Cross-domain data set fault migration health state identification method

Abstract

The invention belongs to the field of information management, and particularly relates to a cross-domain data set fault migration health state recognition method, which comprises the following steps: data acquisition: acquiring a cross-domain data set from a plurality of data sources for cleaning and feature extraction of the acquired data; a deep learning model is constructed, and faults in the cross-domain data set are identified; constructing a graph neural network model, and evaluating the health state of the cross-domain data set; monitoring the state of the cross-domain data set in real time, and triggering an alarm when a fault is detected; and determining the specific position and type of the fault through probability distribution output by the model.

Description

Technical Field

[0001] The present invention relates to the technical field of data set fault migration and health status identification, and in particular to a cross-domain data set fault migration health status identification method. Background Art

[0002] With the rapid development of information technology, big data and cloud computing have become indispensable technical supports for modern enterprises. However, data sets often face various failures during cross-domain transmission and storage, such as network delays, data loss, server failures, etc. These failures seriously affect the stability and reliability of the system. Traditional fault detection and health status identification methods mainly rely on single-domain data and lack the ability to comprehensively monitor and manage cross-domain data sets.

[0003] Complexity of cross-domain datasets: Cross-domain datasets involve multiple data sources and different data formats, which increases the difficulty of data processing and analysis.

[0004] Various fault types: There are various fault types in cross-domain datasets, including network faults, hardware faults, software faults, etc. Traditional fault detection methods are difficult to fully cover.

[0005] High real-time requirements: Fault detection and health status identification of cross-domain data sets need to be performed in real time to promptly discover and handle faults and ensure the normal operation of the system.

[0006] Data security and privacy: During cross-domain data transmission, data security and privacy protection is also an important issue.

[0007] Therefore, developing a system and method that can effectively identify cross-domain dataset failures and perform health status assessment is of great significance to improving the stability and reliability of the system. Summary of the invention

[0008] The purpose of the present invention is to provide a cross-domain data set fault migration health status identification method in order to solve the above problems, which is characterized by comprising:

[0009] S1 data collection, extensively collect raw data, and pre-process the collected raw data to convert data in different formats into numerical matrix or vector form for subsequent processing;

[0010] S2 feature extraction, including the use of principal component analysis combined with local linear embedding for feature extraction of numerical data, the use of word vector model combined with text convolutional neural network for feature extraction of text data, and the use of pre-trained ResNet network for image data and targeted fine-tuning for feature extraction;

[0011] S3 domain adaptation, which uses a domain adaptation algorithm based on a generative adversarial network and dynamic distribution matching. The feature data of the source domain and the target domain after feature extraction are respectively input into the generator and discriminator of the GAN for training, and the training weights of the generator and the discriminator are adjusted using a dynamic distribution matching mechanism, so that the distribution of the source domain data in the feature space is close to the distribution of the target domain data.

[0012] S4 health status recognition, building a health status recognition model based on a bidirectional long short-term memory network and an attention mechanism, inputting the source domain and target domain data after domain adaptation into the BiLSTM-AM model for training, for health status recognition of unknown data, and outputting the health status category of the device and the corresponding confidence score;

[0013] S5 result output and visualization is used to output the results obtained by the health status identification module and visualize the output results in the form of intuitive charts to show the change trend of the health status of the equipment over time and the distribution ratio information of different fault types.

[0014] Furthermore, during the data collection process, for sensor data, the mean and standard deviation of the data are calculated, and data outside the standard deviation range is considered as an outlier and removed; for text data, a text similarity algorithm is used to detect and delete duplicate text records.

[0015] Furthermore, the health status identification also adopts a cross-domain dataset health status assessment model of a graph neural network, abstracting the cross-domain dataset into a graph structure, with nodes representing data sources and servers, and edges representing data transmission paths; extracting the topological features of the graph, using the graph neural network to learn the graph structure, capturing the relationships between nodes, introducing an attention mechanism, highlighting the influence of key nodes and edges, and simultaneously learning fault identification and health status assessment tasks.

[0016] Furthermore, the graph structure data is annotated to mark the healthy and faulty states, the model is trained using the annotated data, the performance of the model is evaluated using the cross-validation method, and the hyperparameters of the model are optimized through grid search and random search methods.

[0017] Furthermore, when the domain adaptation adopts a domain adaptation algorithm based on a generative adversarial network and dynamic distribution matching, the parameters of the generator and discriminator of the GAN are initialized, and the Wasserstein distance between the data generated by the generator and the real target domain data is calculated. The training weights of the generator and the discriminator are dynamically adjusted according to the distance until the distribution difference between the data generated by the generator and the real target domain data in the feature space is minimized, that is, the Wasserstein distance is less than the set threshold, or the preset number of training iterations is reached, and the domain adaptation process is completed.

[0018] Furthermore, when constructing the health status recognition model based on the bidirectional long short-term memory network and the focus mechanism, the BiLSTM network is set to 2 layers, with 128 nodes in each layer, and the additive attention mechanism is adopted. The source domain and target domain data after domain adaptation are input into the BiLSTM-AM model for training. The cross entropy loss function is used as the loss function of the model, and the parameters of the model are continuously adjusted through the back propagation algorithm.

[0019] Furthermore, a domain adaptation algorithm based on generative adversarial networks and dynamic distribution matching is adopted. The feature data of the source domain and the target domain after feature extraction are input into the generator and the discriminator respectively. By calculating the Wasserstein distance between the data generated by the generator and the real target domain data, the training weights of the generator and the discriminator are dynamically adjusted.

[0020] Furthermore, a health status recognition model based on a bidirectional long short-term memory network and attention mechanism is constructed to identify the health status category of the equipment and the corresponding confidence score by bidirectionally processing the temporal information of the data and automatically focusing on the key parts.

[0021] Furthermore, multiple adversarial generative network structures are constructed, each of which is responsible for learning the distribution difference between the source domain and the target domain in a modality. Through adversarial training, the fake target domain data generated by the generator is made similar to the distribution of the real target domain data in this modality, and the discriminator distinguishes them.

[0022] Furthermore, the Wasserstein distance in the optimal transfer theory is introduced to measure the overall distribution difference between the source domain and the target domain in multimodality. The training weights of each adversarial generative network are dynamically adjusted according to the calculated Wasserstein distance, and the training is continuously iterated until the preset conditions are met.

[0023] Beneficial technical effects of the invention:

[0024] The present invention realizes comprehensive monitoring of cross-domain data sets through multi-source data collection and feature extraction, covering multiple levels such as network, hardware, and software. The present invention combines deep learning and graph neural networks for the first time, uses DNN and CNN to process complex features, and uses GNN and attention mechanisms to capture the relationships between nodes, thereby improving the robustness and interpretability of the model. The present invention can quickly discover and locate faults through real-time monitoring and fault detection, thereby improving the response speed and processing power of the system. The present invention improves the comprehensive performance and reliability of the model by learning fault identification and health status assessment tasks simultaneously through multi-task learning. In the process of data transmission and processing, the present invention adopts encryption technology and access control mechanism to ensure data security and privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the process;

[0026] Figure 2 Diagram of algorithm steps. DETAILED DESCRIPTION

[0027] The specific implementation modes of the present invention are described below in conjunction with examples. However, the following examples are only used to illustrate the present invention in detail and are not intended to limit the scope of the present invention in any way.

[0028] Example 1

[0029] Cross-domain dataset fault identification model based on deep learning

[0030] Data collection: Collect cross-domain data sets from multiple data sources, including network logs, server logs, application logs, etc.

[0031] Data cleaning: remove invalid and redundant data, and fill or delete missing values.

[0032] Feature extraction: Extract key features such as network latency, packet loss rate, server response time, etc.

[0033] Model construction: Deep Neural Network (DNN): Build a multi-layer deep neural network to learn complex features of cross-domain datasets. Convolutional Neural Network (CNN): Use convolutional neural network to extract local features of data to improve the robustness of the model. Long Short-Term Memory Network (LSTM): Introduce LSTM network to process time series data and capture the temporal dependency of data.

[0034] Model training: Data labeling: Label the collected data to mark faults and normal states. Training process: Use labeled data to train the model and use cross-validation to evaluate the performance of the model. Hyperparameter optimization: Use grid search and random search methods to optimize the model's hyperparameters and improve the model's generalization ability.

[0035] Fault identification: Real-time monitoring: Deploy the model to the actual system to monitor the status of cross-domain data sets in real time. Fault detection: When the model detects abnormal data, it triggers an alarm and notifies the administrator to handle it. Fault location: Determine the specific location and type of the fault through the probability distribution of the model output.

[0036] Mathematical formula and parameter definition

[0037] Deep Neural Networks (DNN):

[0038] Forward propagation: z (l) =W (l) a (l-1) +b (l)

[0039] a (l) =f(z (l) )

[0040] Among them, z (l) is the weighted input of the (l)th layer, W (l) is the weight matrix of the (l)th layer, b (l) is the bias term of the l(l)th layer, a (l) is the activation output of the (l)th layer, and f is the activation function.

[0041] Convolutional Neural Network (CNN):

[0042] Convolution operation:

[0043] Where I is the input image, K is the convolution kernel, M and N are the sizes of the convolution kernel, and (i, j) are the coordinates of the output image.

[0044] Long Short-Term Memory (LSTM)

[0045] Gating mechanism:

[0046] i t =σ(W i [h t-1 , x t ]+b i )

[0047] O t =σ(W o [h t-1 , x t ]+b o ).

[0048]

[0049]

[0050] Among them, i t is the input gate, f t is the forget gate, Ot is the output gate, is the candidate cell state, Ctt is the cell state, hht is the hidden state, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, and ⊙ represents element-wise multiplication.

[0051] Example 2

[0052] Cross-domain dataset health status assessment model based on graph neural network

[0053] Data modeling: Graph structure construction: abstract the cross-domain dataset into a graph structure, where nodes represent data sources and servers, and edges represent data transmission paths. Graph feature extraction: extract the topological features of the graph, such as node degree, path length, connectivity, etc.

[0054] Model construction: Graph Neural Network (GNN): Use graph neural network to learn graph structure and capture the relationship between nodes. Attention mechanism: Introduce attention mechanism to highlight the influence of key nodes and edges and improve the interpretability of the model. Multi-task learning: Simultaneously learn fault identification and health status assessment tasks to improve the overall performance of the model.

[0055] Model training: Data labeling: Label the graph structure data and mark the health and fault status. Training process: Use the labeled data to train the model and use the cross-validation method to evaluate the performance of the model. Hyperparameter optimization: Use grid search and random search methods to optimize the model's hyperparameters and improve the model's generalization ability.

[0056] Health status assessment: Real-time monitoring: Deploy the model to the actual system to monitor the health status of cross-domain data sets in real time. Health score: Evaluate the overall health status of the system through the health score output by the model. Fault prediction: Based on historical data and model prediction, early warning of potential fault risks.

[0057] Mathematical formula and parameter definition

[0058] Graph Neural Networks (GNN):

[0059] Graph convolution operation:

[0060] Among them, W (l) is the node feature matrix of the lth layer, is the normalized adjacency matrix, W (l) is the weight matrix of layer l, σ is the activation function (such as ReLU)

[0061] Attention Mechanism:

[0062] Attention score: e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0063]

[0064] h i '=σ(∑ i＝N(0) α i W j )

[0065] Among them, eij is the attention score between node i and node j, a is the attention vector, W is the weight matrix, h i is the feature vector of node i, N(i) is the neighbor set of node i, α ij is the normalized attention score, hi′ is the updated node feature vector, and σ is the activation function (such as ReLU).

[0066] Multi-task learning:

[0067] Loss function: L = λ1L fault +λ2L health

[0068] Among them, L fault is the loss function of the fault identification task, L health is the loss function of the health status assessment task, and λ1 and λ2 are the weight coefficients of the task.

[0069] Example 3

[0070] Data preprocessing: Data collection: Collect network logs, server logs, and application logs from multiple data centers. Data cleaning: Use the Pandas library to clean the data, remove invalid and redundant data, and fill or delete missing values. Feature extraction: Extract key features such as network latency, packet loss rate, and server response time.

[0071] Model construction: Deep Neural Network (DNN): Build a DNN model with multiple hidden layers and use the ReLU activation function. Convolutional Neural Network (CNN): Add convolutional layers and pooling layers to extract local features based on DNN. Long Short-Term Memory Network (LSTM): Add LSTM layers to process time series data based on DNN and CNN.

[0072] Model training: Data labeling: Label the collected data to mark faults and normal states. Training process: Use the Keras library to train the model and use the cross-validation method to evaluate the performance of the model. Hyperparameter optimization: Use the GridSearchCV and RandomizedSearchCV methods to optimize the model's hyperparameters and improve the model's generalization ability.

[0073] Fault identification: Real-time monitoring: Deploy the model to the actual system to monitor the status of cross-domain data sets in real time. Fault detection: When the model detects abnormal data, it triggers an alarm and notifies the administrator to handle it. Fault location: Determine the specific location and type of the fault through the probability distribution of the model output.

[0074] Health status assessment of cross-domain datasets based on graph neural networks,Data modeling: Graph structure construction: abstract the cross-domain dataset into a graph structure, where nodes represent data sources and servers, and edges represent data transmission paths. Graph feature extraction: extract the topological features of the graph, such as node degree, path length, connectivity, etc.

[0075] Model construction: Graph Neural Network (GNN): Use GraphSAGE and GAT models to learn the graph structure and capture the relationship between nodes. Attention mechanism: Introduce the attention mechanism to highlight the influence of key nodes and edges and improve the interpretability of the model. Multi-task learning: Simultaneously learn fault identification and health status assessment tasks to improve the overall performance of the model.

[0076] Model training: Data labeling: Label the graph structure data and mark the healthy and faulty states. Training process: Use the PyTorchGeometric library to train the model and use the cross-validation method to evaluate the performance of the model. Hyperparameter optimization: Use the GridSearchCV and RandomizedSearchCV methods to optimize the model's hyperparameters and improve the model's generalization ability.

[0077] Health status assessment: Real-time monitoring: Deploy the model to the actual system to monitor the health status of cross-domain data sets in real time. Health score: Evaluate the overall health status of the system through the health score output by the model. Fault prediction: Based on historical data and model prediction, early warning of potential fault risks.

[0078] Performance testing and comparative experiments

[0079] Dataset: Use public cross-domain datasets, such as AWS CloudWatch logs, Google Cloud Monitoring logs, etc. Build simulated cross-domain datasets that cover a variety of fault types and scenarios.

[0080] Performance indicators: Accuracy: The accuracy of the model in identifying faults. Recall: The recall rate of the model in identifying faults. F1 score: An evaluation indicator that takes into account both accuracy and recall. Response time: The response time of the model in detecting faults. Health score: The health of the model in evaluating the system.

[0081] Comparative experiment: Traditional method: Use traditional fault detection and health status assessment methods, such as threshold method, statistical analysis method, etc. Machine learning method: Use traditional machine learning methods, such as support vector machine (SVM), random forest (RF), etc. Deep learning method: Use traditional deep learning methods, such as DNN, CNN, etc. Graph neural network method: Use traditional graph neural network methods, such as GCN, GAT, etc.

[0082] Experimental results: Accuracy: The deep learning method and graph neural network method of the present invention are significantly superior to traditional methods and machine learning methods in terms of accuracy. Recall rate: The deep learning method and graph neural network method of the present invention are significantly superior to traditional methods and machine learning methods in terms of recall rate. F1 score: The deep learning method and graph neural network method of the present invention are significantly superior to traditional methods and machine learning methods in terms of F1 score. Response time: The deep learning method and graph neural network method of the present invention are significantly superior to traditional methods and machine learning methods in terms of response time. Health score: The graph neural network method of the present invention is significantly superior to traditional methods and machine learning methods in terms of health score.

[0083] Example 4

[0084] Data acquisition module: responsible for collecting raw data from data sources in different fields, including sensor data, equipment operation logs, performance indicator data, etc. These data come from a wide range of sources. For example, in the field of industrial production, data from temperature sensors, pressure sensors, vibration sensors, etc. on production line equipment, as well as log information such as equipment operation time and production output can be collected; in the field of transportation, data such as vehicle speed sensors, engine speed sensors, and fault code records can be obtained. The collected data formats are diverse, which may be structured database data, semi-structured text data, or unstructured image data. The collected data is pre-processed, such as data cleaning, removing noise data, outliers, and duplicate data. For example, for outliers in sensor data that are obviously deviated from the normal range, a statistical analysis-based method can be used to identify and eliminate them; garbled characters and erroneous characters in text data can be corrected or deleted. At the same time, a unified data format conversion is performed to convert data of different formats into a format suitable for subsequent processing, such as converting all data into a numerical matrix form to facilitate mathematical operations and model processing.

[0085] Feature extraction module: Different feature extraction methods are used for different types of data. For numerical sensor data, a method combining principal component analysis (PCA) and local linear embedding (LLE) is used. PCA can effectively reduce the dimension of data, extract the main characteristic components of data, and remove redundant information in data; LLE focuses on mining the local nonlinear structural characteristics of data, and can better retain the local geometric characteristics of data in low-dimensional space, thereby extracting more representative fault features. For example, for vibration sensor data of industrial equipment, PCA can be used to convert high-dimensional vibration frequency, amplitude and other data into low-dimensional principal component features, and then LLE is used to further optimize these features so that they can better reflect the fault status of the equipment. For text-based equipment operation log data, a word vector model combined with a text convolutional neural network (TextCNN) method is used for feature extraction. First, the word vector model is used to convert words in the text into low-dimensional vector representations. For example, the Word2Vec model can learn the semantic vector representation of words based on the text corpus. Then, these word vectors are input into TextCNN, which automatically extracts key semantic features in the text through convolutional layers and pooling layers, such as equipment fault type, fault occurrence time and other related information. For image data, such as images of the appearance of equipment or scanned images of its internal structure, use an image feature extraction network based on deep learning, such as the ResNet network. The residual structure in the ResNet network can effectively solve the gradient vanishing problem in deep neural networks, enabling it to learn the deep features of the image, such as the performance of fault features such as wear and deformation of the equipment in the image.

[0086] Domain Adaptation Module: A new domain adaptation algorithm is proposed, namely, a domain adaptation algorithm based on Generative Adversarial Network and Dynamic Distribution Matching (GAN-DDM). The core idea of ​​this algorithm is to gradually approximate the data distribution of the source domain and the target domain in the feature space through the generator and discriminator structure of the Generative Adversarial Network (GAN). The specific operation process is as follows: First, the feature data of the source domain and the target domain after feature extraction are input into the generator and discriminator of GAN respectively. The task of the generator is to generate data similar to the target domain based on the distribution characteristics of the target domain, and the discriminator is responsible for distinguishing the real target domain data from the data generated by the generator. During the training process, the generator continuously adjusts the generated data to make it closer to the target domain data, and the discriminator continuously improves the distinguishing ability. At the same time, a dynamic distribution matching mechanism is introduced to calculate the dynamic distance between the source domain and the target domain data in the feature space, such as using the Wasserstein distance, and adjust the training weights of the generator and the discriminator according to the distance, so that the distribution of the source domain and the target domain can be more accurately matched during the domain adaptation process. For example, in industrial equipment fault diagnosis, when the equipment failure data in one field is used as the source domain and the failure data of another different type of equipment but with similar failure mechanism is used as the target domain, the GAN-DDM algorithm can effectively convert the features of the source domain data into features suitable for the distribution of the target domain data, thereby realizing the migration of the fault diagnosis model.

[0087] Health status recognition module: Build a health status recognition model based on a bidirectional long short-term memory network and an attention mechanism (BiLSTM-AM). The BiLSTM network can process both the forward and reverse sequence information of the data. For fault data with time series, such as the vibration data sequence and temperature change data sequence during the operation of the equipment, it can better capture the long-term dependencies in the data. For example, when analyzing the vibration data of the engine, the BiLSTM can learn the association between the vibration data at different times, so as to more accurately judge the health status of the engine. The attention mechanism can automatically focus on the more critical part of the data for health status recognition during the data processing of the BiLSTM. For example, in the equipment operation data, abnormal data at certain specific times may play a decisive role in fault diagnosis. The attention mechanism can highlight the importance of these data and improve the accuracy of health status recognition. By training on a large number of cross-domain data sets, the model can learn the mapping relationship between fault features and health status in data from different fields, thereby realizing health status recognition of unknown data.

[0088] Result output and visualization module: Output the results obtained by the health status identification module, including the health status category of the equipment, such as normal, minor fault, major fault, etc., and the corresponding confidence score, indicating the model's credibility of the identification result. For example, for a certain industrial equipment, the model output result is "minor fault, confidence 0.7", which means that the model judges that the equipment is in a minor fault state and is 70% sure of the judgment result. Visualize the output results and use intuitive charts to show the change trend of the equipment's health status over time, the distribution ratio of different fault types, and other information. For example, use a line chart to show the change of the equipment's health status over a period of time, and use a pie chart to show the proportion of various fault types in all faults, so that equipment maintenance personnel can intuitively understand the operating status of the equipment and make maintenance decisions in a timely manner.

[0089] Method flow: Data collection and preprocessing steps: Collect raw data from multiple data sources in different fields, covering various types of data, such as sensor data, text data, image data, etc. Perform comprehensive cleaning on the collected data to remove noise, outliers, and duplicate data. Use a variety of data cleaning techniques, such as outlier detection and removal based on statistical analysis, data deduplication algorithms, etc. For example, for sensor data, by calculating the mean and standard deviation of the data, data that exceeds a certain multiple of the standard deviation range is considered an outlier and removed; for text data, use a text similarity algorithm to detect and delete duplicate text records.

[0090] The cleaned raw data is converted into a unified format and converted into a numerical matrix or vector form suitable for subsequent processing. For example, text data is converted into vector representation through a word vector model, and image data is converted into a pixel matrix, etc., to ensure that different types of data can be operated under the same data processing framework.

[0091] Feature extraction steps: For numerical data, principal component analysis (PCA) is first used to perform preliminary dimensionality reduction and main feature extraction, and the number of principal components to be retained is determined to balance the feature extraction effect and computational cost. For example, the number of principal components that retain more than 90% of the variance is determined based on the variance contribution rate of the data. Then, local linear embedding (LLE) is used to further perform nonlinear feature mining on the principal components extracted by PC, and the feature representation is optimized so that the extracted features can better reflect the fault information in the data. For text data, a word vector model (such as Word2Vec) is first used to map the words in the text into low-dimensional vectors. When training the word vector model, a large-scale text corpus in related fields is used to improve the accuracy and generalization ability of the word vector. Then, the word vector is input into the text convolutional neural network (TextCNN). By adjusting the parameters of TextCNN such as the convolution kernel size, step size, and pooling method, the key semantic features in the text, such as fault type, fault occurrence time, and other information, are extracted. For image data, a pre-trained ResNet network is used for feature extraction. First, the ResNet network is fine-tuned appropriately according to the characteristics of the image data, such as adjusting the size of the network's input layer to adapt to image data of different resolutions, and modifying the number of output categories of the last layer to adapt to the fault classification task. Then, the image data is input into the fine-tuned ResNet network to extract the deep fault features in the image, such as the performance of equipment wear, deformation, cracks and other features in the image.

[0092] Domain adaptation step: A domain adaptation algorithm based on Generative Adversarial Network and Dynamic Distribution Matching (GAN-DDM) is used. First, the parameters of the generator and discriminator of GAN are initialized, and appropriate training parameters such as learning rate and number of iterations are set. For example, the learning rate is set to 0.001 and the number of iterations is set to 10,000. The feature data of the source domain and the target domain after feature extraction are respectively input into the generator and discriminator of GAN for training. During the training process, the generator generates fake target domain data according to the distribution characteristics of the target domain, and the discriminator distinguishes the real target domain data from the data generated by the generator. By calculating the Wasserstein distance between the data generated by the generator and the real target domain data, the training weights of the generator and the discriminator are dynamically adjusted. For example, if the Wasserstein distance between the generated data and the real data is large, the training weight of the generator is increased so that it can learn the distribution characteristics of the target domain faster; if the accuracy of the discriminator is too high, the training weight of the discriminator is appropriately reduced to promote the learning of the generator. The training is continuously iterated until the distribution difference between the data generated by the generator and the real target domain data in the feature space is minimized, that is, the Wasserstein distance is less than the set threshold, or the preset number of training iterations is reached, completing the domain adaptation process.

[0093] Health status recognition steps: Construct a health status recognition model based on a bidirectional long short-term memory network and an attention mechanism (BiLSTM-AM). First, determine the structural parameters of the BiLSTM network, such as the number of layers, the number of nodes in each layer, and the calculation method of the attention mechanism, such as additive attention or dot product attention. For example, set the BiLSTM network to 2 layers, the number of nodes in each layer to 128, and the additive attention mechanism. Input the source domain and target domain data after domain adaptation into the BiLSTM-AM model for training, use the cross entropy loss function as the loss function of the model, and continuously adjust the parameters of the model through the back propagation algorithm, such as the weights and biases of the BiLSTM network, and the parameters of the attention mechanism. For example, during the training process, adjust the update step size of the parameters according to the value of the loss function. When the loss function value decreases slowly, appropriately reduce the update step size to improve the convergence speed and stability of the model. After the training is completed, input the unknown data to be identified into the trained BiLSTM-AM model, and the model outputs the health status category of the device and the corresponding confidence score.

[0094] For the first time, principal component analysis (PCA) and local linear embedding (LLE) are combined and applied to feature extraction of numerical data, giving full play to the advantages of PCA dimensionality reduction and denoising and LLE mining local nonlinear structures. It can more accurately extract fault features hidden in complex numerical data, effectively improve the representativeness and discrimination of fault features, and significantly improve the accuracy of subsequent fault diagnosis and health status recognition compared with traditional single feature extraction methods. For text data, the word vector model combined with the text convolutional neural network (TextCNN) is used for feature extraction. This method can make full use of the good representation ability of the word vector model for text semantics and the ability of TextCNN to extract text structural features, and extract richer and more accurate fault-related semantic information from text-based equipment operation logs and other data, overcoming the shortcomings of traditional text feature extraction methods that are difficult to capture the deep semantics and structural information of text. In terms of image data feature extraction, the pre-trained ResNet network is used and targeted fine-tuning is performed. The residual structure of the ResNet network enables it to effectively learn the deep features of the image. By fine-tuning to adapt to different fault diagnosis tasks, it can better extract fault features in image data, such as the subtle manifestations of equipment wear and deformation in the image, providing more powerful support for fault diagnosis based on image data. The proposed domain adaptation algorithm based on generative adversarial network and dynamic distribution matching (GAN-DDM) is a new domain adaptation strategy. By introducing the dynamic distribution matching mechanism, under the framework of the generative adversarial network, the distribution difference between the source domain and target domain data in the feature space can be more accurately measured, and the training weights of the generator and discriminator can be dynamically adjusted according to the difference, so that the distribution of the source domain data in the feature space can be faster and more accurately approximated to the distribution of the target domain data, effectively solving the problem of inaccurate distribution matching and poor migration effect of traditional domain adaptation algorithms when processing complex cross-domain data. The constructed health status recognition model based on bidirectional long short-term memory network and attention mechanism (BiLSTM-AM) has unique advantages. The BiLSTM network can process the temporal information of data in both directions and capture the long-term dependencies in the data, which is suitable for processing fault data with temporal characteristics. The attention mechanism can automatically focus on the key parts of the data and highlight the data features that have an important impact on health status recognition. The combination of the two can significantly improve the accuracy and reliability of health status recognition, especially when processing complex and long-sequence fault data. The system and method of the present invention can effectively integrate data from different fields and types, such as numerical, textual and image data.By designing a unified data collection, preprocessing, feature extraction and domain adaptation process, various types of data are integrated into a complete fault diagnosis and health status identification framework, making full use of the fault information contained in different types of data, breaking the limitation that traditional fault diagnosis methods can only process a single type of data, and realizing multi-source data-driven cross-domain fault migration health status identification, which greatly improves the comprehensiveness and accuracy of fault diagnosis. The result output and visualization module can not only output the health status category and confidence score of the equipment, but also visualize these results in the form of intuitive charts, such as line charts showing the trend of health status changes over time, and pie charts showing the distribution of fault types. This visualization method can provide equipment maintenance personnel with more intuitive and comprehensive equipment operation information, which helps them to timely discover potential fault risks, formulate reasonable maintenance plans, and improve the efficiency and scientificity of equipment maintenance. It has important practical application value in industrial production, transportation and other fields.

[0095] Example 5

[0096] Data collection and preprocessing Data collection: Collect data from substations, transmission lines and other equipment in the smart grid. Including current and voltage numerical data collected by current transformers and voltage transformers (sampling frequency is 1000 times per second); operation log text data recorded by the equipment operation monitoring system (key operation information is recorded every 5 minutes); equipment temperature distribution image data taken by infrared thermal imagers (taken once every 1 hour).

[0097] Data cleaning: For current and voltage data, an outlier detection method based on dynamic thresholds is used. According to the fluctuation range of current and voltage during normal operation of the equipment, combined with historical data statistical information, the threshold is adjusted dynamically. For example, when the current value exceeds the normal mean ±3 times the standard deviation range and lasts for more than 0.5 seconds, it is determined to be an outlier and removed. For operation log text data, regular expressions in natural language processing technology are used to remove non-critical information (such as parts with irregular date formats, some common meaningless words, etc.), and duplicate records are detected and deleted through semantic similarity algorithms. For infrared thermal imaging image data, a noise removal method based on image contrast is used to enhance image quality.

[0098] Data format conversion: Current and voltage data are organized into numerical matrices in time series, with each row representing the current and voltage values ​​at a sampling moment. Operation log text data is converted into vector representations using a pre-trained domain-specific word vector model (such as the Word2Vec model trained based on a large number of power grid equipment operation manuals), and each text record is converted into a fixed-length vector sequence. Infrared thermal imaging image data is converted into a pixel matrix after pre-processing such as cropping and scaling.

[0099] Feature extraction, feature extraction of numerical data (current, voltage): First, the independent component analysis (ICA) algorithm is used to decompose the multivariate current and voltage data into independent components, and the main independent components are extracted as preliminary features to reduce data redundancy and highlight potential fault-related features. For example, the independent components with the top 80% energy are determined based on the energy contribution rate of the components. Then the local preserving projection (LPP) method is used to further process the features extracted by ICA. LPP can reduce the dimension while maintaining the local structure of the data, so that the extracted features can better reflect the fault characteristics of the power grid equipment under the local operating state. For example, the data is projected into a low-dimensional space (such as 5 dimensions) while retaining the local neighborhood relationship between data points. Feature extraction of text data (operation log): The sequence-to-sequence (Seq2Seq) model based on the recurrent neural network (RNN) is used for feature extraction. The Seq2Seq model can learn the temporal semantic information in the operation log text and convert the input text sequence into a low-dimensional semantic vector representation. For example, the model's encoder uses LSTM units, and the decoder uses an attention mechanism to focus on the key operation steps in the text, outputting a 128-dimensional semantic vector that represents the key features of the operation log. When training the Seq2Seq model, the sliding window technology is used to take a text segment of a certain length (such as 10 words) as input, and gradually slide the window to cover the entire operation log text to improve the accuracy of feature extraction.

[0100] Image data (infrared thermal imaging) feature extraction: Use image enhancement technology based on generative adversarial networks (GAN) to enhance the original infrared thermal imaging images and increase the image details. For example, train a GAN model, the generator is responsible for generating clearer and more detailed images, and the discriminator is used to distinguish between real images and generated images. The quality of generated images is improved through adversarial training. Use densely connected convolutional networks (DenseNet) to extract features from the enhanced images. The densely connected structure in the DenseNet network can make full use of feature information at different levels to extract more discriminative image features. For example, adjust the network depth and growth rate of DenseNet to adapt it to the feature extraction task of infrared thermal imaging images and output a 256-dimensional image feature vector.

[0101] Domain Adaptation A domain adaptation algorithm based on Multimodal Adversarial Learning and Optimal Transfer Theory (MAL-OT) is proposed. First, the feature data of the source domain (such as the test data of power grid equipment in a laboratory environment) and the target domain (such as the data of smart grid equipment in actual operation) are mapped to a common feature space after feature extraction. In this common feature space, multiple adversarial generative network (GAN) structures are constructed, and each GAN is responsible for learning the distribution difference between the source domain and the target domain in a mode (such as the frequency distribution mode of the data, the amplitude distribution mode, etc.). Through adversarial training, the fake target domain data generated by the generator is similar to the distribution of the real target domain data in this mode, and the discriminator tries its best to distinguish them.

[0102] At the same time, the Wasserstein distance in the optimal transfer theory is introduced to measure the overall distribution difference between the source domain and the target domain in multimodality. According to the calculated Wasserstein distance, the training weights of each GAN are dynamically adjusted so that the distribution of the source domain and the target domain can be matched more comprehensively and accurately during the domain adaptation process. For example, if the Wasserstein distance between the source domain and the target domain is large in a certain mode, the training weight of the corresponding GAN is increased to accelerate the distribution matching in this mode. The training is continuously iterated until the distribution difference between the source domain and the target domain in multimodality is minimized (that is, the Wasserstein distance is less than the set threshold), or the preset number of training iterations (such as 15,000) is reached, and the domain adaptation process is completed.

[0103] Health status recognition: A health status recognition model based on gated recurrent unit and hierarchical attention mechanism (GRU-HAM) was constructed. The structural parameters of the GRU network were determined and set to 3 layers of GRU with 64 hidden units in each layer. The hierarchical attention mechanism is divided into two layers. The first layer focuses on the importance between features of different types of data (numerical, textual, and image), and the second layer focuses on the key parts within the features of the same type of data. The source and target domain data after domain adaptation are input into the GRU-HAM model for training. The weighted cross entropy loss function is used to assign different weights according to the importance of different types of data features in fault diagnosis. For example, a higher weight (such as 0.6) is assigned to numerical data features because they directly reflect the electrical operating status of the equipment. During the training process, the model parameter update is optimized through an adaptive learning rate adjustment strategy (such as the Adagrad algorithm). The learning rate is dynamically adjusted according to the trend of the loss function. When the loss function decreases slowly, the learning rate is appropriately reduced to improve the convergence speed and stability of the model. After the training is completed, the data of the smart grid equipment to be identified is input into the model, and the model outputs the health status category of the equipment (normal, minor fault, moderate fault, severe fault) and the corresponding confidence score (ranging from 0 to 1).

[0104] Example 6

[0105] Data collection and preprocessing: Data collection collects data from sensors of aircraft engines, such as temperature sensors (measuring the temperature of various parts of the engine, with a sampling frequency of 50 times per second), pressure sensors (measuring intake pressure, exhaust pressure, etc., with a sampling frequency of 50 times per second), and speed sensors (measuring rotor speed, with a sampling frequency of 100 times per second) to obtain numerical data; obtains text data such as maintenance operation instructions and fault repair history from the engine maintenance record system (recorded after each maintenance operation); and obtains image data of internal engine components through engine endoscopic inspection equipment (inspected once every 100 hours of flight).

[0106] Data cleaning: For numerical data, an outlier detection method based on Hampel identifiers is used. The median and median absolute difference (MAD) of the data are calculated. When the deviation of a data point from the median exceeds a certain multiple (such as 3.5 times) of MAD, it is determined to be an outlier and removed. At the same time, the data is smoothed and the moving average method is used to remove high-frequency noise. For text data, text classification technology is used to classify maintenance instructions and fault repair history into different categories (such as lubrication system maintenance, combustion system fault repair, etc.), and then keyword extraction (using TF-IDF algorithm) is performed on each type of text to remove low-frequency keywords and retain keywords closely related to engine failures and their contextual information. For image data, an image denoising method based on wavelet transform is used to improve image clarity. Image segmentation is then performed to separate the internal components of the engine from the background to facilitate subsequent feature extraction.

[0107] Data format conversion: Numerical data is organized into a multidimensional array according to the sensor type and acquisition time. For example, temperature, pressure, and speed data are arranged in chronological order into a three-dimensional array, with each dimension representing a different sensor type and time point. Text data is converted into a vector representation through the bag-of-words model, with each keyword as a dimension and its word frequency as the vector value. At the same time, the Word2Vec model is used to initialize the keyword vector to incorporate semantic information.

[0108] The image data converts the segmented component images into grayscale images and resizes them to a uniform size (eg, 256×256 pixels), and then organizes each pixel value into a two-dimensional matrix as the format of the image data.

[0109] Feature extraction, feature extraction of numerical data: Use the kernel principal component analysis (KPCA) algorithm to extract features from numerical data. KPCA can map data to a high-dimensional feature space, in which the kernel function (such as the Gaussian kernel function) is used to calculate the principal components of the data and extract nonlinear features. For example, by selecting appropriate kernel parameters through cross-validation, the data is projected into a low-dimensional space (such as 10 dimensions) to retain more than 95% of the variance contribution rate. The features extracted by KPCA are further processed using the Hilbert-Huang transform (HHT). HHT can decompose nonlinear and non-stationary data into a finite number of intrinsic mode functions (IMFs), and by analyzing the energy, frequency and other characteristics of the IMFs, a feature vector that can better reflect the fault characteristics under the dynamic operating state of the engine is extracted. For example, the energy entropy of each IMF is calculated as the eigenvalue to form the final numerical feature vector.

[0110] Feature extraction of text data: Use a text feature extraction method based on a graph convolutional network (GCN). Construct text data into a graph structure, where words are nodes and semantic relationships between words (such as co-occurrence relationships, semantic similarity relationships, etc.) are edges. GCN learns the semantic structural features of text by performing convolution operations on the graph. For example, set the number of convolutional layers of GCN to 3, the number of filters in each layer to 64, and output a 128-dimensional text feature vector, which represents the key features of engine maintenance and fault text.

[0111] When constructing the text graph structure, the knowledge graph technology is used to introduce external knowledge (such as the correlation between engine components, the causal relationship between fault modes, etc.), enrich the semantic information of the text, and improve the accuracy of feature extraction. Image data feature extraction: An image feature extraction method based on variational autoencoder (VAE) is adopted. VAE can learn the latent representation space of image data, encode the image into mean and variance vectors through the encoder, and then generate a reconstructed image through the decoder. The reconstruction error is used to optimize the parameters of the encoder and decoder to extract the latent features of the image. For example, the image data is encoded into a 128-dimensional latent vector, which can reflect the structure and state characteristics of the internal components of the engine.

[0112] Perform cluster analysis on the features extracted by VAE (such as using K-means clustering algorithm) to divide the image features into different categories, each category representing a state mode of internal engine components (such as normal wear mode, abnormal wear mode, etc.), further refining the image feature representation.

[0113] Domain adaptation, a domain adaptation algorithm based on conditional adversarial domain adaptation network and manifold alignment (CADAN-MA) is proposed. First, the feature data of the source domain (such as engine ground test data) and the target domain (such as actual engine operation data in flight) after feature extraction are input into the CADAN-MA network. In the network, a conditional adversarial generative network (cGAN) structure is constructed, in which the generator generates feature data similar to the target domain based on the source domain features and a condition vector (such as engine operating conditions, such as flight altitude, speed, etc.), and the discriminator determines whether the input data is real target domain data or data generated by the generator based on the condition vector. Through adversarial training, the generator can generate feature data that is more consistent with the target domain distribution according to different conditions. At the same time, the manifold alignment technology is introduced to map the feature data of the source domain and the target domain to a common manifold space. In this space, the data of the source domain and the target domain are kept consistent in the manifold structure by minimizing the manifold distance (such as geodesic distance) between the source domain and the target domain. For example, the Laplace eigenmap algorithm is used to calculate the manifold distance and add it as a constraint condition to the training process of cGAN, adjusting the parameters of the generator and discriminator so that the distribution of the source domain and the target domain in the manifold space is more matched while considering the conditional factors.

[0114] The training is continuously iterated until certain convergence conditions are met (such as the discrimination accuracy between the data generated by the generator and the real target domain data reaches more than 95%, or the manifold distance is less than the set threshold), completing the domain adaptation process.

[0115] Health status recognition, build a health status recognition model based on spatiotemporal convolutional network and multi-head attention mechanism (STCN-MAM). Design the STCN-MAM network structure. The spatiotemporal convolutional network is used to process the time series of numerical data and the spatial structure information of image data. Set the convolution kernel size and step size of the spatiotemporal convolution layer to adapt to the characteristic scale of the data. For example, for the time series of numerical data, a 3×1 convolution kernel with a step size of 1 is used to extract the characteristic relationship between different time points; for image data, a 3×3 convolution kernel with a step size of 2 is used to extract the local and global features of the image.

[0116] The multi-head attention mechanism is used to fuse the features of numerical, textual and image data. The number of heads is set to 4, each head focuses on different aspects of feature information, and then the outputs of the multiple heads are spliced ​​to obtain the final fused feature representation. The source domain and target domain data after domain adaptation are input into the STCN-MAM model for training. A multi-label classification loss function is used because the engine health state may have multiple fault types at the same time (such as temperature anomalies and vibration anomalies may occur at the same time). The model parameters are adjusted through the back propagation algorithm and the stochastic gradient descent optimization algorithm (such as SGD with momentum) so that the model can accurately identify the health state of the engine. After training, the aircraft engine data to be identified is input into the model, and the model outputs the engine health state category (such as normal, temperature anomaly, pressure anomaly, mechanical component wear, etc.) and the corresponding confidence score (value range is 0-1).

[0117] Efficiency test comparison, comparison index setting Fault diagnosis accuracy: the proportion of equipment fault types accurately identified. Health status identification accuracy: the proportion of equipment health status categories (such as normal, different degrees of fault, etc.) accurately judged. False alarm rate: the proportion of normal equipment mistakenly judged as faulty equipment. Missing alarm rate: the proportion of faulty equipment mistakenly judged as normal equipment. Average fault identification time: the average time from data collection to accurate fault identification.

[0118] Experimental design and data collection. In the smart grid equipment test, 100 substation equipment were selected as test objects, of which 50 devices used traditional fault diagnosis methods (based on single type data and simple threshold judgment), and 50 devices used the method (implementation) of the present invention. During the 6-month test period, the actual operation data and fault occurrence of the equipment were collected. For the aircraft engine test, 30 aircraft engines of the same model were selected, 15 engines used traditional methods (based on empirical rules in the engine manual and single sensor data diagnosis), and 15 engines used the method (implementation) of the present invention. During a complete maintenance cycle of the engine (such as 1000 flight hours), the engine's operating data and fault events were recorded.

[0119] Results analysis, smart grid equipment test results: the fault diagnosis accuracy of the method of the present invention reaches 92%, while the traditional method is only 75%. In terms of health status recognition accuracy, the method of the present invention is 88%, and the traditional method is 68%. The false alarm rate of the method of the present invention is 5%, while the traditional method is 15%; the missed alarm rate is 3%, and the traditional method is 10%. In terms of average fault identification time, the method of the present invention is 2 minutes, and the traditional method is 5 minutes. Aircraft engine test results: The fault diagnosis accuracy of the method of the present invention reaches 90%, while the traditional method is 70%. In terms of health status recognition accuracy, the method of the present invention is 85%, and the traditional method is 65%. In terms of false alarm rate, the method of the present invention is 6%, and the traditional method is 18%; the missed alarm rate is 4%, and the traditional method is 12%. In terms of average fault identification time, the method of the present invention is 10 minutes, and the traditional method is 20 minutes. From the above efficiency test comparison, it can be seen that the method of the present invention is significantly superior to the traditional method in terms of fault diagnosis accuracy, health status recognition accuracy, false alarm rate, missed alarm rate and average fault identification time, and can more effectively identify the faults and health status of the equipment, providing more reliable support for equipment maintenance and operation.

[0120] The present invention has been described in detail above in conjunction with the embodiments. However, those skilled in the art will appreciate that, without departing from the spirit of the present invention, the specific parameters in the above embodiments may be changed to form a plurality of specific embodiments, which are all within the common variation range of the present invention and will not be described in detail herein.

Claims

1. A cross-domain dataset fault migration health status identification method, characterized in that: include: S1 data collection, extensively collect raw data, and pre-process the collected raw data, converting data in different formats into numerical matrix or vector form for subsequent processing; S2 feature extraction, including the use of principal component analysis combined with local linear embedding for feature extraction of numerical data, the use of word vector model combined with text convolutional neural network for feature extraction of text data, and the use of pre-trained ResNet network for image data and targeted fine-tuning for feature extraction; S3 domain adaptation uses a domain adaptation algorithm based on a generative adversarial network and dynamic distribution matching. The feature data of the source domain and the target domain after feature extraction are input into the generator and discriminator of the GAN for training respectively. The training weights of the generator and the discriminator are adjusted using a dynamic distribution matching mechanism so that the distribution of the source domain data in the feature space is close to that of the target domain data. S4 health status recognition, building a health status recognition model based on a bidirectional long short-term memory network and an attention mechanism, inputting the source domain and target domain data after domain adaptation into the BiLSTM-AM model for training, for health status recognition of unknown data, and outputting the health status category of the device and the corresponding confidence score; S5 result output and visualization is used to output the results obtained by the health status identification module and visualize the output results in the form of intuitive charts to show the change trend of the health status of the equipment over time and the distribution ratio information of different fault types.

2. The cross-domain dataset fault migration health status identification method according to claim 1 is characterized in that: During the data collection process, for sensor data, the mean and standard deviation of the data are calculated, and data beyond the standard deviation range is regarded as outliers and removed; for text data, a text similarity algorithm is used to detect and delete duplicate text records.

3. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: The health status identification also adopts a cross-domain dataset health status assessment model of a graph neural network, abstracting the cross-domain dataset into a graph structure, with nodes representing data sources and servers, and edges representing data transmission paths; extracting the topological features of the graph, using the graph neural network to learn the graph structure, capturing the relationships between nodes, introducing an attention mechanism, highlighting the influence of key nodes and edges, and simultaneously learning fault identification and health status assessment tasks.

4. The cross-domain dataset fault migration health status identification method according to claim 3 is characterized in that: It also includes labeling graph structure data, marking health and fault states, training models using labeled data, evaluating model performance using cross-validation methods, and optimizing model hyperparameters through grid search and random search methods.

5. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: When the domain adaptation adopts the domain adaptation algorithm based on the generative adversarial network and dynamic distribution matching, the parameters of the generator and the discriminator of the GAN are initialized, and the Wasserstein distance between the data generated by the generator and the real target domain data is calculated. The training weights of the generator and the discriminator are dynamically adjusted according to the distance until the distribution difference between the data generated by the generator and the real target domain data in the feature space is minimized, and the domain adaptation process is completed.

6. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: When constructing the health status recognition model based on the bidirectional long short-term memory network and the focus mechanism, the BiLSTM network is set to 2 layers, with 128 nodes in each layer, and an additive attention mechanism is adopted. The source domain and target domain data after domain adaptation are input into the BiLSTM-AM model for training. The cross entropy loss function is used as the loss function of the model, and the parameters of the model are continuously adjusted through the back propagation algorithm.

7. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: A domain adaptation algorithm based on generative adversarial networks and dynamic distribution matching is adopted. The feature data of the source domain and the target domain after feature extraction are input into the generator and the discriminator respectively. The training weights of the generator and the discriminator are dynamically adjusted by calculating the Wasserstein distance between the data generated by the generator and the real target domain data.

8. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: It also includes building a health status recognition model based on a bidirectional long short-term memory network and attention mechanism, which identifies the health status category of the equipment and the corresponding confidence score by bidirectionally processing the timing information of the data and automatically focusing on the key parts.

9. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: It also includes constructing multiple adversarial generative network structures, each of which is responsible for learning the distribution difference between the source domain and the target domain in a modality. Through adversarial training, the fake target domain data generated by the generator is similar to the distribution of the real target domain data in this modality, and the discriminator distinguishes them.

10. The cross-domain dataset fault migration health status identification method according to claim 1, characterized in that: It also includes the introduction of the Wasserstein distance in the optimal transfer theory to measure the overall distribution difference between the source domain and the target domain under multimodality, dynamically adjusting the training weights of each adversarial generative network according to the calculated Wasserstein distance, and continuously iterating the training until the preset conditions are met.

Images