Fault Diagnosis Method for Secure Transmission of Industrial Data Based on Transfer Learning

Abstract

The present invention discloses a fault diagnosis method for secure transmission of industrial data based on transfer learning, belonging to the technical field of data supervision. The method includes: realizing secure supervision and diagnosis of industrial data by constructing a transfer learning model, adopting an end-to-end feature learning strategy, integrating a CBAM hybrid attention module, dynamically strengthening fault-sensitive features through channel and spatial attention mechanisms, and suppressing noise interference; overcoming the defect of limited feature coverage of traditional single-scale networks and alleviating the problem of gradient disappearance in deep networks through multi-scale parallel processing and residual skip connections. The combination of channel and spatial attention mechanisms and multi-scale feature extraction enables the network to dynamically strengthen fault-sensitive features, significantly improving the robustness and accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention provides a fault diagnosis method for secure transmission of industrial data based on transfer learning, belonging to the technical field of data supervision. Background Art

[0002] With the rapid development of the industrial Internet, the complexity of industrial systems and the diversity of devices are increasing continuously, the amount of data is growing exponentially, and the fault modes have become more complex and diverse. Due to technical limitations, existing fault diagnosis methods often rely on a single data source or a fixed model, making it difficult to adapt to the dynamically changing industrial environment and unable to effectively process multi-source heterogeneous data and complex fault scenarios with high real-time requirements. This makes traditional fault diagnosis methods gradually expose problems such as low diagnosis efficiency, high misjudgment rate, and insufficient adaptability to new fault modes when facing the increasing demand for fault classification, and it is difficult to meet the urgent need for efficient, accurate, and real-time fault diagnosis in the modern industrial Internet environment.

[0003] Current fault diagnosis and classification methods include knowledge-based methods, model-based methods, deep learning-based methods, and signal processing-based methods; among them, knowledge-based methods use the experience accumulated by experts in long-term practice to establish a knowledge base and diagnose faults through the knowledge base. However, the coverage of the knowledge base is limited and it highly depends on the experience of experts, with poor accuracy and low diagnosis efficiency; model-based methods detect faults by establishing an accurate mathematical model and using the residual between the model prediction and the actual measurement data. However, this method requires high accuracy of the model and cannot adapt to the dynamic randomness and multi-source uncertainty of complex systems; deep learning-based methods achieve fault classification by using a multi-hidden layer network to gradually learn the features of the input data and perform multi-level abstraction to form a high-level representation. However, this method requires high quality and quantity of training data, and the network structure is complex, with poor real-time diagnosis performance; signal processing-based methods extract the features of faults through signal processing techniques and classify faults in combination with statistical analysis. However, this method is sensitive to signal noise, has poor anti-interference ability, and the diagnosis results depend on the accuracy of signal processing. In addition, existing fault diagnosis methods fail to ensure the security of data transmission, resulting in possible data leakage during the transmission process, which affects the final diagnosis results.

[0004] Therefore, with the explosion of data volume and the complexity of the network environment, how to achieve accurate fault diagnosis of industrial equipment on the premise of ensuring the security and effectiveness of industrial data during transmission has become the core challenge for improving the intelligent level of industrial equipment. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention proposes a fault diagnosis method for secure transmission of industrial data based on transfer learning, and the technical solutions adopted include:

[0006] Step 1: Collect the original data from different sources in the industrial process, and classify different types of data into different levels according to the importance of the original data;

[0007] Step 2: Construct a data storage module by defining an ontology model, and map the original data in Step 1 into the ontology model;

[0008] Step 3: Preprocess the data mapped in Step 2;

[0009] Step 4: Encrypt and transmit the data preprocessed in Step 3;

[0010] Step 5: Construct a transfer learning model, and input the data encrypted and transmitted in Step 4 into the transfer learning model for training;

[0011] Step 6: Use the transfer learning model obtained in Step 5 to implement fault diagnosis under the condition of secure transmission of industrial data.

[0012] Optionally, in Step 1, the data is classified and graded according to the "Industrial Data Classification and Grading Guidelines", including the R & D, production, operation and maintenance, management, and external data domains, and at the same time, it is divided into different levels according to the potential impact of the data;

[0013] Optionally, the criteria for dividing the data levels in Step 1 are: the data can be divided into three levels according to different impacts,

[0014] Among them, the leakage or failure of level-three data may trigger serious safety accidents, major environmental incidents, or significant economic losses;

[0015] Problems with level-two data may lead to safety accidents or cascading effects, have a significant impact on enterprise operations, and may lead to the leakage of supplier or customer resources;

[0016] Level-one data has limited impact on production operations, low recovery costs, and controllable negative impacts.

[0017] Optionally, Step 2 includes:

[0018] Step 2.1: Define the ontology model. In this step, the ontology model is described by OWL (Web Ontology Language) to define concepts such as devices and operations, and the semantics of data are represented in RDF (Resource Description Framework) format. Each element in the data model will have a unique URI for identification and is saved as a read-only file in XML format. These definitions ensure the semantic consistency of subsequent data storage and query, making data management more standardized.

[0019] Step 2.2: Design the original database. This stage is the design and construction of a relational database, which is usually used to store raw data without semantic processing. Design tables (such as device tables, diagnosis tables, etc.) and define appropriate columns for each table (such as device ID, device type, fault type). Data is stored according to the standard table structure without involving complex semantic associations, and the data is managed in a traditional way.

[0020] Step 2.3: Map the database to the ontology through the D2RQ tool. In this step, the table structure and data in the original database are mapped to the defined ontology model through the D2RQ tool. Each database table is mapped to an entity in the ontology, the column name is mapped to the property of the entity, and each row of data in the database corresponds to an instance in the ontology. This mapping enables relational data to be semantically managed through the ontology model and supports subsequent semantic queries.

[0021] Optionally, in Step 3, through data cleaning and formatting, noise, missing values, and outliers in the raw data are removed to ensure the accuracy and integrity of the data. At the same time, data from different sources and in different formats are converted into a unified standard format to ensure data consistency and availability. This process can improve data quality and provide a reliable basis for subsequent analysis and fault diagnosis; specifically, it includes:

[0022] Step 3.1: Data cleaning; process the noise in the data through wavelet transform and Fourier transform to identify and remove irrelevant or random error data; supplement missing data through methods such as interpolation and deletion; identify outliers that significantly deviate from the mean or distribution through the Z-score method or the IQR method, and at the same time, in combination with the mutation points or abnormal fluctuation characteristics in time series data, identify and correct extreme values or incorrect readings in the data.

[0023] Step 3.2: Data formatting; data standardization or normalization is the process of converting data from different sources into a unified standard format. Data from different sources may use different timestamps, units, or field names, so unified conversion is required to ensure that the data is compatible with each other and there will be no conflicts in subsequent processing.

[0024] Step 3.3: Data conversion converts data from one format to another, such as converting a CSV file to a database table, to ensure data compatibility between different systems.

[0025] Optionally, Step 4 includes:

[0026] Step 4.1: Develop a security policy and encrypt the data. According to the data type and application scenario, develop a suitable security policy and use DES encryption technology to protect the confidentiality of the data and prevent it from being intercepted during transmission;

[0027] First, generate sub-keys. Perform a key permutation on the 64-bit key, ignoring the 8th bit of each byte. The DES key is reduced from 64 bits to 56 bits. The 56-bit key is divided into two parts, the first 28 bits are and the last 28 bits are , and its expression is:

[0028]

[0029]

[0030] Next, according to the number of rounds (taking integer values in [1,16]), and respectively undergo a circular left shift of 1 bit or 2 bits, for a total of 16 circular shifts. The and obtained after the circular left shift are then compressed and permuted to obtain the sub-key .

[0031] After generating the sub-keys, the DES algorithm first performs an initial permutation on the 64-bit input data. Then, the first 32 bits of the data are defined as the left half and the last 32 bits are defined as the right half, and enter 16 rounds of iteration. In each round, the right half of the data first passes through an expansion permutation, then performs an exclusive OR operation with the sub-key of the current round, and then passes through the encryption function F for processing to generate a 32-bit output. This output is exclusive ORed with the left half of the data and exchanged as the input for the next round. After 16 rounds of processing, the data in the left and right parts are combined and an inverse initial permutation is performed to finally obtain the ciphertext output M, and its expression is:

[0032]

[0033] The decryption process is: Swap the positions of the initial transposition table and the final transposition table, and use the sub-key for the first time, for the second time, and so on for 16 times, and finally obtain the decrypted ciphertext.

[0034] Step 4.2: Implement a secure transmission protocol (such as SSL / TLS) through the Industrial Data Space (IDS) framework to establish an encrypted channel to prevent data from being tampered with or eavesdropped. Finally, the IDS framework ensures that cross-organizational data exchange complies with privacy protection and compliance requirements, supporting secure data sharing and interoperability;

[0035] The IDS framework provides standardized connection and usage control for the data security supervision platform through the First Data Connector and the Second Data Connector, allowing trusted applications to be executed in an isolated and authenticated environment. The APP Store in the data security supervision platform has functions such as basic data provision, data service and management, vocabulary management, and software guardianship, and these applications can be certified by a certification body recognized by the IDS. The Broker is responsible for functions such as management of data sources, search, data exchange protocol, and data exchange monitoring. The Data Provider can control the access and use of data by data consumers, that is, allowing consumers to access data under specific purposes and models. The Data Consumer searches for data from different Data Providers through the Broker and can use their data after establishing a security protocol with the Data Provider. The Data Sink is responsible for receiving data from the data source and converting it into a format suitable for further processing.

[0036] Step 4.3: Define the data scope, sign a security agreement, and evaluate the security capabilities of the recipient. The data provider should clarify the scope, category, conditions, and processing procedures of the data, and at the same time sign a data security agreement with the recipient to ensure that both parties have a clear agreement on the use, security requirements, and responsibilities of the data. In the agreement, the data provider also needs to evaluate the data security protection capabilities of the recipient to ensure that it has sufficient technologies and measures to protect the security of the data;

[0037] Step 4.4: Encrypt the data during the transmission process by using DES encryption technology to ensure that the data will not be intercepted or tampered with during the transmission process, thus ensuring the confidentiality and integrity of the data;

[0038] Step 4.5: Build a standardized framework for data exchange through the Industrial Data Space (IDS) framework, allowing data providers to maintain control over the data, ensuring secure sharing of data in a cross-organizational and cross-platform environment, while complying with privacy protection and compliance requirements;

[0039] Step 4.6: Design access control and permission management. The data provider needs to set strict access control policies to ensure that only authorized users can access and use sensitive data, and prevent unauthorized access through authentication and permission management.

[0040] Optionally, the transfer learning model in Step 5 includes a feature extraction module, a domain adaptation module, and a diagnostic classification module;

[0041] In the transfer learning model, the feature extraction module uses a multi-scale residual neural network combined with an attention mechanism. This network adopts an end-to-end feature learning strategy. The source domain data and the target domain data are preprocessed through two-level enhanced basic convolutional layers in sequence, and each level adopts a "Conv-BN-ReLU-Pool" normalization structure: the first level maps single-channel data to a 32-dimensional feature space through a 1×7 convolutional kernel, and the second level is extended to 64 channels, enhancing the feature expression ability through batch normalization and non-linear activation. The preprocessed 64-channel features are input into an improved multi-scale residual neural network, which is composed of three parallel multi-scale feature extraction units, and heterogeneous receptive fields are constructed using three different sizes of convolutional kernels, namely 3, 5, and 7, to simultaneously capture the local details and global trend features of the fault data. Each unit embeds a two-level residual structure: the first level maps 64 channels to 64 channels, and the second level is extended to 128 channels, and a CBAM hybrid attention module is integrated to dynamically strengthen the fault-sensitive features and suppress noise interference through channel and spatial attention mechanisms. The outputs of each unit are concatenated across channels to form a 384-dimensional feature tensor, compressed to a 256-dimensional high-density representation through a 1×3 fusion convolution, and finally a 256-dimensional global feature vector is generated through global average pooling, and fault classification is achieved through a fully connected layer.

[0042] Specifically, it includes:

[0043] Step 5.1: The feature extraction module extracts features through a multi-scale residual neural network combined with an attention mechanism;

[0044] The encrypted transmitted data passes through two layers of 1×7 convolutional layers and a 2×2 max pooling layer to extract basic features; these basic features will be input into three multi-scale convolutional sub-blocks for processing. Each multi-scale convolutional sub-block contains two cascaded residual blocks, where the sizes of the convolutional kernels are 1×3, 1×5, and 1×7 respectively. In each residual block, the number of channels is 64 and 128 in sequence, and finally the features are weighted through the CBAM hybrid attention mechanism. The output feature channel number of each multi-scale convolutional sub-block is 128. The outputs of each unit are concatenated across channels to form a 384-dimensional feature tensor, compressed to a 256-dimensional high-density representation through a 1×3 fusion convolution, and the final feature map is input into a fully connected layer, and the classification result is output after processing.

[0045] Step 5.2: Domain Adaptation. An improved metric difference module that constructs the Joint Maximum Mean Discrepancy (JMMD) and Correlation Alignment (CORAL) is used to collaboratively align the cross-domain feature distributions at the global and local levels, effectively reducing the inter-domain differences and improving the classification accuracy of unlabeled samples in the target domain.

[0046] The Joint Maximum Mean Discrepancy adjusts the global statistical characteristics and local feature relationships between the source domain and the target domain by jointly matching the marginal distribution and the conditional distribution. Its expression is:

[0047]

[0048] Where, represents the data distribution of the source domain, which is the distribution of the labeled data used during model training. represents the data distribution of the target domain, which is the distribution of the unlabeled data that the model needs to adapt to. represents the expectation of all samples in the source domain distribution , that is, the statistical average of the source domain data. represents the expectation of all samples in the target domain distribution , that is, the statistical average of the target domain data. represents the norm. is the selected kernel function.

[0049] is the Reproducing Kernel Hilbert Space (RKHS) of this kernel.

[0049] Correlation Alignment reduces the distribution difference between the source domain and the target domain by aligning the covariance matrices of the features in the source domain and the target domain. By assuming that the data features of the source domain and the target domain are and respectively, domain alignment is achieved by minimizing the difference between the covariance matrices of the source domain and the target domain. The expression is:

[0050]

[0051] Where, represents the covariance matrix of the source domain. represents the covariance matrix of the target domain. represents the Frobenius norm, which measures the difference between the covariance matrices.

[0052] By introducing a correlation alignment method based on the joint maximum mean discrepancy, the local distribution difference between the source domain and the target domain is further reduced through covariance matrix alignment. The joint maximum mean discrepancy adopts the kernel mean embedding method to globally compare the feature distributions of the source domain and the target domain, and jointly match the marginal distribution and the conditional distribution, thereby adjusting the global statistical characteristics of the entire feature space. In contrast, correlation alignment makes the internal correlations of features in different domains more consistent by aligning the covariance matrices, emphasizing the matching of local feature relationships.

[0053] Step 5.3: Use the joint maximum mean discrepancy loss function, the correlation alignment loss function, and the source domain classification loss function to jointly form the total loss function, and perform backpropagation to update the model;

[0054] The expression of the joint maximum mean discrepancy loss function is:

[0055]

[0056] where, represents the kernel function, which is used to measure the similarity between two samples; represents the number of samples in the source domain, that is, the total number of samples drawn from the source domain distribution and and represent the sample pairs in the source domain, where and represent the th and j th samples in the source domain respectively , represents the number of samples in the target domain, and represent the sample pairs in the target domain, where and represent the th and j th samples in the target domain respectively .

[0057] The correlation alignment loss function can be expressed as:

[0058]

[0059] where, is the dimension of the feature space, represents the covariance matrix of the source domain, represents the covariance matrix of the target domain, represents the Frobenius norm, which measures the difference between covariance matrices. By minimizing this loss function, correlation alignment can align the covariance matrices of the source domain and the target domain to make the internal correlations of their features more consistent.

[0060] The total loss function can be expressed as:

[0061]

[0062] where is the JMMD loss between the source domain and the target domain; is the CORAL loss between the source domain and the target domain; is the cross-entropy loss for source domain classification, and its expression is:

[0063]

[0064] where is the hyperparameter for adjusting the weights of each loss term in the joint maximum mean discrepancy, is the hyperparameter for adjusting the weights of each loss term in the correlation alignment, represents the number of samples in the source domain, represents the total number of classification categories. represents the true label of the -th sample in the source domain for the -th class, usually represented in one-hot encoding; represents the predicted probability of the model for the -th sample in the source domain for the -th class, usually output by the softmax function;

[0065] Therefore, the expression of the total loss function is:

[0066]

[0067] The optimization process of the total loss function will simultaneously consider the source domain classification task and the distribution alignment between the source domain and the target domain, so as to achieve better cross-domain transfer learning and classification effects.

[0068] Step 6: Use the transfer learning model obtained in Step 5 to implement the security supervision and fault diagnosis of industrial data;

[0069] Input the preprocessed industrial data into the transfer learning model through the transmission method in Step 4 to achieve the security supervision and fault diagnosis of industrial data.

[0070] The beneficial effects of the present invention are:

[0071] The present invention proposes an industrial data security supervision and diagnosis method based on transfer learning, which realizes the security supervision and diagnosis of industrial data by constructing a transfer learning model. It adopts an end-to-end feature learning strategy, integrates a CBAM hybrid attention module, and dynamically strengthens fault-sensitive features through channel and spatial attention mechanisms to suppress noise interference. Through multi-scale parallel processing and residual skip connections, it overcomes the defect of limited feature coverage of traditional single-scale networks and alleviates the problem of gradient disappearance in deep networks. The combination of channel and spatial dual attention mechanisms and multi-scale feature extraction enables the network to dynamically strengthen fault-sensitive features, significantly improving the robustness and accuracy of fault diagnosis. In addition, the difference between the source domain and the target domain is reduced by combining the joint maximum mean discrepancy and correlation alignment to ensure the effect of cross-domain transfer learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0073] Figure 1 is a schematic flowchart of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0074] Figure 2 is a schematic diagram of data mapping in the data storage module of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0075] Figure 3 is a schematic diagram of the principle of DES encryption technology in the data storage module of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0076] Figure 4 is a schematic diagram of the IDS component in the data storage module of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0077] Figure 5 is a schematic diagram of the transfer learning model architecture in the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0078] Figure 6 is a classification confusion matrix result diagram of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0079] Figure 7 is a classification result scatter diagram of the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention;

[0080] Figure 8 It is a schematic diagram of the structural parameters of the multi-scale residual neural network combined with the attention mechanism in the fault diagnosis method for industrial data secure transmission based on transfer learning proposed by the present invention. Specific implementation manners

[0081] To make the objectives, technical solutions and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the accompanying drawings.

[0082] Embodiment 1

[0083] This embodiment provides a fault diagnosis method for industrial data secure transmission based on transfer learning. The flow of this method is as Figure 1 shown, and this method includes:

[0084] Step 1: Collect the original data from different sources in the industrial process, and classify different types of data into different levels according to the importance of the original data;

[0085] Step 2: Construct a data storage module by defining an ontology model, and map the original data in Step 1 into the ontology model;

[0086] Step 3: Preprocess the data mapped in Step 2;

[0087] Step 4: Encrypt and transmit the preprocessed data in Step 3;

[0088] Step 5: Construct a transfer learning model, and input the data encrypted and transmitted in Step 4 into the transfer learning model for training;

[0089] Step 6: Use the transfer learning model obtained in Step 5 to realize fault diagnosis under the condition of industrial data secure transmission.

[0090] In Step 1, the data is classified and graded according to the "Industrial Data Classification and Grading Guidelines", including the R & D, production, operation and maintenance, management, and external data domains, and at the same time, it is divided into different levels according to the potential impact of the data;

[0091] The division of data levels in Step 1 is as follows: The data can be divided into three levels according to different importance degrees,

[0092] Among them, the leakage or failure of level-three data may trigger serious safety accidents, major environmental incidents or significant economic losses;

[0093] The problems of level-two data may lead to safety accidents or cascading effects, have a significant impact on the enterprise operation, and may lead to the leakage of supplier or customer resources;

[0094] The impact of primary data on production operations is limited, the recovery cost is low, and the negative impact is controllable.

[0095] Step 2 includes:

[0096] Step 2.1: Define the ontology model. In this step, the ontology model is described by OWL (Web Ontology Language) to define concepts such as devices and operations, and the semantics of data are represented in RDF (Resource Description Framework) format. Each element in the data model will have a unique URI for identification and is saved as a read-only file in XML format. These definitions ensure the semantic consistency of subsequent data storage and query, making data management more standardized.

[0097] Step 2.2: Original database design. This stage is the design and construction of a relational database, usually used to store raw data that has not been semantically processed. Design tables (such as device tables, diagnostic tables, etc.) and define appropriate columns for each table (such as device ID, device type, fault type). The data is stored according to the standard table structure without involving complex semantic associations, and the data is only managed in the traditional way.

[0098] Step 2.3: Map the database to the ontology through the D2RQ tool. As Figure 2 shown, in this step, the table structure and data in the original database are mapped to the defined ontology model through the D2RQ tool. Each database table is mapped to an entity in the ontology, the column name is mapped to the attribute of the entity, and each row of data in the database corresponds to an instance in the ontology. This mapping enables relational data to be semantically managed through the ontology model, supporting subsequent semantic queries.

[0099] In Step 3, through data cleaning and formatting, noise, missing values, and outliers in the original data are removed to ensure the accuracy and integrity of the data. At the same time, data from different sources and formats are converted into a unified standard format to ensure data consistency and availability. This process can improve data quality and provide a reliable basis for subsequent analysis and fault diagnosis; specifically includes:

[0100] Step 3.1: Data cleaning; Process the noise in the data through wavelet transform and Fourier transform to identify and remove irrelevant or random error data; Supplement the missing data through methods such as interpolation and deletion; Identify abnormal points that deviate significantly from the mean or distribution through statistical methods (such as Z-score method, IQR method), and at the same time, combined with the mutation points or abnormal fluctuation characteristics in the time series data, to identify and correct the extreme values or incorrect readings in the data.

[0101] Step 3.2: Data formatting process; Data formatting (such as standardization, normalization) is the process of converting data from different sources into a unified standard format. Data from different sources may use different timestamps, units, or field names, so unified conversion is required to ensure that the data is compatible with each other and there are no conflicts in subsequent processing.

[0102] Step 3.3: Data conversion converts data from one format to another, such as converting a CSV file to a database table, to ensure compatibility between different systems.

[0103] Step 4 includes:

[0104] Step 4.1: Develop a security policy and encrypt the data. According to the data type and application scenario, develop a suitable security policy and use DES encryption technology to protect the confidentiality of the data and prevent it from being intercepted during transmission;

[0105] As Figure 3 shown, first generate sub-keys. Perform a key permutation on the 64-bit key, ignoring the 8th bit of each byte. The DES key is reduced from 64 bits to 56 bits. The 56-bit key is divided into two parts, the first 28 bits are and the last 28 bits are , and its expression is:

[0106]

[0107]

[0108] Then, according to the number of rounds (taking integer values in [1,16]), and respectively undergo a circular left shift of 1 bit or 2 bits, for a total of 16 circular shifts. The and obtained after the circular left shift are then compressed and permuted to obtain the sub-key .

[0109] After generating the sub-keys, the DES algorithm first performs an initial permutation on the 64-bit input data. Then, the data is divided into left and right parts and enters 16 rounds of iteration. In each round, the right half of the data first undergoes an expansion permutation, then performs an exclusive OR operation with the sub-key of the current round, and then is processed through the encryption function F to generate a 32-bit output. This output is exclusive ORed with the left half of the data and swapped as the input for the next round. After 16 rounds of processing, the data in the left and right parts is combined and an inverse initial permutation is performed to finally obtain the ciphertext.

[0110] The decryption process is: Swap the positions of the initial transposition table and the final transposition table, and use the sub-key , for the second time , and so on for 16 times, and finally the decrypted ciphertext is obtained.

[0111] Step 4.2: Implement a secure transmission protocol (such as SSL / TLS) through the Industrial Data Space (IDS) framework to establish an encrypted channel to prevent data from being tampered with or eavesdropped. Finally, the IDS framework ensures that cross-organizational data exchange complies with privacy protection and compliance requirements, supporting secure data sharing and interoperability;

[0112] Such as Figure 4 As shown, the IDS framework provides standardized connection and usage control for the data security supervision platform through the First Data Connector and the Second Data Connector, allowing trusted applications to be executed in an isolated and authenticated environment. The APP Store in the data security supervision platform has functions such as basic data provision, data service and management, vocabulary management, and software guardianship, and these applications can be certified by a certification agency recognized by the IDS. The Broker is responsible for functions such as managing data sources, searching, data exchange protocols, and data exchange monitoring. The Data Provider can control the access and use of data by data consumers, that is, allowing consumers to access data under specific purposes and models. The Data Consumer searches for data from different data providers through the Broker and can use their data after establishing a security protocol with the data provider. The Data Sink is responsible for receiving data from the data source and converting it into a format suitable for further processing.

[0113] Step 4.3: Define the data scope, sign a security agreement, and evaluate the security capabilities of the recipient. The data provider should clarify the scope, category, conditions, and processing procedures of the data, and at the same time sign a data security agreement with the recipient to ensure that both parties have clear agreements on the use, security requirements, and responsibilities of the data. In the agreement, the data provider also needs to evaluate the data security protection capabilities of the recipient to ensure that it has sufficient technologies and measures to protect the security of the data;

[0114] Step 4.4: Encrypt the data during the transmission process by using DES encryption technology to ensure that the data will not be intercepted or tampered with during the transmission process, thus ensuring the confidentiality and integrity of the data;

[0115] Step 4.5: Build a standardized framework for data exchange through the Industrial Data Space (IDS) technology, allowing the data provider to maintain control over the data, ensuring the secure sharing of data in a cross-organizational and cross-platform environment, while complying with privacy protection and compliance requirements;

[0116] Step 4.6: Design access control and permission management. The data provider needs to set strict access control policies to ensure that only authorized users can access and use sensitive data, and prevent unauthorized access through authentication and permission management.

[0117] In Step 5, by constructing a transfer learning model, such as Figure 5 shown, the transfer learning model includes a feature extraction module, a domain adaptation module, and a diagnostic classification module;

[0118] Step 5.1: The feature extraction module extracts features through a multi-scale residual neural network combined with an attention mechanism;

[0119] The encrypted transmitted data passes through a 1×7 convolutional layer and a 2×2 max pooling layer to extract basic features; these basic features are input into three multi-scale convolutional sub-blocks for processing. Each multi-scale convolutional sub-block contains two cascaded residual blocks, where the kernel sizes of the convolutional layers are 1×3, 1×5, and 1×7 respectively. In each residual block, the number of channels is 64 and 128 in sequence, and finally the features are weighted through the CBAM hybrid attention mechanism. The output feature channel number of each multi-scale convolutional sub-block is 128. The features of three different scales are concatenated together in the channel dimension after global average pooling to form a feature map of size 1×1 with 384 channels. This final feature map is input into the fully connected layer, and the classification result is output after processing. The specific structural parameters of the multi-scale residual neural network combined with the attention mechanism are as Figure 8 shown.

[0120] Step 5.2: Domain adaptation, by constructing an improved metric difference module that combines maximum mean discrepancy and correlation alignment to synergistically align cross-domain feature distributions at the global and local levels, effectively reducing the domain difference and improving the classification accuracy of unlabeled samples in the target domain;

[0121] The joint maximum mean discrepancy adjusts the global statistical characteristics and local feature relationships of the source domain and the target domain by jointly matching the marginal distribution and the conditional distribution; its expression is:

[0122]

[0123] where represents the data distribution of the source domain, which is the distribution of the labeled data used during model training; represents the data distribution of the target domain, which is the distribution of the unlabeled data that the model needs to adapt to; represents the expectation of all samples in the source domain distribution , that is, the statistical average of the source domain data; Represents the expectation of all samples in the target domain distribution , that is, the statistical average of the target domain data; Represents the norm, is the selected kernel function, and is the Reproducing Kernel Hilbert Space (RKHS) of this kernel.

[0124] The expression of the joint maximum mean discrepancy loss function is:

[0125]

[0126] Among them, represents the kernel function, which is used to measure the similarity between two samples; represents the number of samples in the source domain, that is, the total number of samples drawn from the source domain distribution ; and respectively represent the -th and j -th samples in the source domain, ; represents the number of samples in the target domain, where and respectively represent the -th and j -th samples in the target domain .

[0127] The relevant alignment reduces the distribution difference between the source domain and the target domain by aligning the covariance matrices of the source domain and target domain features. By assuming that the data features of the source domain and target domain are and respectively, the domain alignment is achieved by minimizing the difference between the covariance matrices of the source domain and target domain, and the expression is:

[0128]

[0129] Among them, represents the covariance matrix of the source domain, represents the covariance matrix of the target domain, represents the Frobenius norm, which measures the difference between covariance matrices.

[0130] The expression of the relevant alignment loss function is:

[0131]

[0132] Among them, represents the dimension of the feature space;

[0133] By minimizing the relevant alignment loss function, the covariance matrices of the source domain and the target domain can be aligned, making the internal correlations of their features more consistent.

[0134] By introducing the relevant alignment method on the basis of the joint maximum mean discrepancy, the local distribution differences between the source domain and the target domain are further reduced through covariance matrix alignment. The joint maximum mean discrepancy uses the kernel mean embedding method to compare the feature distributions of the source domain and the target domain as a whole, and jointly matches the marginal distribution and the conditional distribution, thereby adjusting the global statistical characteristics of the entire feature space. In contrast, relevant alignment makes the internal correlations of features in different domains more consistent by aligning the covariance matrices, emphasizing the matching of local feature relationships.

[0135] Step 5.3: Use the joint maximum mean discrepancy loss function, the relevant alignment loss function, and the source domain classification loss function to jointly form the total loss function, and perform backpropagation updates on the model;

[0136] The total loss function can be expressed as:

[0137]

[0138] where is the joint maximum mean discrepancy loss between the source domain and the target domain; is the relevant alignment loss between the source domain and the target domain; is the cross-entropy loss for source domain classification, and its expression is:

[0139]

[0140] where, is the hyperparameter used to adjust the weights of each loss term in the joint maximum mean discrepancy, is the hyperparameter used to adjust the weights of each loss term in the relevant alignment, represents the total number of classification categories; represents the -th sample in the source domain and its true label in the -th class, usually represented in one-hot encoding; represents the predicted probability of the model for the -th sample in the source domain in the -th class, usually output by the softmax function.

[0141] Therefore, the expression of the total loss function is:

[0142]

[0143] The optimization process of the total loss function simultaneously considers the source domain classification task and the distribution alignment between the source domain and the target domain (through the joint maximum mean difference and correlation alignment), thereby achieving better cross-domain transfer learning and classification effects.

[0144] The transfer learning model is trained through the total loss function to achieve reverse update of the transfer learning model.

[0145] Step 6: Use the model obtained in step 5 to implement security supervision and fault diagnosis of industrial data;

[0146] The preprocessed industrial data is input into the transfer learning model through the transmission method of step 4 to realize the security supervision and fault diagnosis of the industrial data.

[0147] Embodiment 2

[0148] This embodiment provides a fault diagnosis method based on secure transmission of industrial data through transfer learning. The method is implemented based on the transfer learning model provided in the first embodiment. The data of the cement roller press is monitored and diagnosed through the method. The method specifically includes:

[0149] Step 1: Data collection and storage; Collect historical data from two cement roller presses at different locations in a cement plant. Collect data by installing four stress wave sensors SNR-10800, and use a data collector to store and process the data, and then upload the data to the server.

[0150] Step 2: Data processing: First, clean and format the data. Then, the first 1300k points of each working condition of the two roller presses are segmented according to 3072 sampling points as one data sample. The sliding step is set to 1024. Each working condition is cut off after generating 1000 samples. Roller press A generates 4000 samples as a training set, and roller press B generates 4000 samples as a test set, for a total of 8000 samples.

[0151] Step 3: Ensure the secure transmission of data; use encryption algorithms and deploy IDS components to achieve trusted data transmission. Roller A, as the data provider, uploads the labeled source domain data, while roller B, as the data user, transmits the unlabeled target domain data. The data of both parties are aggregated to the same server through the data connector Connector.

[0152] Step 4: The data uploaded to the server is trained using the transfer learning model to generate diagnostic results, so that the unlabeled data of roller press B can be parsed and its actual operating status can be diagnosed. According to the diagnostic results, the abnormal status of roller press B equipment is adjusted in time, thereby improving the operating efficiency and stability of the production line.

[0153] The confusion matrix of the classification results is as follows Figure 6 shown. It can be seen that there is a certain degree of confusion in the sample data of the two working conditions of predicted label 1 and predicted label 3. Most of the mispredicted samples are confused between working conditions 0 and 1. However, the vast majority of samples are concentrated on the diagonal line, that is, the vast majority of samples are correctly classified, and the accuracy rate is 99.3%. Therefore, it can be proved that the transfer learning model proposed by the present invention has good effectiveness in identifying the fault types of the unlabeled target domain.

[0154] To verify the effectiveness of the method proposed by the present invention, the t-SNE method is used to reduce the dimensionality of the feature data, and a scatter plot of the classification results is drawn, as shown in Figure 7 shown; where category 0 represents heavy load of the moving roller, category 1 represents light load of the moving roller, category 2 represents heavy load of the fixed mixing, and category 3 represents light load of the fixed roller. As can be seen from Figure 7 this, the feature points of each working condition show a relatively clear distribution in the reduced-dimensional space, with clear distinguishability between different categories and no obvious mixing phenomenon. This indicates that the model can effectively extract and identify the feature patterns under each working condition. The feature points show strong clustering characteristics in the projection space with clear boundaries, further proving the superiority of the transfer learning model proposed by the present invention in terms of generalization and robustness, and having good fault type identification ability.

[0155] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0156] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault diagnosis method based on transfer learning for secure transmission of industrial data, characterized in that: The method comprises: Step 1: Collect raw data from different sources in the industrial process and classify different types of data into different levels according to the importance of the raw data; Step 2: Build a data storage module by defining an ontology model and map the original data in step 1 into the ontology model; Step 3: Preprocess the data mapped in step 2; Step 4: Encrypt and transmit the pre-processed data in step 3; Step 5: Build a transfer learning model and input the encrypted data transmitted in step 4 into the transfer learning model for training; Step 6: Use the transfer learning model obtained in step 5 to realize fault diagnosis under the condition of secure transmission of industrial data; The transfer learning model in step 5 includes: a feature extraction module, a domain adaptation module and a diagnosis classification module; step 5 includes: Step 5.1: Construct a feature extraction module to extract features from encrypted transmitted data, weight them through the CBAM hybrid attention mechanism, concatenate feature maps in the channel dimension, and finally output the classification results; Step 5.2: construct a domain adaptation module, which measures the cross-domain feature distribution at the global and local levels by combining the maximum mean difference and the relevant alignment metrics; Step 5.3: Construct a total loss function, which is composed of the joint maximum mean difference loss function, the correlation alignment loss function, and the source domain classification loss function, and perform a reverse update on the model. The domain adaptation module includes joint maximum mean difference and correlation alignment, which realizes the alignment of cross-domain feature distribution at global and local levels; In step 5.2, the processing process of the domain adaptation module includes: The joint maximum mean difference adjusts the global statistical characteristics and local feature relationship between the source domain and the target domain by jointly matching the marginal distribution and the conditional distribution; its expression is: in, Represents the data distribution of the source domain, which indicates the distribution of labeled data used in model training; Represents the data distribution of the target domain, indicating the distribution of unlabeled data that the model needs to adapt to; Represents the source domain distribution All samples in , which is the statistical average of the source domain data; Represents the distribution of the target domain All samples in The expectation of , that is, the statistical average of the target domain data; represents the norm, is the selected kernel function, is the regenerated Hilbert space of the kernel; The correlation alignment is performed by aligning the covariance matrices of the source domain and the target domain features; assuming that the data features of the source domain and the target domain are and , inter-domain alignment is achieved by minimizing the difference between the covariance matrices of the source domain and the target domain, expressed as: in, represents the covariance matrix of the source domain, represents the covariance matrix of the target domain, stands for the Frobenius norm, which measures the difference between covariance matrices.

2. The method according to claim 1, characterized in that The feature extraction module includes a convolution layer, a maximum pooling layer, a residual block, a multi-scale splicing layer, a fused convolution layer, an adaptive pooling layer, a fully connected layer and a CBAM hybrid attention mechanism. In step 5.1, the processing process of the feature extraction module includes: The encrypted transmitted data passes through a 1×7 convolution layer and a 2×2 maximum pooling layer to extract the basic features and input them into three multi-scale convolution sub-blocks for processing; each multi-scale convolution sub-block contains two cascaded residual blocks, where the sizes of the convolution kernels are 1×3, 1×5 and 1×7 respectively; in each residual block, the number of channels is 64 and 128 respectively, and the features are weighted by the CBAM hybrid attention mechanism, and the number of output feature channels of each multi-scale convolution sub-block is 128; after global mean pooling, the features of the three different scales are spliced ​​together in the channel dimension to form a 1×1 feature map with 384 channels. The final feature map is input into the fully connected layer and the classification result is output after processing.

3. The method according to claim 1, characterized in that The expression of the joint maximum mean difference loss function in step 5.3 is: in, represents the kernel function, which is used to measure the similarity between two samples; The number of samples representing the source domain, that is, the distribution from the source domain The total number of samples drawn from and Represent the source domain and j samples, ; represents the number of samples in the target domain, where and Represent the target domain and j Samples ; The expression of the related alignment loss function is: in, represents the dimension of the feature space, represents the covariance matrix of the source domain, represents the covariance matrix of the target domain, stands for Frobenius norm; The total loss function is: in is the JMMD loss between the source domain and the target domain; It is the CORAL loss between the source domain and the target domain; is the cross entropy loss of source domain classification, which is expressed as: in, is a hyperparameter used to adjust the weights of each loss term in the joint maximum mean difference, is a hyperparameter used to adjust the weights of each loss term in the correlation alignment. represents the total number of classification categories; Represents the source domain The sample in The true label of the class; The representative model is the source domain The sample in Prediction probability on class; Therefore, the expression of the total loss function is: Using the total loss function Update the transfer learning model; When the total loss function does not decrease as the training progresses, the training ends.

4. The method according to claim 1, characterized in that: The step 2 comprises: Step 2.1: Describe the ontology model by defining devices and operations through OWL, build the data storage module using the ontology model, and use RDF format to represent the semantics of the original data; Step 2.2: Design the original database and store the original data in a standard table structure; Step 2.3: Use D2RQ to map the table structure and data to the ontology model in step 2.1 to achieve data storage.

5. The method according to claim 1, characterized in that The step 4 comprises: Step 4.1: Encrypt the data to be transmitted using DES encryption technology; Step 4.2: Implement secure transmission protocol through industrial data space framework and establish encrypted channel; The industrial data space framework includes a data security supervision platform, a first data connector, a second data connector and a data receiver, wherein the data security supervision platform includes an application store and an agent; the first data connector and the second data connector are responsible for providing standardized connection and use control for the data security supervision platform, allowing trusted applications to be executed in an isolated and authenticated environment; the application store in the data security supervision platform has basic data provision, data service and management, vocabulary management and software monitoring functions; the agent is responsible for providing management, search, data exchange protocol and data exchange monitoring functions for the data source; the data provider controls the data user's access to and use of the data, that is, allows the user to access the data under specific purposes and models; the data user searches for data from different data providers through the agent and uses their data after establishing a security agreement with the data provider; the data receiver is responsible for receiving data from the data source and converting it into a format suitable for further processing; Step 4.3: The data provider ensures data security by signing a data security agreement with the data user; Step 4.4: Encrypt the data during transmission by using DES encryption technology; Step 4.5: Provide a standardized framework for data exchange through the Industrial Data Space Framework to ensure data providers have control over their data; Step 4.6: Ensure data security by setting up access control and permission management; The data security protocol includes one of SSL / TLS, IPSec, PGP or S / MIME.

6. The method according to claim 1, characterized in that The step 3 comprises: Step 3.1: Data cleaning: Process the noise in the data through wavelet transform or Fourier transform, fill the missing values ​​through interpolation, and detect and process outliers through Z-score method or IQR method; Step 3.2: Data formatting: converting data from different sources into a unified standard format through standardization or normalization; Step 3.3: Convert the data format to ensure compatibility between different systems.

Images