signal; is the vibration time domain signal; is an imaginary unit, satisfying ; is a negative exponential function used to map the time domain signal to the frequency domain. Indicates the vibration frequency.

[0075] The low-pass filtering method is to use Keep the data when greater than Remove this data when:

[0076]

[0077] in, is a low-pass filter, is the cutoff frequency, is the vibration frequency, where Take 2000Hz.

[0078] The method for removing outliers with a set threshold is as follows: for vibration frequency domain data, current data, and temperature data, calculate the mean of each data separately. and variance , the threshold is , those outside this range are considered as outliers and removed.

[0079] The three data are standardized using the following methods:

[0080]

[0081] in, Indicates the original data The value of a feature, For the corresponding The standardized value, and Indicates the original data The maximum and minimum values ​​of the features.

[0082] Step 3: Use the MMD distance with adaptive weight assignment to optimize the multi-source domain transfer learning model on the preprocessed source domain dataset, integrate the cross entropy loss on the source domain and the MMD distance regularization term to obtain the transfer learning function, and use the transfer learning function to perform transfer learning on the target domain data to obtain the feature representation of the target domain data after transfer learning.

[0083] The calculation formula of the MMD distance with adaptive weight assignment is:

[0084] ;

[0085] In the formula, The MMD distance for assigning adaptive weights, represents the source domain; represents the target domain, is the number of source domain data, is the source domain dataset, and represents the i-th source domain data and the j-th source domain data in the source domain dataset, is the Gaussian kernel function, indicating and The distance between is the number of target domain data, is the target domain dataset, For the Target domain data, is the jth target domain data, is the Gaussian kernel function, indicating and The distance between is the tth iteration Target domain data The weight of .

[0086] The expression of the transfer learning function is as follows:

[0087] ;

[0088] The constraints are:

[0089] ;

[0090] In the formula, is the transfer learning function, is the weight of the i-th source domain data, which is used to measure the importance of source domain data in the transfer learning process. is the cross entropy loss function on the i-th source domain data, is the SVM classifier, is the high-dimensional representation of the source domain data obtained by the feature extractor, It is The feature matrix of source domain data, is the parameter mapped by the SVM kernel function, It is The label vector of the source domain data, is the weight of the MMD distance regularization term.

[0091] The process of using the transfer learning function to transfer the target domain data to obtain the feature representation of the target domain data after transfer learning is expressed by the formula:

[0092] ;

[0093] In the formula, is the feature representation of the target domain data after transfer learning, is the target domain dataset, express The ath feature of Indicates the number of target domain features.

[0094] Step 4: Build a fault prediction model that integrates Transformer and SVM, and use the preprocessed target domain data set and the feature representation of the target domain data after transfer learning to train the fault prediction model that integrates Transformer and SVM.

[0095] The standard Transformer architecture (consisting of an encoder and a decoder) is extended by adding an embedding layer and a position encoding layer. A sparse self-attention mechanism is designed based on a learnable sparse graph structure to improve the encoder layer, forming a time-enhanced model based on the Transformer architecture to better capture the temporal dependencies in time series data.

[0096] The fault prediction model integrating Transformer and SVM includes input layer, embedding layer, position encoding layer, encoder and SVM classifier;

[0097] The encoder is improved based on the sparse self-attention mechanism of the learnable sparse graph structure, and is specifically composed of multiple identical layers stacked. This embodiment takes 6 stacked layers, each of which is composed of a multi-head sparse self-attention sublayer and a feedforward fully connected network sublayer. The feedforward fully connected network sublayer is composed of 2 hidden layers. The number of neurons in the hidden layer is set to 1024, and the ReLU activation function is used. The multi-head sparse self-attention sublayer and the feedforward fully connected network sublayer are connected through residual connection and normalization operation. Specifically, the output of each sublayer is first added to the input of the sublayer (i.e., residual connection), and then the layer normalization operation is performed. The multi-head attention layer is used to obtain the weighted value matrix , the feed-forward fully connected sublayer is linearly transformed through the fully connected layer.

[0098] The fault prediction model that integrates Transformer and SVM is trained by using the preprocessed target domain dataset and the feature representation of the target domain data after transfer learning. Specifically, the following are included:

[0099] Step 4.1: Initialize Transformer hyperparameters and SVM classifier parameters; as shown in Table 1.

[0100] Table 1 Model initialization parameter values

[0101]

[0102] Step 4.2: Preprocess the target domain dataset Before Continuous sampling is performed to obtain the sampled target domain data sequence , the input layer receives the sampled target domain data sequence And the feature representation of the target domain data after transfer learning , concatenate and fuse the two to form the input tensor :

[0103] ;

[0104] In the formula, Express and Perform splicing and fusion;

[0105] Step 4.3: Input tensor Input the embedding layer and get the output of the embedding layer for:

[0106] ;

[0107] in, is the weight matrix of the embedding layer;

[0108] Step 4.4: Output of the Embedding Layer Enter the position encoding layer to add position information to each input vector. The formula is as follows:

[0109] ;

[0110] in, is the position index, is the dimension index, It means that the sine function is used to give the Input the added location information, is the output of the embedding layer Dimensions;

[0111] The output of the position encoding layer is:

[0112]

[0113] in, is the input to the position encoding layer The new position information after adding the position information is used as the output of the position encoding layer. The input of the position encoding layer Original location information; is the input to the position encoding layer Added location information;

[0114] Step 4.5: In the encoder, the output of the position encoding layer Convert to nodes and introduce a learnable sparse adjacency matrix , a sparse adjacency matrix Elements Representation Node and nodes The connection strength between them is the sparse adjacency matrix With the query matrix and key matrix Multiply the dot product results of to obtain a sparse attention score matrix. The score matrix of an attention head is expressed as :

[0115] ;

[0116] in, Represents element-wise multiplication; is the dimension of the key; represents the softmax function, which is used to normalize the attention score into a probability distribution;

[0117] For each attention head, the attention score matrix Sum Matrix Multiply and linearly transform to get its attention output:

[0118] ;

[0119] In the formula, is the attention output of the i-th attention head;

[0120] Concatenate the outputs of all attention heads to get the final attention output :

[0121] ;

[0122] Where E is the identity matrix, is the output weight matrix, Indicates splicing;

[0123] A two-layer feedforward neural network is used to obtain the output of the encoder as the time-enhanced feature representation H:

[0124] ;

[0125] in, is the weight matrix of a two-layer feedforward neural network, is the bias term,

[0126] Step 4.6: Convert the time-enhanced feature representation H into a classification feature as the input of the SVM classifier to train the SVM classifier with the goal of minimizing the loss function value. Backpropagate the loss function value and update the Transformer parameters until the loss converges. The loss function is specifically:

[0127] ;

[0128] Where L represents the total loss, , and is the balance coefficient used to balance the weights between sparsity regularization, self-supervision loss, and SVM classification loss. is the sparsity regularization term, is the self-supervisory loss term, and is a positive sample pair, yes The positive sample of , , , Respectively time, Moment and The target host operation sample data collected at all times; Express , Take the cosine distance to find the similarity between the two; , They are the weight vector and bias term of the SVM classifier respectively; is the feature representation after time enhancement No. elements; yes The target host is running sample data at this moment The true label of is the regularization parameter of the SVM classifier, and N is the number of samples of the target host running sample data processed by the SVM classifier.

[0129] Step 5: Input the real-time operation data of the new target host into the trained fault prediction model integrating Transformer and SVM to obtain the classification results of the host fault prediction. The fault classification results include no fault, bearing wear fault, bearing damage fault, bearing overheating fault, motor rotor fault and motor stator fault.

[0130] The fault diagnosis results at 10 moments are taken to verify the prediction performance of the model, with accuracy, precision, recall, and F1 value as evaluation indicators. The calculation of each indicator is:

[0131]

[0132]

[0133]

[0134]

[0135] Among them, TP represents the number of samples predicted as positive; TN represents the number of samples predicted as negative; FP represents the number of samples predicted as positive, i.e., false positives; FN represents the number of samples predicted as negative, i.e., missed negatives. is the total number of samples, which is set to 10.

[0136] The obtained index values ​​are shown in Table 2. It can be seen that the prediction accuracy of the model is 93.6%, which is a significant effect.

[0137] Table 2 Model prediction performance evaluation index values

[0138]

[0139] like Figure 3 , Figure 4 and Figure 5 They are the vibration data characteristic diagram, current data characteristic diagram and temperature data characteristic diagram after preprocessing of the verification sample, among which the 10th sample is identified as a bearing wear fault by the proposed model. It can be seen that the vibration main frequency, current mean and peak value, temperature mean and peak value in the 10th sample are all high, which indirectly explains the result of bearing wear.

[0140] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0141] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for predicting failure of hot and cold rolling mill main machine, characterized in that: The following steps are involved: Step 1: Take the historical fault data of the host in the same production line as the target host and the historical fault data of the host in other production lines with the same production process as the source domain, and the real-time operation data of the target host as the target domain; collect the vibration data, current data and temperature data of the hosts in the source domain and the target domain; label the vibration data, current data and temperature data of the host in the source domain to obtain the source domain data set, and the vibration data, current data and temperature data of the host in the target domain constitute the target domain data set; Step 2: Convert the vibration data from the time domain to the frequency domain to obtain vibration frequency domain data, and perform data preprocessing on the vibration frequency domain data, current data, and temperature data, including removing noise and outliers, filling missing values, and data standardization; Obtain the preprocessed source domain dataset and target domain dataset; Step 3: Use the MMD distance with adaptive weight assignment to optimize the multi-source domain transfer learning model on the preprocessed source domain data set, integrate the cross entropy loss on the source domain and the MMD distance regularization term to obtain the transfer learning function, use the transfer learning function to perform transfer learning on the target domain data, and obtain the feature representation of the target domain data after transfer learning; Step 4: Build a fault prediction model integrating Transformer and SVM, and use the preprocessed target domain data set and the feature representation of the target domain data after transfer learning to train the fault prediction model integrating Transformer and SVM; Step 5: Input the real-time operation data of the new target host into the trained fault prediction model integrating Transformer and SVM to obtain the classification result of host fault prediction.

2. The method for predicting failure of hot and cold rolling mill main machine according to claim 1, characterized in that: In step 1, the source domain dataset is represented as ,in Indicates A sample pair consisting of source domain data and its label, is the number of source domain data; Indicates The source domain data is a vector composed of vibration data, current data and temperature data. yes Corresponding labels, the target domain dataset is represented as ,in, Indicates the amount of target domain data, For the The target domain data is a vector composed of vibration data, current data and temperature data, and has no corresponding labels.

3. The method for predicting failure of hot and cold rolling mill main machine according to claim 1, characterized in that: In step 3, the specific process of using the MMD distance with adaptive weight assignment to optimize the multi-source domain transfer learning model on the preprocessed source domain data set, integrating the cross entropy loss on the source domain and the MMD distance regularization term, and obtaining the transfer learning function is: The calculation formula of the MMD distance with adaptive weight assignment is: ； In the formula, The MMD distance for assigning adaptive weights, represents the source domain; represents the target domain, is the number of source domain data, is the source domain dataset, and Represents the first source domain data and j-th source domain data, is the Gaussian kernel function, indicating and The distance between is the number of target domain data, is the target domain dataset, For the Target domain data, is the jth target domain data, is the Gaussian kernel function, indicating and The distance between is the tth iteration Target domain data The weight of The expression of the transfer learning function is as follows: ； The constraints are: ； In the formula, is the transfer learning function, It is The weight of the source domain data is used to measure the importance of the source domain data in the transfer learning process. For the The cross entropy loss function on the source domain data is: is the SVM classifier, is the high-dimensional representation of the source domain data obtained by the feature extractor, It is The feature matrix of source domain data, is the parameter mapped by the SVM kernel function, It is The label vector of the source domain data, is the weight of the MMD distance regularization term; In step 3, the process of using the transfer learning function to perform transfer learning on the target domain data to obtain the feature representation of the target domain data after transfer learning is expressed by the formula: ； In the formula, is the feature representation of the target domain data after transfer learning, is the target domain dataset, express The ath feature of Indicates the number of target domain features.

4. The method for predicting failure of hot and cold rolling mill main machine according to claim 1, characterized in that: In step 4, the fault prediction model integrating Transformer and SVM includes an input layer, an embedding layer, a position encoding layer, an encoder and an SVM classifier; The encoder is improved based on the sparse self-attention mechanism of the learnable sparse graph structure, and is specifically composed of multiple identical layers stacked, each layer consists of a multi-head sparse self-attention sublayer and a feedforward fully connected network sublayer, the multi-head sparse self-attention sublayer and the feedforward fully connected network sublayer are connected through residual connection and normalization operation, and the multi-head attention layer is used to obtain a weighted value matrix , the feed-forward fully connected sublayer is linearly transformed through the fully connected layer.

5. The method for predicting failure of hot and cold rolling mill main machine according to claim 4, characterized in that: In step 4, The method of training the fault prediction model integrating Transformer and SVM by using the feature representation of the preprocessed target domain data set and the target domain data after transfer learning specifically includes: Step 4.1, initialize Transformer hyperparameters and SVM classifier parameters; Step 4.2: Preprocess the target domain dataset Before Continuous sampling is performed to obtain the sampled target domain data sequence , the input layer receives the sampled target domain data sequence And the feature representation of the target domain data after transfer learning , concatenate and fuse the two to form the input tensor : ； In the formula, Express and Perform splicing and fusion; Step 4.3: Input tensor Input the embedding layer and get the output of the embedding layer for: ； in, is the weight matrix of the embedding layer; Step 4.4: Output of the Embedding Layer Enter the position encoding layer to add position information to each input vector. The formula is as follows: ； in, is the position index, is the dimension index, It means that the sine function is used to give the Input the added location information, is the output of the embedding layer Dimensions; The output of the position encoding layer is: ； in, is the input to the position encoding layer The new position information after adding the position information is used as the output of the position encoding layer. The input of the position encoding layer Original location information; is the input to the position encoding layer Added location information; Step 4.5: In the encoder, the output of the position encoding layer Convert to nodes and introduce a learnable sparse adjacency matrix , a sparse adjacency matrix Elements Representation Node and nodes The connection strength between them is the sparse adjacency matrix With the query matrix and key matrix Multiply the dot product results of to obtain a sparse attention score matrix. The score matrix of an attention head is expressed as : ； in, Represents element-wise multiplication; is the dimension of the key; represents the softmax function, which is used to normalize the attention score into a probability distribution; For each attention head, the attention score matrix Sum Matrix Multiply and linearly transform to get its attention output: ； In the formula, It is The attention output of each attention head; Concatenate the outputs of all attention heads to get the final attention output : ； Where E is the identity matrix, is the output weight matrix, Indicates splicing; A two-layer feedforward neural network is used to obtain the output of the encoder as the time-enhanced feature representation H: ； in, is the weight matrix of a two-layer feedforward neural network, is the bias term, Step 4.6: Convert the time-enhanced feature representation H into classification features as the input of the SVM classifier to train the SVM classifier with the goal of minimizing the value of the loss function. Backpropagate the loss function value and update the Transformer parameters until the loss converges.

6. The method for predicting failure of hot and cold rolling mill main machine according to claim 4, characterized in that: In step 4.6, the loss function is specifically: ； Where L represents the total loss, , and is the balance coefficient used to balance the weights between sparsity regularization, self-supervision loss, and SVM classification loss. is the sparsity regularization term, is the self-supervisory loss term, and is a positive sample pair, yes The positive sample of , , , Respectively time, Moment and The target host operation sample data collected at all times; Express , Take the cosine distance to find the similarity between the two; , They are the weight vector and bias term of the SVM classifier respectively; is the feature representation after time enhancement No. elements; yes The target host is running sample data at this moment The true label of is the regularization parameter of the SVM classifier, and N is the number of samples of the target host running sample data processed by the SVM classifier.

Images