Hot rolling process composite fault tracing method under variable working conditions

Abstract

The invention is suitable for the technical field of industrial process fault diagnosis, and relates to a hot rolling process composite fault tracing method under a variable working condition, which comprises the following steps: S10, fusing the generality and personality characteristics of a composite fault under the variable working condition by adopting multi-task learning and deep learning; s20, on the basis of extraction and fusion of generalization and personality characteristics of the composite fault under variable working conditions, adopting transfer learning and a convolutional neural network to realize intelligent diagnosis of the composite fault of working condition transfer and generalization; and S30, adopting an ensemble learning and transfer learning method to realize compound fault accurate tracing under variable working conditions. The method is simple in process and convenient to operate, and the accuracy of compound fault tracing in the hot rolling process under the variable working conditions is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial process fault diagnosis, and particularly relates to a hot rolling process compound fault tracing method under variable working conditions. BACKGROUND

[0002] Under the influence of raw materials, product demand, external environment, equipment state change and control action compensation and other factors, the hot rolling process may run under different steady and non-steady working conditions, and may also dynamically switch between different running working conditions. Facing the compound faults with characteristics such as randomness, concurrency, concealment and coupling, there may be a large difference between the distribution characteristics of fault data under a single working condition and the distribution characteristics of fault data under an actual working condition, especially the compound faults under extreme working conditions such as catastrophic accidents. The compound fault characteristics of the hot rolling process under variable working conditions are often similar but have large differences, making it challenging to trace the compound faults of the hot rolling process under variable working conditions.

[0003] Therefore, how to provide a hot rolling process compound fault tracing method under variable working conditions with high accuracy is a problem to be solved by those skilled in the art. SUMMARY

[0004] In view of the deficiencies of the prior art, the purpose of the present application is to provide a hot rolling process compound fault tracing method under variable working conditions to solve the problem of low accuracy of tracing the compound faults of the hot rolling process under variable working conditions in the prior art.

[0005] In order to solve the above technical problems, the application adopts the following technical solutions:

[0006] The application provides a hot rolling process compound fault tracing method under variable working conditions, comprising the following steps:

[0007] S10, adopt multi-task learning and deep learning to fuse the common and individual characteristics of compound faults under variable working conditions;

[0008] S20, on the basis of extracting and fusing the common and individual characteristics of compound faults under variable working conditions, adopt transfer learning and convolutional neural network to realize intelligent diagnosis of compound faults under working condition transfer and generalization;

[0009] S30, adopt ensemble learning and transfer learning method to realize accurate tracing of compound faults under variable working conditions.

[0010] Further, in the S10, according to the thermal rolling process data characteristics under each single working condition, the composite fault features under each single working condition are extracted by using the convolutional neural network, the variational mode decomposition, the bidirectional gated recurrent unit and the bidirectional minimum gated unit method, the composite fault feature vectors are obtained, the composite fault individual features and common features under each single working condition are obtained by using the individual encoder composed of the feedforward neural network and the common encoder composed of the multi-task learning method, the common features are subjected to common constraint to reduce the difference in spatial distribution of multiple feature vectors, and the composite fault common and individual features under variable working conditions are extracted.

[0011] Further, the Transformer encoder composed of the multi-head attention mechanism and the feedforward neural network is used to realize information reinforcement of the composite fault common features on the individual features under single working condition, and relevant vector representations are obtained, the individual vectors fused with the composite fault common features under single working condition are taken as the input of the gated recurrent unit, the relevant relationship inside the time series is extracted, the output thereof is taken as the final composite fault feature representation, and the composite fault common and individual features under variable working conditions are fused.

[0012] Further, in the S20, the labeled thermal rolling process historical working condition data and the unlabeled target working condition data are collected, the edge probability distributions of the source domain and the target domain are different, the training data set is constructed by using the labeled normal and single fault data of the source domain and the unlabeled data of the target domain, the composite fault samples of the source domain and all types of samples of the target domain are used to construct the test data set, the convolutional neural network features and the capsule layer features of the source domain and the target domain samples are extracted in sequence by taking the training data set as the input, the probability vector and the loss term of the source domain sample are calculated by using the capsule layer feature vector, the model parameters are optimized and updated by using the Adam algorithm, the intelligent diagnosis model is constructed, the unknown test samples of the source domain and the target domain are classified by using the intelligent diagnosis model, and the composite fault intelligent diagnosis oriented to working condition migration is realized.

[0013] Further, the monitoring data of multiple running working conditions of the thermal rolling process are collected, the data collected under a working condition is selected as the training data set, the data collected under other working conditions is selected as the test data set, the features are extracted in sequence by using the convolutional neural network and the weight-sharing capsule layer by taking the training data set as the input, the probability vector and the loss term are calculated by using the last capsule layer feature vector, the model parameters are optimized and updated by using the Adam algorithm, the intelligent diagnosis model is constructed, the test samples are classified by using the intelligent diagnosis model, and the composite fault intelligent diagnosis oriented to working condition generalization is realized.

[0014] Further, in the S30, the composite fault intelligent diagnosis model based on deep fusion of multi-source heterogeneous data uses the common and individual characteristics information of the composite fault under variable working conditions to construct a composite fault diagnosis model reflecting the characteristics of each working condition, and uses the ensemble learning and Bayesian fusion method to fuse the diagnosis information of each model, thereby realizing accurate tracing of the composite fault under variable working conditions.

[0015] Further, the strong feature learning capability of the convolutional neural network is used to map the source domain and the target domain data to a high-dimensional feature space, to obtain the data distribution in the high-dimensional feature space, to use the constructed composite fault tracing model to discriminate the high-dimensional features, to design a non-adversarial field discriminator, to learn the similar features of the source domain and the target domain, to reduce the distribution difference between the two domains to match the features, to evaluate the similarity between the target domain sample data and the source domain data through a weighted discrimination mechanism, to discriminate the transferability of the data, and finally to construct a composite fault intelligent diagnosis framework for new fault modes by weighting the categories to which the data belong, so as to realize accurate tracing of the composite fault under variable working conditions.

[0016] Further, an adversarial convolutional autoencoder is designed to extract the interaction features among the three types of feature information, and a bidirectional gate unit and a multilayer perceptron are used to fuse the extracted features, a joint field separation network is used to extract the common and individual characteristics among the three types of feature information, a multi-scale dynamic Bayesian fusion method is used to realize adaptive dynamic fusion, and the interaction information among the three types of feature information is obtained, the fused multi-source heterogeneous feature information, the constructed composite fault classifier under unbalanced data, and the composite fault decoupling model under incomplete data are deeply fused, the model is trained using the multi-source heterogeneous composite fault data of the hot rolling process, the predicted values of the model are combined and used as the input of a deep ensemble learner, the deep ensemble learner is trained, the learning of the decision fusion parameters is automatically completed, and the hot rolling process composite fault intelligent diagnosis based on deep fusion of multi-source heterogeneous data is realized.

[0017] Further, the offline data set is standardized and divided into a training set and a validation set, the features and label information of the training set are selected as the state, the classification performance of the validation set is selected as the reward, the reward of sample selection is maximized as the target, a single-state Markov decision process of sample selection is constructed in the training set, the sample selection process is optimized by designing a loss function, and an error density function is introduced to realize adaptive selection of unbalanced composite fault samples; on the basis of the adaptive selection result of the unbalanced composite fault samples, a hierarchical clustering method is used to divide the complex student network for multi-class unbalanced classification into a plurality of student networks, i.e., under the guidance of the teacher network, the multi-class unbalanced composite fault classification problem is converted into a plurality of multi-classification problems, and the initial parameters of the knowledge distillation network are obtained by using the method of Boltzmann machine; a sample weight update strategy and a reward signal are designed by using a deep reinforcement learning method, and a fine-grained composite fault classification model based on knowledge distillation and deep reinforcement learning is constructed to realize the construction of the composite fault classifier under unbalanced data.

[0018] Further, the collected monitoring data is divided into a plurality of samples according to the sample length, and is randomly divided into a training data set and a test data set, wherein the training data set only contains hot rolling process normal operation state and single fault mode samples, and is labeled one by one, and the test data set contains unknown samples of composite fault modes; on the basis of composite fault feature extraction, a deep decoupling convolutional neural network is constructed, the training data set is used as input, a dynamic routing protocol algorithm is used to train and optimize the capsule layer in the decoupling classifier, and a minimum boundary loss function is designed to train and optimize the network model, so that the sample missing problem in the construction process of the hot rolling process composite fault decoupling model under incomplete data is solved; based on the fine-grained composite fault classification result, the attribute description of the hot rolling process single fault and composite fault is summarized, including the influence of the fault, the position of the fault and the reason of the fault, each attribute defines a dimension of a vector space, and 0 and 1 are used to represent whether the attribute exists in the description of a certain fault mode, the fault attribute is used as auxiliary information for training of a support vector machine, a random forest and a k-nearest neighbor model, the attribute knowledge is migrated from a source domain to a target domain, and the intelligent decoupling of the composite fault with simultaneous sample and label missing is realized.

[0019] Compared with the prior art, the application has at least the following beneficial effects:

[0020] The present application has the advantages that the present application has simple process and convenient operation, the composite fault intelligent diagnosis model based on multi-source heterogeneous data deep fusion fully utilizes the common and individual feature fusion information of the composite fault under variable working conditions, constructs a composite fault diagnosis model capable of fully reflecting the characteristics of each working condition, and utilizes integrated learning and other methods to fuse the diagnosis information of each model, so that the precise tracing of the composite fault under variable working conditions is realized. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the scheme of the present application, the drawings needed in the following embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings.

[0022] Figure 1 A flowchart of a variable working condition hot rolling process composite fault tracing method provided by the embodiment of the present application is shown in the figure.

[0023] Figure 2 A research idea diagram of a variable working condition hot rolling process composite fault tracing method provided by the embodiment of the present application is shown in the figure.

[0024] Figure 3 A technical roadmap of composite fault commonness and individuality feature fusion under variable working conditions of a variable working condition hot rolling process composite fault tracing method provided by the embodiment of the present application is shown in the figure.

[0025] Figure 4 A technical roadmap of working condition migration oriented composite fault intelligent diagnosis of a variable working condition hot rolling process composite fault tracing method provided by the embodiment of the present application is shown in the figure.

[0026] Figure 5 A technical roadmap of variable working condition composite fault precise tracing of a variable working condition hot rolling process composite fault tracing method provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0027] For the purpose of facilitating the understanding of the present application, a more complete description of the present application will be provided below with reference to the relevant drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be implemented in many different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the specification of the present application herein is only for the purpose of describing specific embodiments and is not intended to limit the present application.

[0029] The present application provides a hot rolling process compound fault tracing method under variable working conditions, which is applied to intelligent diagnosis of compound faults under variable working conditions of the hot rolling process. The hot rolling process compound fault tracing method under variable working conditions comprises the following steps:

[0030] S10, adopting multi-task learning and deep learning fusion to extract common and individual characteristics of compound faults under variable working conditions; S20, on the basis of extraction and fusion of common and individual characteristics of compound faults under variable working conditions, adopting transfer learning and convolutional neural network to realize intelligent diagnosis of compound faults under working condition transfer and generalization; S30, adopting ensemble learning and transfer learning method to realize accurate tracing of compound faults under variable working conditions.

[0031] The present application has simple process and convenient operation, and effectively improves the accuracy of tracing compound faults under variable working conditions of the hot rolling process.

[0032] In order to make the persons skilled in the art better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings.

[0033] The present application provides a hot rolling process compound fault tracing method under variable working conditions, which is applied to intelligent diagnosis of compound faults under variable working conditions of the hot rolling process. In combination with Figures 1 to 5 In the present embodiment, the hot rolling process compound fault tracing method under variable working conditions comprises the following steps:

[0034] S10, adopting multi-task learning and deep learning fusion to extract common and individual characteristics of compound faults under variable working conditions.

[0035] Specifically, in this embodiment, the composite fault features of the hot rolling process under variable working conditions are often similar but have great differences, making the extraction and intelligent diagnosis of the composite fault features of the hot rolling process under variable working conditions challenging. Based on multi-task learning and deep learning theory, the extraction and fusion method of common and individual features of the composite fault of the hot rolling process under variable working conditions is studied. Specifically, first, to solve the problem of poor generalization performance of the composite fault feature extraction model under variable working conditions, the difference information extraction method of common and individual features of the composite fault under variable working conditions is studied. According to the data characteristics of the hot rolling process under each single working condition, the convolutional neural network, variational modal decomposition, bidirectional gated recurrent unit, bidirectional minimum gated unit, etc. are used to extract the composite fault features under each single working condition, and the composite fault feature vector is obtained. The individual encoder composed of a feedforward neural network and the common encoder composed of a multi-task learning method are used to obtain the individual features and common features of the composite fault under each single working condition, and the common features are constrained to reduce the difference in spatial distribution of multiple feature vectors, realizing the extraction of common and individual features of the composite fault under variable working conditions.

[0036] Further, in this embodiment, the correlation and coupling relationship between the common and individual features of the composite fault under variable working conditions is considered, the influence of individual features on common features is analyzed, and the common and individual features are fused. Specifically, the Transformer encoder composed of multi-head attention mechanism and feedforward neural network is used to realize information reinforcement of the common features of the composite fault under single working condition on individual features, and to obtain the related vector representation. On this basis, the individual vector fused with the common features of the composite fault under single working condition is taken as the input of the gated recurrent unit, the correlation between the time series is extracted, and the output is taken as the final composite fault feature representation, realizing the fusion of common and individual features of the composite fault under variable working conditions. The research method and technical route are as shown in Figure 3

[0037] S20, on the basis of the extraction and fusion of common and individual features of the composite fault under variable working conditions, transfer learning and convolutional neural network are used to realize the intelligent diagnosis of the composite fault under working condition transfer and generalization.

[0038] ​Specifically, in this embodiment, constructing a complex fault intelligent diagnosis model for the hot rolling process under varying operating conditions may lead to significant distribution differences between the test data and training data, failing to satisfy the independent and identically distributed assumption and exhibiting domain transfer problems. Based on the extraction and fusion of common and unique features of complex faults under varying operating conditions, and grounded in transfer learning and convolutional neural networks, this paper studies a complex fault intelligent diagnosis method oriented towards operating condition transfer and generalization. Specifically, firstly, addressing the issue of poor generalization performance of the existing complex fault intelligent decoupling and diagnosis model under varying operating conditions, a complex fault intelligent diagnosis method oriented towards operating condition transfer is studied. Labeled historical operating condition data (source domain) and unlabeled target operating condition data (target domain) of the hot rolling process are collected, ensuring different marginal probability distributions between the source and target domains. A training dataset is constructed using labeled normal and single fault data from the source domain and unlabeled data from the target domain. A test dataset is constructed using composite fault samples from the source domain and samples of all types from the target domain. Using the training dataset as input, convolutional neural network features and capsule layer features are extracted from samples in both the source and target domains. The probability vector and loss term of the source domain samples are calculated using the capsule layer feature vectors. The Adam algorithm is then used to optimize and update the model parameters, constructing an intelligent diagnostic model. Based on this, the intelligent diagnostic model is used to classify unknown test samples in both the source and target domains, achieving intelligent diagnosis of complex faults oriented towards changing operating conditions. The research methods and technical routes are as follows: Figure 4 As shown.

[0039] Furthermore, in this embodiment, addressing the problem that complex fault data under extreme conditions such as catastrophic accidents cannot be collected in advance for decoupling and diagnostic model training, a complex fault intelligent diagnosis method generalized to operating conditions is studied. Specifically, monitoring data from multiple operating conditions in the hot rolling process are collected. Data collected under a specific operating condition is selected as the training dataset (labeled data), and data collected under other operating conditions is used as the test dataset. Using the training dataset as input, features are extracted sequentially using a convolutional neural network and weight-sharing capsule layers (a weight-sharing mechanism is introduced between adjacent capsule layers to improve generalization performance). The probability vector and loss term are calculated using the feature vector of the last capsule layer, and the Adam algorithm is used to optimize and update the model parameters to construct an intelligent diagnostic model. Based on this, the intelligent diagnostic model is used to classify test samples, achieving complex fault intelligent diagnosis generalized to operating conditions.

[0040] S30. Employ ensemble learning and transfer learning methods to achieve accurate tracing of complex faults under varying operating conditions.

[0041] Specifically, in this embodiment, the complex and variable operating conditions of the hot rolling process make it difficult for traditional intelligent diagnostic methods under single operating conditions to accurately trace complex faults. Based on ensemble learning and transfer learning methods, accurate tracing of complex faults under varying operating conditions is achieved. Specifically, firstly, based on the existing intelligent diagnostic model for complex faults that deeply fuses multi-source heterogeneous data, the common and unique characteristics of complex faults under varying operating conditions are fully utilized to construct a complex fault diagnostic model that can fully reflect the characteristics of each operating condition. Then, using ensemble learning, Bayesian fusion, and other methods, the diagnostic information from each model is fused to achieve accurate tracing of complex faults under varying operating conditions.

[0042] In this embodiment, the strong feature learning capability of convolutional neural networks is utilized to map source and target domain data to a high-dimensional feature space, obtain the data distribution in the high-dimensional feature space, and use the constructed composite fault tracing model to discriminate the high-dimensional features. A non-adversarial domain discriminator is designed to learn similar features between the source and target domains, reduce the distribution difference between the two domains to match features, evaluate the similarity between target domain sample data and source domain data through a weighted discrimination mechanism, and determine the data transferability. Finally, the category to which the data belongs is determined by weighted discrimination, and a composite fault intelligent diagnosis framework for new fault modes is constructed to achieve accurate tracing of composite faults under varying operating conditions.

[0043] In this embodiment, an adversarial convolutional autoencoder is designed to extract the interaction features among three types of feature information. The extracted features are then fused using a bidirectional gating unit and a multilayer perceptron. A joint domain separation network is used to extract the common and individual features among the three types of feature information. An adaptive dynamic fusion method is used to achieve adaptive dynamic fusion and obtain the interaction information among the three types of feature information. The fused multi-source heterogeneous feature information, the constructed composite fault classifier under imbalanced data, and the composite fault decoupling model under incomplete data are deeply fused. The model is trained using multi-source heterogeneous composite fault data from the hot rolling process, and the predicted values ​​of the model are merged and used as input to a deep ensemble learner to train the deep ensemble learner. The deep ensemble learner automatically completes the learning of decision fusion parameters, realizing intelligent diagnosis of composite faults in the hot rolling process through deep fusion of multi-source heterogeneous data.

[0044] 9. A method for tracing complex faults in hot rolling processes under varying operating conditions, as described in claim 8, is characterized in that: the offline dataset is standardized and divided into a training set and a validation set; the features and label information of the training set are selected as the state; the classification performance of the validation set is used as the reward; maximizing the reward for sample selection is used as the objective; a single-state Markov decision process for sample selection is constructed in the training set; the sample selection process is optimized by designing a loss function; and an error density function is introduced to achieve adaptive selection of imbalanced complex fault samples; based on the adaptive selection results of imbalanced composite fault samples, a hierarchical clustering method is used to divide the complex student network used for multi-class imbalanced classification into several student networks, that is, under the guidance of the teacher network, the classification problem of multi-class imbalanced composite faults is transformed into several multi-class problems; the initial parameters of the knowledge distillation network are obtained using the Boltzmann machine method; a sample weight update strategy and reward signal are designed using deep reinforcement learning methods; a fine-grained composite fault classification model based on knowledge distillation and deep reinforcement learning is constructed to realize the construction of a composite fault classifier under imbalanced data.

[0045] In this embodiment, the collected monitoring data is divided into multiple samples according to sample length, and randomly divided into training datasets and test datasets. The training dataset contains only samples of normal operation status and single fault mode in the hot rolling process, and each sample is labeled. The test dataset contains unknown samples of compound fault modes. Based on the feature extraction of compound faults, a deep decoupled convolutional neural network is constructed. Using the training dataset as input, a dynamic routing protocol algorithm is used to train and optimize the capsule layer in the decoupled classifier. The network model is trained and optimized by minimizing the boundary loss function to solve the problem of missing samples in the construction of the decoupled model of compound faults in the hot rolling process under incomplete data. Based on the fine-grained compound fault classification results, the attribute descriptions of single faults and compound faults in the hot rolling process are summarized, including the impact of the fault, the location of the fault, and the cause of the fault. Each attribute defines a dimension of the vector space, and 0 and 1 are used to indicate whether the attribute exists in the description of a certain fault mode. The fault attributes are used as auxiliary information for training support vector machines, random forests, and k-nearest neighbor models to transfer attribute knowledge from the source domain to the target domain, realizing intelligent decoupling of compound faults where both samples and labels are missing.

[0046] Furthermore, in this embodiment, addressing the issue of potential new failure modes arising during hot rolling under varying operating conditions—specifically, the target domain exhibiting more failure modes than the source domain—a composite fault accurate tracing method for these new failure modes is researched based on an adversarial domain adaptation algorithm. Specifically, leveraging the strong feature learning capability of convolutional neural networks, source and target domain data are mapped to a high-dimensional feature space to obtain the data distribution within this space. Building upon this, the constructed composite fault tracing model is used to discriminate the high-dimensional features. A non-adversarial domain discriminator is designed to learn similar features between the source and target domains, reducing the distribution differences between the two domains to match features. Since the target domain may contain new failure modes, a weighted discrimination mechanism is designed to evaluate the similarity between target domain sample data and source domain data, determining data transferability. Finally, by weightedly determining the data category, a composite fault intelligent diagnosis framework for new failure modes is constructed to achieve accurate tracing of composite faults under varying operating conditions. The research methods and technical routes are as follows: Figure 5 As shown.

[0047] The above-described method for tracing complex faults in the hot rolling process under varying operating conditions has lower accuracy compared to existing technologies. This invention features a simple process and convenient operation. Based on a multi-source heterogeneous data deep fusion intelligent diagnostic model for complex faults, it fully utilizes the fusion information of common and unique characteristics of complex faults under varying operating conditions to construct a comprehensive diagnostic model that fully reflects the characteristics of each operating condition. Furthermore, it employs ensemble learning and other methods to fuse the diagnostic information from each model, achieving accurate tracing of complex faults under varying operating conditions. Building upon this, and addressing the potential emergence of new fault modes in the hot rolling process under varying operating conditions, this invention researches a domain-adaptation-based method for accurate tracing of complex faults under varying operating conditions, based on the adversarial domain adaptation algorithm in transfer learning. This method effectively improves the accuracy of tracing complex faults in the hot rolling process under varying operating conditions.

[0048] Obviously, the embodiments described above are merely preferred embodiments of the present invention, and not all embodiments. The accompanying drawings illustrate preferred embodiments of the present invention, but do not limit the scope of the patent. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this invention.

Claims

1. A method of tracing compound faults in a hot rolling process under varying operating conditions, characterized by, Comprise the following steps: S10, adopt multi-task learning and deep learning fusion under variable working condition composite fault commonness and individuality characteristics; S20, on the basis of variable working condition composite fault commonness and individuality characteristics extraction and fusion, adopt transfer learning and convolutional neural network, realize working condition migration and generalization of composite fault intelligent diagnosis; S30, adopt ensemble learning and transfer learning method, realize variable working condition composite fault precision tracing.

2. The method according to claim 1, wherein, In the S10, according to the data characteristics of hot rolling process under each single working condition, the convolutional neural network, variational mode decomposition, bidirectional gated recurrent unit and bidirectional minimum gated unit method are used to extract the composite fault characteristics under each single working condition, obtain the composite fault characteristic vector, and the individuality encoder composed of feedforward neural network and the commonness encoder composed of multi-task learning method are used to obtain the individuality characteristics and commonness characteristics of each single working condition, and the commonness characteristics are constrained to reduce the difference in spatial distribution of multiple feature vectors, realize the extraction of commonness and individuality characteristics of composite fault under variable working condition.

3. The method according to claim 2, wherein, The Transformer encoder composed of multi-head attention mechanism and feedforward neural network is used to realize the information strengthening of single working condition composite fault commonness characteristics to individuality characteristics, and the related vector representation is obtained, the individuality vector fused with single working condition composite fault commonness characteristics is taken as the input of gated recurrent unit, the related relationship in time series is extracted, and the output is taken as the final composite fault feature representation, realizing the fusion of commonness and individuality characteristics of composite fault under variable working condition.

4. The method according to claim 1, wherein, In the S20, collect hot rolling process historical working condition data with labels and target working condition data without labels, meet the different edge probability distribution of source domain and target domain, use the labeled normal and single fault data of source domain and unlabeled data of target domain to construct training data set, the composite fault samples of source domain and all types of samples of target domain construct test data set; Take the training data set as input, extract the convolutional neural network features and capsule layer features of source domain and target domain samples in turn, and calculate the probability vector and loss term of source domain sample by using capsule layer feature vector, optimize and update model parameters by Adam algorithm, construct intelligent diagnosis model, and use the intelligent diagnosis model to classify unknown test samples of source domain and target domain, realize the intelligent diagnosis of composite fault facing working condition migration.

5. The method according to claim 4, wherein, Collect monitoring data of multiple running working conditions of hot rolling process, select the data collected under a certain working condition as training data set, and the data collected under other working conditions as test data set, take the training data set as input, use convolutional neural network and weight sharing capsule layer to extract features in turn, calculate probability vector and loss term by using the last capsule layer feature vector, and use Adam algorithm to optimize and update model parameters, construct intelligent diagnosis model, use the intelligent diagnosis model to classify test samples, realize the intelligent diagnosis of composite fault facing working condition generalization.

6. The method of claim 1, wherein, In the S30, the composite fault intelligent diagnosis model based on deep fusion of multi-source heterogeneous data utilizes the common and individual characteristic information of the composite fault under variable working conditions to construct a composite fault diagnosis model reflecting the characteristics of each working condition, and utilizes the integrated learning and Bayesian fusion method to fuse the diagnosis information of each model, thereby realizing the accurate tracing of the composite fault under variable working conditions.

7. The method according to claim 6, wherein, The strong feature learning capability of the convolutional neural network is utilized to map the source domain and the target domain data to a high-dimensional feature space, to obtain the data distribution in the high-dimensional feature space, to utilize the constructed composite fault tracing model to discriminate the high-dimensional features, to design a non-adversarial field discriminator, to learn the similar features of the source domain and the target domain, to reduce the distribution difference between the two domains to match the features, to evaluate the similarity between the target domain sample data and the source domain data through a weighted discrimination mechanism, to discriminate the transferability of the data, and finally to construct a composite fault intelligent diagnosis framework for new fault modes by weighting the categories to which the data belong, so as to realize the accurate tracing of the composite fault under variable working conditions.

8. The method of claim 6, wherein the method is characterized by, An adversarial convolutional autoencoder is designed to extract the interaction features among the three types of feature information, and a bidirectional gate unit and a multilayer perceptron are utilized to fuse the extracted features, a joint field separation network is adopted to extract the common and individual features among the three types of feature information, a multi-scale dynamic Bayesian fusion method is utilized to realize adaptive dynamic fusion, and the interaction information among the three types of feature information is obtained, the fused multi-source heterogeneous feature information, the constructed composite fault classifier under unbalanced data, and the composite fault decoupling model under incomplete data are deeply fused, the model is trained by the multi-source heterogeneous composite fault data of the hot rolling process, and the predicted values of the model are combined as the input of the deep ensemble learner, the deep ensemble learner is trained, the learning of the decision fusion parameters is automatically completed, and the hot rolling process composite fault intelligent diagnosis based on deep fusion of multi-source heterogeneous data is realized.

9. The method of claim 8, wherein, The offline data set is standardized and divided into a training set and a validation set, the features and label information of the training set are selected as the state, the classification performance of the validation set is selected as the reward, the reward of sample selection is maximized as the target, a single-state Markov decision process of sample selection is constructed in the training set, the sample selection process is optimized by designing a loss function, and an error density function is introduced to realize adaptive selection of unbalanced composite fault samples; on the basis of the adaptive selection result of the unbalanced composite fault samples, a hierarchical clustering method is utilized to divide the complex student network for multi-class unbalanced classification into a plurality of student networks, i.e., under the guidance of the teacher network, the multi-class unbalanced composite fault classification problem is converted into a plurality of multi-classification problems, and a Boltzmann machine method is used to obtain the initial parameters of the knowledge distillation network; a sample weight update strategy and a reward signal are designed by utilizing a deep reinforcement learning method, a fine-grained composite fault classification model based on knowledge distillation and deep reinforcement learning is constructed, and the construction of the composite fault classifier under unbalanced data is realized.

10. The method of claim 8, wherein the method is characterized by: The collected monitoring data is divided into multiple samples according to the sample length, and is randomly divided into a training data set and a test data set, wherein the training data set only contains hot rolling process normal operation state and single fault mode samples, and is labeled one by one, and the test data set contains unknown samples of composite fault modes; on the basis of composite fault feature extraction, a deep decoupling convolutional neural network is constructed, the training data set is used as input, the dynamic routing protocol algorithm is used to train and optimize the capsule layer in the decoupling classifier, and a minimum boundary loss function is designed to train and optimize the network model, so as to solve the sample missing problem in the construction process of the hot rolling process composite fault decoupling model under the condition of incomplete data; based on the fine-grained composite fault classification result, the attribute description of the hot rolling process single fault and composite fault is summarized, including the influence of the fault, the position of the fault and the reason of the fault, each attribute defines a dimension of the vector space, and 0 and 1 represent whether the attribute exists in the description of a certain fault mode, the fault attribute is used as auxiliary information for training of support vector machine, random forest and k-neighbor model, the attribute knowledge is migrated from the source domain to the target domain, and the intelligent decoupling of the composite fault with simultaneous sample and label missing is realized.

Images