A remote acquisition system for gas equipment operation data

Abstract

The present invention provides a remote collection system for gas equipment operation data, which relates to the field of data processing technology, including: by learning the long-term dependency in the sequence, capturing the law of the change of the gas equipment state over time, and generating key features containing time series information; fusing the time series feature representation output by the long short-term memory network with the equipment attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive state of the gas equipment; comprehensively analyzing the feature vector to obtain hidden patterns related to the gas equipment risk; based on the hidden patterns related to the gas equipment risk, using a multi-classification support vector machine to predict the risk level to obtain the final risk level prediction result. The present invention can realize the risk assessment of the gas equipment operation status, thereby improving the safety and operation efficiency of the equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a remote collection system for gas equipment operation data. Background Art

[0002] With the advancement of science and technology and the rapid development of industry, the application of gas equipment in daily life and industrial production is indeed becoming more and more extensive. However, since the operation of gas equipment involves dangerous factors such as high pressure and flammability, its safety and stability have become important concerns. In order to solve this problem, real-time monitoring and risk assessment of the operating status of gas equipment has become an indispensable link.

[0003] Traditional gas equipment monitoring methods, such as manual inspections and regular maintenance, can ensure the safe operation of equipment to a certain extent, but this method has obvious limitations. First, manual inspections are inefficient and difficult to cover all equipment and operating periods, which can easily lead to missed inspections and false inspections. Second, although regular maintenance can conduct comprehensive inspections and repairs on equipment, it is difficult to respond in time when equipment fails or abnormal conditions occur, which can easily cause safety hazards. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a remote collection system for gas equipment operating data, which can realize risk assessment of the operating status of the gas equipment, thereby improving the safety and operating efficiency of the equipment.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A remote collection system for gas equipment operation data, comprising:

[0007] A graph construction module is used to perform preliminary feature extraction on preprocessed data using a convolutional neural network model to capture local features in gas equipment operation data, wherein the preprocessed data is preprocessed gas equipment operation data, including pressure, temperature, flow rate and leakage; extract deep features through a deep residual network based on local features; combine the deep features with equipment entities, attributes, relationships and timestamp information to construct a preliminary knowledge graph; jointly annotate entities and relationships in the preliminary knowledge graph using a conditional random field model, optimize the preliminary knowledge graph using a graph convolutional network, and update nodes through iteration;

[0008] A risk prediction module is connected to the graph construction module and receives a preliminary knowledge graph as input data; the long short-term memory network uses the time series data in the preliminary knowledge graph as sequence input data, captures the law of changes in the state of the gas equipment over time by learning the long-term dependencies in the sequence, and generates key features containing time series information; the time series feature representation output by the long short-term memory network is integrated with the equipment attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive state of the gas equipment; the feature vector is comprehensively analyzed to obtain hidden patterns related to the risks of the gas equipment; based on the hidden patterns related to the risks of the gas equipment, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction result;

[0009] A remote transmission module, connected to the risk prediction module, for remotely transmitting the final risk level prediction result and raw data to a cloud server or a user terminal via a wireless network;

[0010] The user interaction module is connected to the remote transmission module and is used to provide a visual interface to the user to display the real-time operating status of the gas equipment, the final risk level prediction results and historical data.

[0011] Furthermore, the convolutional neural network model is used to perform preliminary feature extraction on the preprocessed data to capture the local features in the gas equipment operation data, including:

[0012] Determine the basic architecture of the convolutional neural network model, including convolutional layers, pooling layers, and fully connected layers;

[0013] Define the parameters of the genetic algorithm, including population size, crossover rate, mutation rate, and number of generations;

[0014] Encode the hyperparameters of the convolutional neural network model as genes of the genetic algorithm to form an initial population. The hyperparameters include learning rate, batch size, and number of convolution kernels.

[0015] The preprocessed data is divided into training set, validation set and test set. For each generation of population, the following steps are performed:

[0016] Decode each individual in the population into the hyperparameter settings of a convolutional neural network model;

[0017] Train the convolutional neural network model using the training set and the current individual hyperparameter settings. In each training cycle, forward propagation, loss calculation, backpropagation and parameter update are performed;

[0018] After each training cycle, the validation set is used to evaluate the performance of the convolutional neural network model and the accuracy is recorded;

[0019] According to the accuracy rate on the validation set, the corresponding individuals are selected to enter the next generation population. Two individuals in the population are randomly selected, and some of their genes are exchanged to generate new individuals. The new individuals are used to replace some individuals in the original population to form the next generation population, until the preset evolutionary generation is reached and the genetic algorithm cycle is stopped.

[0020] Select the corresponding individuals from the last generation of population as the final hyperparameters, and retrain the convolutional neural network model using the final hyperparameter settings to obtain the trained convolutional neural network model;

[0021] The convolutional layer of the trained convolutional neural network model is used as a feature extractor, the input data is forward propagated through the convolutional layer, and the output of the convolutional layer is obtained as the extracted local feature.

[0022] Furthermore, the accuracy The calculation formula is:

[0023] ;

[0024] in, Yes Category The true positives, i.e., the categories The number of samples correctly predicted as positive; Yes Category The true negative of The number of samples correctly predicted as negative class; Yes Category The false positive of the category The number of negative samples that are incorrectly predicted as positive; Yes Category The false negatives of The number of positive samples that are incorrectly predicted as negative; is the total number of categories; is the mean of the Gaussian distribution, indicating a certain central location; is the standard deviation; Yes Category The number of samples; is the index value of the category.

[0025] Furthermore, based on the local features, deep features are extracted through a deep residual network; the deep features are combined with device entities, attributes, relationships and timestamp information to construct a preliminary knowledge graph, including:

[0026] The deep residual network is trained using local features, and the network parameters are optimized through back propagation to obtain the trained deep residual network;

[0027] Apply the trained deep residual network to the data of all local features to extract deep features;

[0028] Integrate the deep features with the corresponding device entity information, attributes, relationships and timestamp information, so that the feature data and metadata correspond one-to-one to obtain the integrated data;

[0029] Based on the integrated data, a preliminary equipment knowledge graph is constructed, in which nodes represent equipment entities, edges represent the relationships between entities, attribute information is attached to the corresponding nodes, and timestamp information is used to describe the dynamic changes of entities and relationships.

[0030] Furthermore, the conditional random field model is used to jointly annotate the entities and relationships in the preliminary knowledge graph, and the graph convolutional network is used to optimize the preliminary knowledge graph, and the nodes are updated through iteration, including:

[0031] Extract features for the conditional random field model from the preliminary knowledge graph, including part of speech, context information, and entity type;

[0032] Use the trained conditional random field model to jointly annotate entities and relationships on the preliminary knowledge graph and identify the relationships between entities in the preliminary knowledge graph.

[0033] The knowledge graph annotated by the conditional random field model is converted into graph structure data, where nodes represent entities and edges represent relationships. A feature matrix is ​​constructed for each node, which includes the attributes of the node and the information of adjacent nodes.

[0034] Use graph convolutional networks to optimize the graph and capture the deep connections and features between nodes through multi-layer graph convolution operations;

[0035] According to the deep associations and features between nodes, the entities and relationships in the graph are adjusted and optimized to obtain the optimization results; based on the optimization results, the status and attributes of each node in the graph are updated;

[0036] The graph convolutional neural network is repeatedly used for multiple iterations. Each iteration adjusts the graph according to the results of the previous one until the preset number of iterations is reached. After the iteration is completed, the graph is finally optimized and adjusted.

[0037] Furthermore, the long short-term memory network uses the time series data in the preliminary knowledge graph as sequence input data, captures the law of gas equipment status changes over time by learning the long-term dependencies in the sequence, and generates key features containing time series information, including:

[0038] Extracting time-series related data from the preliminary knowledge graph, including device status records and sensor readings arranged in chronological order;

[0039] Read the data related to the time series and set a corresponding time window size; divide the time series data according to the set time window, form a sequence of data in each window, and standardize each sequence to obtain the processed time series data sequence;

[0040] Determine the number of LSTM network layers, set the number of hidden units in each LSTM network layer, and determine the input and output dimensions of the network. The input dimension matches the dimension of the processed time series data sequence.

[0041] Use the deep learning framework to create an LSTM model, convert the processed time series data sequence into the tensor format required by the LSTM model, and divide the processed time series data sequence into training set, validation set, and test set;

[0042] Use the training set data to train the LSTM model. After each training round, use the validation set to evaluate the performance of the LSTM model. Adjust the learning rate and optimizer parameters based on the performance of the validation set. Use the back-propagation algorithm to update the weights and biases of the LSTM network to obtain the trained LSTM model.

[0043] The processed time series data sequence is input into the trained LSTM model, and the trained LSTM model outputs the hidden state of each time step. The hidden state contains the long-term dependencies and key features in the time series data.

[0044] Furthermore, the temporal feature representation output by the long short-term memory network is integrated with the equipment attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive status of the gas equipment. The feature vector is comprehensively analyzed to obtain hidden patterns related to the risks of the gas equipment, including:

[0045] Obtain the temporal feature representation of the LSTM network output, and extract static attributes and relational information related to gas equipment from the knowledge graph, where static attributes include equipment model, manufacturer, and installation date; relational information includes connection relationship with the equipment and maintenance records;

[0046] The temporal features output by the long short-term memory network are concatenated with the static features and relational features extracted from the knowledge graph to form a fused feature vector;

[0047] K-means is used to cluster the fused feature vectors to find groups with similar device status and behavior patterns;

[0048] The isolation forest algorithm is used to identify feature vectors that are different from the normal pattern. Different feature vectors represent potential risks.

[0049] Hidden patterns associated with gas equipment risks are identified based on static attributes and relationship information related to gas equipment, groups with similar equipment status and behavior patterns, and potential risks.

[0050] Furthermore, based on the hidden patterns related to gas equipment risks, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction results, including:

[0051] Assign a risk level label to each latent pattern feature vector based on historical data;

[0052] Set the parameters of the SVM model, including the kernel function, penalty coefficient C, and kernel function parameters;

[0053] Train the SVM model using the latent pattern feature set with risk level labels;

[0054] By dividing the training set and the validation set, the performance and generalization ability of the SVM model are evaluated to obtain the validation results;

[0055] Adjust the SVM model parameters according to the verification results to obtain the final SVM model;

[0056] For new gas equipment data, extract its hidden pattern feature vector;

[0057] The hidden pattern feature vector is input into the final SVM model to predict the risk level to obtain the final risk level prediction result.

[0058] Furthermore, the final risk level prediction result and the original data are remotely transmitted to the cloud server or the user terminal via a wireless network, including:

[0059] According to the network environment and transmission requirements, a particle swarm is initialized, and each particle represents a data transmission scheme, including transmission path, data compression rate and transmission protocol parameters;

[0060] Define a fitness function to evaluate the performance of each transmission scheme;

[0061] Update the speed and position of each particle according to the individual final solution and the global final solution, and obtain a data transmission scheme suitable for the current network environment and transmission requirements through multiple iterations;

[0062] After the iteration is completed, the optimized transmission scheme corresponding to the global final solution is selected;

[0063] Using the optimized transmission scheme, the final risk level prediction results and original data are remotely transmitted to the cloud server or user terminal via the wireless network.

[0064] Furthermore, the fitness function The calculation formula is:

[0065] ;

[0066] in, Indicates the normalized value of actual transmission time; Indicates the amount of data successfully transferred; Indicates the total amount of data transferred; Indicates the normalized value of the bandwidth used; Indicates the actual transmission rate; Indicates the maximum transmission rate of the network.

[0067] The above scheme of the present invention includes at least the following beneficial effects.

[0068] Through the graph construction module, the present invention can use convolutional neural networks and deep residual networks to efficiently and accurately extract features from the operating data of gas equipment, which not only captures the local features in the data, but also further extracts deep-level features.

[0069] The risk prediction module combines the long short-term memory network with the knowledge graph, enabling the system to capture the law of gas equipment status changes over time and generate key features containing time series information. This fusion method improves the ability to describe the comprehensive status of gas equipment, thereby more accurately predicting hidden patterns related to equipment risks and providing risk level prediction results.

[0070] The design of the remote transmission module enables the system to transmit risk level prediction results and raw data to the cloud server or user terminal in real time. This provides users with a convenient remote monitoring function, so that they can keep abreast of the operating status and risk of gas equipment no matter where they are.

[0071] The user interaction module provides a visual interface, allowing users to intuitively view the real-time operating status of gas equipment, risk level prediction results and historical data. This design not only improves the user experience, but also enhances the interactivity of the system, making it easier for users to perform corresponding operations and adjustments according to actual needs.

[0072] Through the implementation of the present invention, potential risks in the operation of gas equipment can be discovered in time, and corresponding preventive measures can be taken, thereby effectively avoiding the occurrence of equipment failures or safety accidents and improving the safety of gas equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a schematic diagram of a remote collection system for gas equipment operation data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0075] like Figure 1 As shown, an embodiment of the present invention provides a remote collection system for gas equipment operation data, comprising:

[0076] A graph construction module is used to perform preliminary feature extraction on preprocessed data using a convolutional neural network model to capture local features in gas equipment operation data, wherein the preprocessed data is preprocessed gas equipment operation data, including pressure, temperature, flow rate and leakage; extract deep features through a deep residual network based on local features; combine the deep features with equipment entities, attributes, relationships and timestamp information to construct a preliminary knowledge graph; jointly annotate entities and relationships in the preliminary knowledge graph using a conditional random field model, optimize the preliminary knowledge graph using a graph convolutional network, and update nodes through iteration;

[0077] A risk prediction module is connected to the graph construction module and receives a preliminary knowledge graph as input data; the long short-term memory network uses the time series data in the preliminary knowledge graph as sequence input data, captures the law of changes in the state of the gas equipment over time by learning the long-term dependencies in the sequence, and generates key features containing time series information; the time series feature representation output by the long short-term memory network is integrated with the equipment attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive state of the gas equipment; the feature vector is comprehensively analyzed to obtain hidden patterns related to the risks of the gas equipment; based on the hidden patterns related to the risks of the gas equipment, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction result;

[0078] A remote transmission module, connected to the risk prediction module, for remotely transmitting the final risk level prediction result and raw data to a cloud server or a user terminal via a wireless network;

[0079] The user interaction module is connected to the remote transmission module and is used to provide a visual interface to the user to display the real-time operating status of the gas equipment, the final risk level prediction results and historical data.

[0080] In an embodiment of the present invention, through the graph construction module, the present invention can use convolutional neural networks and deep residual networks to efficiently and accurately extract features from the operating data of gas equipment, which not only captures local features in the data, but also further extracts deep features. The risk prediction module combines the long short-term memory network with the knowledge graph, so that the system can capture the law of the change of the state of the gas equipment over time and generate key features containing time series information. This fusion method improves the ability to describe the comprehensive state of the gas equipment, so that the hidden patterns related to the equipment risk can be more accurately predicted, and the prediction results of the risk level can be given. The design of the remote transmission module enables the system to transmit the risk level prediction results and the original data to the cloud server or user terminal in real time. This provides users with a convenient remote monitoring function, so that they can grasp the operating status and risk situation of the gas equipment in a timely manner no matter where they are. The user interaction module enables users to intuitively view the real-time operating status, risk level prediction results and historical data of the gas equipment by providing a visual interface. This design not only improves the user experience, but also enhances the interactivity of the system, making it convenient for users to perform corresponding operations and adjustments according to actual needs. Through the implementation of the present invention, potential risks in the operation of gas equipment can be discovered in time, and corresponding preventive measures can be taken, thereby effectively avoiding the occurrence of equipment failures or safety accidents and improving the safety of gas equipment.

[0081] The pre-processed data is the operating data of the gas equipment after pre-processing, including pressure, temperature, flow and leakage, which involves the pre-processing process of the original gas equipment operating data and the specific data types contained after pre-processing. The following is a detailed explanation of this step:

[0082] Raw data collection:

[0083] First, collect raw operating data from gas equipment, which is usually unprocessed data directly recorded by sensors or other monitoring devices.

[0084] The collected raw data often contains noise, outliers, missing values, or inconsistent formats, which may affect subsequent data analysis and model training. Therefore, the raw data needs to be preprocessed to improve the quality and usability of the data.

[0085] Preprocessing operations include data cleaning (such as removing noise and filling missing values), data transformation (such as normalization and standardization), data compression (such as dimensionality reduction), feature selection, etc. The purpose of these operations is to make the data more suitable for subsequent machine learning or deep learning models.

[0086] After preprocessing, the operating data of gas equipment mainly includes the following categories:

[0087] Pressure data:

[0088] Gas equipment will generate various pressure data during operation, such as pipeline pressure, tank pressure, etc.

[0089] Temperature data:

[0090] Temperature is another key parameter in the operation of gas equipment, for example, the operating temperature of gas turbines, burners and other components.

[0091] Traffic data:

[0092] Gas flow data reflects the flow of gas through pipelines or equipment. This data helps monitor gas consumption, detect leaks, and evaluate equipment efficiency.

[0093] Leakage data:

[0094] Leakage is a safety issue that requires special attention in the operation of gas equipment. Leakage data may include leakage signals detected by specific sensors, leakage rates, etc.

[0095] In summary, the preprocessed gas equipment operation data is a cleaned, transformed and optimized data set, which mainly contains four key information: pressure, temperature, flow and leakage.

[0096] In a preferred embodiment of the present invention, a convolutional neural network model is used to perform preliminary feature extraction on the preprocessed data to capture local features in the gas equipment operation data, including:

[0097] Determine the basic architecture of the convolutional neural network model, including convolutional layers, pooling layers, and fully connected layers, including: selecting the appropriate convolution kernel size (such as 3×3, 5×5, etc.) and stride; determining the number of convolutional layers and the number of feature maps output by each layer; after each convolutional layer, select an activation function (such as ReLU) to increase the nonlinearity of the model; after the convolutional layer, add a pooling layer (such as maximum pooling, average pooling) to reduce the dimension of the feature map and enhance the translation invariance of the features, and set the window size and stride of the pooling layer; after multiple convolutional and pooling layers, use a fully connected layer to integrate features and prepare for the final classification or regression task, and determine the number of fully connected layers and the number of neurons in each layer. Design a suitable output layer according to the task type (classification, regression, etc.), such as using the softmax function for multi-classification tasks.

[0098] Define the parameters of the genetic algorithm, including population size, crossover rate, mutation rate, and evolutionary generations. Specifically, set an appropriate population size, that is, the number of individuals in each generation, according to the complexity of the problem and computing resources; the crossover rate determines the probability of a crossover operation between two individuals in the genetic algorithm. The crossover operation helps to combine the excellent genes of different individuals to generate new possible solutions. Set a moderate crossover rate to balance the ability of exploration and development; the mutation rate determines the probability of random mutation of individual genes in the genetic algorithm. The mutation operation helps to introduce new genetic mutations, increase the diversity of the population, and prevent premature convergence. Set a lower mutation rate to avoid destroying the existing excellent gene structure; the evolutionary generation represents the number of iterations of the genetic algorithm. Set a sufficiently large evolutionary generation to ensure that the algorithm has enough time to search for excellent solutions. At the same time, computing resources and time costs should also be considered.

[0099] The hyperparameters of the convolutional neural network model are encoded as genes of the genetic algorithm to form an initial population. The hyperparameters include learning rate, batch size and number of convolution kernels. Specifically, the hyperparameters that have a significant impact on the performance of the CNN model are selected for encoding, such as learning rate, batch size and number of convolution kernels; for each hyperparameter, its value range and encoding method are determined. For example, the learning rate can take a continuous value in the range of [0.0001, 0.1], which can be encoded as a real gene; the batch size and number of convolution kernels can be discrete integer values, which can be encoded as integer genes; the initial population is randomly generated according to the population size and gene encoding method. Each individual represents a set of hyperparameter settings for the CNN model, ensuring the diversity of the initial population to cover different hyperparameter combinations and increase the possibility of searching for the global optimal solution.

[0100] The preprocessed data is divided into training set, validation set and test set. For each generation of population, the following steps are performed:

[0101] Decode each individual in the population into the hyperparameter settings of the convolutional neural network model, specifically including: for each individual in the population, decode it into the corresponding hyperparameter value according to its gene encoding method. For example, if the learning rate is encoded as a real number gene, directly use the gene value as the learning rate; if the batch size or the number of convolution kernels is encoded as an integer gene, decode it into the corresponding integer value; set the decoded hyperparameter value as the training parameter of the CNN model, including the learning rate, batch size, and the number of convolution kernels.

[0102] The convolutional neural network model is trained using the training set and the current individual hyperparameter settings. In each training cycle, forward propagation, loss calculation, backpropagation, and parameter update are performed, including: loading data from the training set and dividing the data into multiple small batches according to the batch size; initializing the network structure and parameters according to the basic architecture of the CNN model; for each training cycle (epoch), the following operations are performed:

[0103] Forward propagation: A small batch of data is input into the CNN model, and forward calculation is performed through the convolutional layer, pooling layer, and fully connected layer to obtain the output of the model.

[0104] Loss calculation: Calculate the loss function value based on the model's output and the true label.

[0105] Back propagation: Perform back propagation based on the loss function value, calculate the gradient and update the model parameters.

[0106] Parameter Update: Use an optimization algorithm (such as SGD) and a learning rate to update the parameters of the model.

[0107] After each training cycle, the validation set is used to evaluate the performance of the convolutional neural network model and record the accuracy, which includes: loading data from the validation set; inputting the validation set data into the trained CNN model to obtain the model's prediction results. Then, performance indicators such as accuracy are calculated based on the prediction results and the true labels, and the calculated accuracy is recorded for subsequent selection of excellent individuals.

[0108] According to the accuracy on the validation set, select the corresponding individuals to enter the next generation population, randomly select two individuals in the population, exchange some of their genes, generate new individuals, and replace some individuals in the original population with new individuals to form the next generation population, until the preset evolutionary generation is reached, and the genetic algorithm cycle is stopped, specifically including: according to the accuracy on the validation set, select individuals with better performance to enter the next generation population. You can select the top N individuals with the highest accuracy, or use selection strategies such as roulette. Randomly select two individuals in the population, exchange some of their genes at a certain crossover rate to generate new individuals. The crossover operation can use single-point crossover, multi-point crossover or uniform crossover methods. Perform gene mutation on the newly generated individuals at a certain mutation rate to increase the diversity of the population. Mutation operations can include bit reversal, bit exchange or real number mutation. Replace some individuals in the original population with the newly generated individuals to form the next generation population; after the genetic algorithm reaches the preset evolutionary generation, stop the algorithm cycle, and select the best performing individual from the last generation population as the final hyperparameter setting.

[0109] Select the corresponding individuals from the last generation population as the final hyperparameters, and use the final hyperparameter settings to retrain the convolutional neural network model to obtain the trained convolutional neural network model, specifically including: setting the finally selected hyperparameters as the training parameters of the CNN model; using all the training set data, retraining the CNN model according to the training process in step 2 until the preset number of training cycles is reached or other stopping conditions are met.

[0110] The convolution layer of the trained convolutional neural network model is used as a feature extractor, and the input data is forward propagated through the convolution layer to obtain the output of the convolution layer as the extracted local features. Specifically, the convolution layer of the trained CNN model is used as a feature extractor, and for new input data, only forward propagation is performed through the convolution layer to obtain the output of the convolution layer as the extracted local features.

[0111] In an embodiment of the present invention, the hyperparameters (such as learning rate, batch size and number of convolution kernels) of the convolutional neural network model are automatically searched and optimized by a genetic algorithm, thereby avoiding the tedious process of manual parameter adjustment, and at the same time, it is more likely to find a global optimal or near-global optimal hyperparameter combination, thereby improving the performance and generalization ability of the model. The convolutional neural network has a strong feature learning ability and can automatically extract useful local features from the original data. The convolution layer of the CNN model optimized by the genetic algorithm can more effectively capture the key local features in the gas equipment operation data, and provide high-quality feature input for subsequent tasks (such as risk prediction). The genetic algorithm continuously evolves more adaptable individuals in the population by simulating natural selection and genetic mechanisms. This evolutionary process helps the CNN model maintain high robustness and stability when facing complex and changeable gas equipment operation data. The optimized CNN model can more accurately capture the potential patterns in the data, thereby improving the accuracy of risk prediction. Compared with traditional hyperparameter optimization methods such as trial and error or grid search, the genetic algorithm combined with the CNN method can find a better hyperparameter combination in a shorter time, thereby reducing the consumption of computing resources and time cost.

[0112] In one embodiment of the present invention, the accuracy The calculation formula is:

[0113] ;

[0114] in, Yes Category The true positives, i.e., the categories The number of samples correctly predicted as positive; Yes Category The true negative of The number of samples correctly predicted as negative class; Yes Category The false positive of the category The number of negative samples that are incorrectly predicted as positive; Yes Category The false negatives of The number of positive samples that are incorrectly predicted as negative; is the total number of categories; is the mean of the Gaussian distribution, indicating a certain central location; is the standard deviation; Yes Category The number of samples; is the index value of the category.

[0115] In the embodiment of the present invention, the formula not only considers the prediction accuracy of the classification model for different categories (represented by TP and TN), but also considers the performance of the model in avoiding wrong predictions (represented by FP and FN). Therefore, it can more comprehensively evaluate the performance of the model. The accuracy of each category is weighted so that the category with more samples has a relatively smaller impact on the overall accuracy, while the category with fewer samples has a relatively larger impact. This helps to balance the deviation caused by the difference in the number of samples of different categories, making the evaluation result more fair. The Gaussian distribution parameter in the formula is and It can be adjusted according to actual conditions to adapt to different data sets and classification tasks. This flexibility makes the accuracy calculation formula have a wider range of application scenarios. By calculating the true positive, true negative, false positive and false negative for each category separately, and combining the Gaussian distribution weights, we can gain a deeper understanding of the performance of the model in each category. This helps to find out which categories the model performs well and which categories need improvement, thereby providing targeted guidance for subsequent model optimization.

[0116] In one embodiment of the present invention, deep features are extracted through a deep residual network based on local features; the deep features are combined with device entities, attributes, relationships and timestamp information to construct a preliminary knowledge graph, including:

[0117] Use local features to train a deep residual network, optimize network parameters through back propagation, and obtain a trained deep residual network. Specifically, it includes: collecting a data set containing local features and performing necessary preprocessing, such as normalization and denoising; designing the structure of the residual block, which includes two or more convolutional layers and a skip connection to achieve residual learning. Determine the depth of the network, that is, the number of residual blocks, and the convolutional layer settings within each residual block, and set the input and output feature map size, number of channels and other parameters.

[0118] Build the model framework:

[0119] Start building a model using a deep learning framework (such as TensorFlow, PyTorch, etc.). Create a class that inherits from the model base class provided by the framework (such as torch.nn.Module in PyTorch). In the constructor of the model class, add convolutional layers (Conv2d) in sequence according to the designed network structure, and set parameters such as convolution kernel size, step size, and padding. Add a batch normalization layer (BatchNorm2d) after each convolutional layer to help the model converge faster and improve generalization ability. Select a suitable activation function (such as ReLU) and add it after the batch normalization layer. Implement residual connections to jump the input directly to the output of the residual block to ensure dimension matching. Define the forward method in the model class, which describes the forward propagation process of data in the network. Call the forward propagation function of each layer in sequence according to the order in which the network layers are added, and ensure the correct implementation of the residual connection. The framework will automatically initialize the parameters in the network layer (such as convolution kernel weights, biases, etc.); set the optimizer (such as Adam, SGD, etc.) and the learning rate. Define the loss function for error calculation during training.

[0120] The local feature data is input into the deep residual network, and the output of the network is obtained after calculations in each layer; the loss function value is calculated based on the network output and the true label. Commonly used loss functions include cross entropy loss, mean square error loss, etc.; the loss function value is used for back propagation, the gradient is calculated and the network parameters are updated to minimize the loss function; the process of forward propagation, loss calculation and back propagation is repeated until the preset number of training rounds is reached or other stopping conditions are met; after the training is completed, the trained deep residual network model and its parameters are saved.

[0121] Apply the trained deep residual network to the data of all local features to extract deep features, including: Load the previously trained deep residual network model and its parameters using the functions or methods provided by the deep learning framework. This usually involves specifying the path and name of the model file. For example, in PyTorch, you can use the torch.load() function to load the model state dictionary (state_dict). Make sure that all local feature data has been preprocessed and prepared for input to the network. Normalize, resize, or convert the data to the appropriate tensor format according to the input requirements of the model. Pass the prepared local feature data as input to the loaded deep residual network model. Execute the forward propagation method of the model (usually the forward method), which will calculate through the layers of the network. During the forward propagation process, select one or more intermediate layers of the network and obtain the output from these layers as deep features. These intermediate layers should be able to capture higher-level and abstract information of the input data. Save the deep features obtained from the selected intermediate layers, which can be stored in variables in memory, written to files or databases.

[0122] Integrate the deep features with the corresponding device entity information, attributes, relationships and timestamp information to make the feature data correspond to the metadata one by one to obtain the integrated data, including:

[0123] The deep features extracted from the deep residual network are kept consistent with the corresponding device entity information, attributes, relationships, and timestamp information in terms of data dimension and order. Data alignment can be achieved through index matching, data sorting, or using data loaders provided by the framework (such as PyTorch's DataLoader).

[0124] Concatenate deep features with metadata such as device entity information, attributes, and relationships. This can be achieved by expanding on the feature dimension, such as using np.concatenate (NumPy) or torch.cat (PyTorch), so that the numerical range and normalization of the features are taken into account when concatenating to avoid some features overwhelming other features numerically. The concatenated feature vector or data record should contain more comprehensive information that can reflect multiple aspects of the device entity.

[0125] Associate timestamp information with the corresponding device entities and relationships. This can be achieved by adding timestamps as additional feature dimensions or attributes to data records so that the timestamp format is uniform and takes into account factors such as time zone, time unit, etc. If the timestamp is a continuous numeric value (such as Unix timestamp), consider converting it to more meaningful features such as time of day, day of the week, etc.

[0126] Combine all the integrated data records into a dataset. This can be done by creating a list, array, or using a data structure provided by the framework (such as PyTorch's TensorDataset). So that each record in the dataset contains a complete feature vector and related timestamp information. Divide the dataset appropriately, such as training set, validation set, and test set, so that the knowledge graph model can be built and evaluated later. Save the constructed dataset to disk for subsequent use. You can choose to save it as a CSV file, a NumPy array file (such as .npy), or a framework-specific format (such as PyTorch's .pt file), and back up the dataset regularly to prevent data loss or corruption.

[0127] Based on the integrated data, a preliminary equipment knowledge graph is constructed, in which nodes represent equipment entities, edges represent relationships between entities, attribute information is attached to corresponding nodes, and timestamp information is used to describe the dynamic changes of entities and relationships, including:

[0128] For each device entity in the dataset, create a node. This can be achieved through a graph database or graph processing framework (such as Neo4j, JanusGraph, NetworkX, etc.); each node should contain the entity's unique identifier (such as ID, serial number, etc.), name, and other core attribute information, such as device type, manufacturer, etc. According to the relationship information in the integrated data, establish edges between device entity nodes. The type of edge should reflect the actual relationship between entities, such as "connected", "belongs to", "controls", etc. The attributes of the edge can include the strength, weight, direction, etc. of the relationship, which are defined according to the application scenario. Attach other attribute information of the device entity (such as location, status, configuration parameters, etc.) to the corresponding node. These attributes can be stored directly as part of the node, or exist as associated attribute nodes, depending on the design of the graph database and the complexity of the data model. For each node and edge, add a time attribute to describe its life cycle or effective time range according to its related timestamp information. If the dataset contains time series data, you can create time series nodes or attributes to represent the evolution of entity states. For example, you can create a time series node to record the monitoring data of the device such as temperature and humidity over a certain period of time. For dynamic changes in edges, timestamps can be used to mark the start and end time of the relationship, thereby indicating the timeliness of the relationship between entities. After building the knowledge graph, verify the graph structure to ensure the correctness, completeness, and consistency of the nodes and edges. Optimize the graph according to actual needs, such as merging duplicate nodes and deleting invalid edges. Create appropriate indexes for the nodes and edges in the graph according to common query patterns to improve query efficiency, and update the index regularly to reflect the latest status of the graph.

[0129] In an embodiment of the present invention, the deep residual network has a powerful feature learning ability and can effectively extract deeper features from local features. These deep features contain richer semantic information, which helps to improve the accuracy and completeness of the knowledge graph. By integrating the deep features with device entity information, attributes, relationships and timestamp information, a more comprehensive data representation can be formed. This integration not only enriches the dimensions of the data, but also establishes a one-to-one correspondence between feature data and metadata, which facilitates subsequent data analysis and knowledge reasoning. The constructed device knowledge graph displays the relationship and attribute information between device entities in an intuitive way. This graphical representation method is not only easy for humans to understand, but also helps to discover patterns and laws hidden in the data. The introduction of timestamp information enables the knowledge graph to describe the dynamic changes of entities and relationships. The method based on the deep residual network and the knowledge graph has good scalability and flexibility. With the continuous accumulation of data and the continuous optimization of the model, the accuracy and coverage of the knowledge graph can be further improved.

[0130] In one embodiment of the present invention, a conditional random field model is used to jointly annotate entities and relationships in a preliminary knowledge graph, a graph convolutional network is used to optimize the preliminary knowledge graph, and nodes are updated iteratively, including:

[0131] Extract features for the conditional random field model from the preliminary knowledge graph, including part of speech, context information, and entity type, including:

[0132] According to the task requirements, a series of feature functions are defined to capture the intrinsic relationship between the observation sequence (such as part of speech, context, entity type) and the state sequence (such as entity label, relationship label). The feature function can include: the part of speech of the current word, the part of speech of the previous and next words, whether the current word is part of the entity, whether the previous and next words are entities, entity type, etc. Define all possible entity and relationship labels, such as entity labels such as "name", "place name", "institution name", and possible relationship labels, and create a label set for the model to select the most likely label during training. Use the defined feature function set and label set to build a CRF model. The CRF model is essentially a probabilistic undirected graph model, in which nodes represent random variables and edges represent dependencies between random variables. A weight parameter is assigned to each feature function, which will be adjusted during the training process. The weight parameter can be initialized randomly or with some prior knowledge to set the initial value. The features such as part of speech, context information, and entity type extracted from the preliminary knowledge graph are integrated into a training data set. Each training sample should contain an observation sequence (feature) and a corresponding state sequence (label). Choose a training algorithm suitable for the CRF model, such as L-BFGS, SGD (stochastic gradient descent) and other optimization algorithms, which are used to adjust the model parameters during the training process to minimize the loss function.

[0133] The CRF model is trained using training data in an iterative manner. In each iteration, the conditional probability distribution of each training sample is calculated based on the current model parameters, and the model parameters are updated using the selected optimization algorithm to maximize the joint probability of the observation sequence and the corresponding state sequence. During the training process, the validation set is used regularly to evaluate the performance of the model, and the model parameters, such as learning rate, regularization strength, etc., are adjusted based on the evaluation results. If the performance reaches the preset standard or the number of iterations reaches the upper limit, the training is stopped. After the training is completed, the parameters of the CRF model are saved for subsequent use. The model parameters can be saved as files or database records so that they can be loaded and applied when needed. Among them, in the conditional random field (CRF) model, the feature function is used to capture the intrinsic relationship between the observation sequence and the state sequence. The feature function is defined as a binary function whose input includes a part of the observation sequence (such as the current word, previous and next words, etc.) and a part of the state sequence (such as the current label, previous and next labels, etc.), and the output is a real value (usually 0 or 1, indicating whether the feature exists). The following are some specific feature function formulas:

[0134] Part of speech features of the current word:

[0135] If the part of speech of a is a certain part of speech (such as a noun) and s is a certain label (such as an entity label), then 1; otherwise 0; among them, represents the current word in the observation sequence, Represents the current label in the state sequence.

[0136] Part of speech characteristics of the preceding and following words:

[0137] if , and The part of speech satisfies a certain pattern and is a specific tag, then 1, otherwise 0; among them, , , Respectively represent the previous word, current word and next word in the observation sequence, Represents the current label in the state sequence.

[0138] Whether the current word is part of an entity:

[0139] if Is Entity part of ;otherwise ;in, Represents an entity. This feature function can be calculated on each word in the observation sequence to determine whether it belongs to a certain entity.

[0140] Whether the preceding and following words are entities:

[0141] if , or At least one of them is an entity part of 1; otherwise 0.

[0142] Entity Type Characteristics:

[0143] if The entity type is (such as "name of a person", "place name", etc.), then ;otherwise ;

[0144] in, Represents a specific entity type. This feature function is used to capture the entity type information of words in the observation sequence.

[0145] Use the trained conditional random field model to jointly annotate entities and relationships on the preliminary knowledge graph, and identify the relationships between entities in the preliminary knowledge graph. Specifically, use the corresponding library (such as CRF++, sklearn-crfsuite, etc.) to load the trained conditional random field model, extract features such as part of speech, context information, and entity type from the preliminary knowledge graph, and these features will be used as the input of the conditional random field model. The extracted features are sorted in the format required by the model, which requires converting the features into a specific data structure, such as a feature vector or matrix. Input the formatted feature data into the loaded conditional random field model. The conditional random field model will use the input features to predict the most likely annotation of each word, that is, the entity type (such as a person's name, a place name, etc.) or the relationship type (such as belonging to, located in, etc.), and obtain the prediction results from the model. These results are usually a sequence of annotations that correspond one to one with the words in the input data. According to the annotation results, identify the entities in the preliminary knowledge graph and the relationships between them. For example, consecutive words with the same entity type annotation can form an entity, while specific types of relationship annotations indicate the relationship between entities. Traverse the annotation results and extract all entities and their relationships. This can be achieved through rule matching or more complex parsing algorithms.

[0146] The knowledge graph annotated by the conditional random field model is converted into graph structured data, where nodes represent entities and edges represent relationships. A feature matrix is ​​constructed for each node. The feature matrix includes the attributes of the node and the information of adjacent nodes, including: Traversing the results of the CRF model annotation, identifying all entities; treating each identified entity as a node in the graph; and treating the relationships between entities as edges in the graph based on the relationships annotated by the CRF model. For example, if CRF marks "A is the father of B", an edge from node A to node B is created in the graph, and the label of the edge is "father"; ensure that each node and edge in the graph has a unique identifier for subsequent processing. Construct a feature matrix for each node:

[0147] Initialize an empty feature matrix, whose size is determined by the number and type of required features; for each node (entity) in the graph, collect its attribute information, such as entity type (person name, place name, organization name, etc.), text description, etc., and encode this information as part of the feature matrix. For example, one-hot encoding can be used to represent entity types; collect node information adjacent to the current node. This includes the type of adjacent nodes, the type of relationship with the current node, etc., which can also be encoded as part of the feature matrix; if the number of adjacent nodes is not fixed, consider setting a maximum number of adjacent nodes and filling the insufficient part (such as filling with 0); merge the node's own attributes and adjacent node information into the feature matrix. Standardization and processing of feature matrix:

[0148] The feature matrix is ​​standardized to ensure the consistency of dimensions and ranges between different features. If needed, the feature matrix can be further processed, such as dimensionality reduction, denoising, etc. The feature matrix of each node is saved as a file or database record for use in subsequent models such as graph convolutional networks (GCN).

[0149] Use graph convolutional networks to optimize the graph, and use multi-layer graph convolution operations to capture the deep connections and features between nodes. Specifically, convert the previously constructed graph structure data into a format suitable for GCN processing, which is an adjacency matrix or adjacency list, representing the connection relationship between nodes. Arrange the feature matrix of each node into a two-dimensional tensor, where each row corresponds to the feature vector of a node. Construct a graph convolutional network model:

[0150] Define the model structure of GCN, including the number of graph convolution layers, the number of neurons in each layer, activation function, etc. Initialize the parameters of the GCN model, which will be updated during the training process. Feed the feature matrix and adjacency matrix of the node as input to the first layer of GCN. In each layer of graph convolution, perform neighborhood information aggregation operation on each node. This is usually achieved by weighted averaging (or other forms of aggregation) the feature vector of the node with the feature vectors of its neighboring nodes. The aggregated feature vector undergoes a linear transformation (defined by the weight matrix and bias term of the layer) and an activation function (such as ReLU) to obtain the output feature representation of the node in the current layer. Use the output of the current layer as the input of the next layer and repeat the graph convolution operation until the preset number of graph convolution layers is reached. After each layer of graph convolution, the feature representation of the node is updated so that it can perceive more distant context information. Through the superposition of multiple layers of graph convolution, the deep associations and features between nodes are gradually captured. After the last layer of graph convolution, the final feature representation of each node is obtained. These feature representations can be used for subsequent tasks such as node classification, link prediction, etc. Define the loss function according to the specific task, and use optimization algorithms such as gradient descent to update and optimize the parameters of the GCN model.

[0151] According to the deep connections and features between nodes, the entities and relationships in the graph are adjusted and optimized to obtain the optimization results; based on the optimization results, the state and attributes of each node in the graph are updated, including: obtaining the feature representation of each node output by the graph convolution network; parsing these feature representations to understand the meaning of each feature dimension and how they are captured through the deep connection of the graph convolution operation. Identify and correct incorrect entity type or relationship annotations:

[0152] Using the feature representation of the nodes, possible entity type errors can be identified through clustering, classification, or other machine learning methods. For example, the features of some nodes may be closer to another entity type than the currently labeled type. For relationship annotation, possible relationship annotation errors can be found by analyzing the feature similarities and differences between node pairs and comparing with other node pairs. Correct the identified errors manually or using automated tools.

[0153] Optimize graph structure:

[0154] Based on the feature representation of the nodes and the strength of the association between them, add or delete some edges to optimize the structure of the graph. For example, if the feature representations between two nodes are very similar and closely related, but they are not directly connected in the graph, you can consider adding an edge. On the contrary, if the association between two nodes is weak or their feature representations are very different, but they are connected by an edge in the graph, you can consider deleting this edge. Update the state and attributes of each node in the graph:

[0155] Based on the revised entity type and relationship annotations, update the attribute information of each node, such as entity type, text description, etc. If some attributes of the node can be inferred or predicted through its feature representation (such as through a machine learning model), then update these attributes to ensure that the status and attributes of all nodes are up-to-date and accurate.

[0156] Verify and evaluate optimization results:

[0157] Use some quantitative indicators (such as the accuracy of entity type classification, the F1 score of relationship extraction, etc.) to evaluate the quality of the optimized graph. You can also verify the effectiveness of the optimization results through expert review or comparison with other reliable data sources, and persist the optimized graph structure, node status, attributes and other information to the database or file system.

[0158] The graph convolutional neural network is repeatedly used for multiple iterations. Each iteration adjusts the graph according to the result of the previous iteration until the preset number of iterations is reached. After the iteration, the graph is finally optimized and adjusted, including: setting a preset number of iterations N, which will serve as the termination condition of the iteration process, preparing the initial graph data, including the graph structure and node feature matrix; for the i-th iteration (i starts from 1 until the preset number of iterations N is reached), perform the following steps:

[0159] The current graph data and node feature matrix are used as input to the graph convolutional neural network. Through the processing of the graph convolutional neural network, the updated feature representation of each node is obtained; according to the updated feature representation of the node, the graph is adjusted, including correcting the entity type, relationship annotation, adding and deleting edges to optimize the graph structure, etc. The state and attributes of each node in the graph are updated to ensure that they reflect the latest adjustment results. Check the iteration termination condition:

[0160] If i is less than the preset number of iterations N, the iteration continues; if i is equal to the preset number of iterations N, the iteration stops. Final optimization after the iteration ends:

[0161] Perform a final check and refinement of the map, which includes:

[0162] Delete isolated nodes (i.e., nodes that are not connected to other nodes), identify and delete redundant edges (for example, there may be multiple edges of the same type between two nodes), perform a rationality check on the overall structure of the graph to ensure that there are no logical errors or inconsistencies, and save the knowledge graph that has been optimized and improved many times as a file or database record.

[0163] In an embodiment of the present invention, the conditional random field model can comprehensively consider features such as part of speech, context information and entity type, and jointly annotate entities and relationships in the preliminary knowledge graph. Compared with the traditional independent annotation method, this method can better capture the dependency between entities and relationships, thereby improving the accuracy of annotation. By optimizing the graph through a graph convolution network, the deep associations and features between nodes can be captured. The graph convolution operation can aggregate the neighborhood information of nodes layer by layer, so that each node can perceive farther context information, thereby more accurately representing the complex relationship between entities and relationships. In the iterative process, the graph is adjusted according to the results of the previous time, and the graph structure can be dynamically optimized. This method can gradually correct the errors and incomplete information in the preliminary knowledge graph, making the graph more accurate and rich. Through multiple iterations and the final optimization adjustment, the quality of the knowledge graph can be significantly improved. The optimized graph can more accurately reflect the relationship between entities. The method has good scalability and can adapt to knowledge graph optimization tasks of different scales and complexities. With the increase in data volume and the continuous improvement of the model, this method is expected to achieve better optimization results in the future.

[0164] In one embodiment of the present invention, the long short-term memory network uses the time series data in the preliminary knowledge graph as sequence input data, captures the law of changes in the state of the gas equipment over time by learning the long-term dependencies in the sequence, and generates key features containing time series information, including:

[0165] Extract time-series related data from the preliminary knowledge graph, including device status records and sensor readings arranged in chronological order, specifically including: accessing the database or data storage system of the preliminary knowledge graph, using a query language (such as Cypher, SPARQL or SQL) to retrieve time-series related data, including device status records and sensor readings, ensuring that the retrieved data contains timestamp information so that it can be subsequently arranged in chronological order, and exporting the retrieved data into a format suitable for subsequent processing, such as CSV or JSON.

[0166] Read the time series-related data and set a corresponding time window size; split the time series data according to the set time window, form a sequence for the data in each window, and standardize each sequence to obtain the processed time series data sequence, including: use programming languages ​​(such as Python) and related libraries (such as pandas) to read the exported time series data, sort the data by timestamp to ensure the correctness of the time series, and set a time window size, for example, every hour, day, or week; split the data into multiple sequences according to the time window size, each sequence contains the data in a time window; standardize each sequence, such as Z-score standardization, to eliminate the dimensional differences between different features.

[0167] Determine the number of LSTM network layers, set the number of hidden units in each LSTM network layer, and determine the input and output dimensions of the network. The input dimension should match the dimension of the processed time series data sequence. Specifically, the number of LSTM network layers should be determined based on the complexity of the problem and the size of the data. You can start with one layer and increase or decrease it based on performance. Set the number of hidden units in each LSTM layer. This number can be determined based on experience or experiments. Determine the input dimension of the network to match the dimension of the processed time series data sequence. For example, if each time series data point contains 10 features, the input dimension should be 10. Determine the output dimension based on the task requirements. If it is a prediction task, the output dimension may be the same as the input dimension. If it is a classification task, the output dimension may correspond to the number of categories.

[0168] Use a deep learning framework to create an LSTM model, convert the processed time series data sequence into the tensor format required by the LSTM model, and divide the processed time series data sequence into training set, validation set, and test set. Specifically, it includes: Use a deep learning framework (such as TensorFlow or PyTorch) to create an LSTM model, convert the processed time series data sequence into the tensor format required by the model, which involves converting the data into a floating-point array and ensuring that the shape of the data matches the model input requirements. Divide the data set into training set, validation set, and test set. Random partitioning or chronological partitioning can be used, depending on task requirements and data characteristics.

[0169] The LSTM model is trained using the training set data. After each round of training, the performance of the LSTM model is evaluated using the validation set. The learning rate and optimizer parameters are adjusted according to the performance of the validation set. The weights and biases of the LSTM network are updated using the back-propagation algorithm to obtain the trained LSTM model, which includes: training the LSTM model using the training set data, which involves specifying the loss function (such as mean square error or cross entropy loss) and the optimizer (such as Adam or RMSprop). After each round of training, the performance of the model is evaluated using the validation set. The model parameters and learning rate can be adjusted based on indicators such as loss and accuracy on the validation set. The learning rate and optimizer parameters are adjusted according to the performance of the validation set, which can be achieved by manual adjustment or by using automatic parameter adjustment techniques (such as grid search or random search). The weights and biases of the LSTM network are updated using the back-propagation algorithm, which is accomplished by calculating the gradient of the loss function with respect to the model parameters and updating the parameters accordingly.

[0170] The processed time series data sequence is input into the trained LSTM model, and the trained LSTM model outputs the hidden state of each time step. The hidden state contains the long-term dependencies and key features in the time series data. Specifically, the processed time series data sequence is input into the trained LSTM model, and the model outputs the hidden state of each time step. These hidden states contain the long-term dependencies and key features in the time series data. These hidden states can be further processed according to task requirements, such as for predicting future equipment status, detecting anomalies, or performing classification tasks.

[0171] In an embodiment of the present invention, it is possible to ensure that the model focuses on data related to time series changes, thereby more effectively capturing the law of changes in device status over time. By setting the time window size and splitting the data, the model can process data in a continuous time period, which helps to capture state changes and trends within a specific time range. At the same time, standardization can eliminate dimensional differences in the data, making it easier for the model to learn the inherent laws in the data. Determining the number of LSTM layers and the number of hidden units can affect the learning ability and complexity of the model, while determining the input and output dimensions ensures that the model can correctly process time series data. The LSTM model can be efficiently constructed and trained using a deep learning framework. By training the model with a training set, the model can learn long-term dependencies in time series data. The evaluation of the validation set helps to adjust the learning rate and optimizer parameters in a timely manner to prevent the model from overfitting or underfitting, thereby improving the prediction performance of the model. The trained LSTM model can process new time series data and output the hidden state of each time step. These hidden states contain long-term dependencies and key features in the time series data, which are crucial for subsequent tasks such as device state prediction and anomaly detection. By capturing these key features, the future state of the equipment can be predicted more accurately, allowing for timely maintenance and intervention to improve the operating efficiency and safety of the equipment.

[0172] In one embodiment of the present invention, the temporal feature representation output by the long short-term memory network is integrated with the device attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive state of the gas equipment; the feature vector is comprehensively analyzed to obtain hidden patterns related to the risk of the gas equipment, including:

[0173] Obtain the temporal feature representation output by the long short-term memory network, and extract the static attributes and relational information related to the gas equipment from the knowledge graph, where the static attributes include the equipment model, manufacturer, and installation date; the relational information includes the connection relationship and maintenance record with the equipment, specifically including: loading the trained long short-term memory network (LSTM) model; inputting the preprocessed time series data (such as equipment status records, sensor readings, etc.) into the LSTM model; running the LSTM model to obtain the temporal feature representation output by the model. These feature representations capture the long-term dependencies and dynamic behavior patterns in the data; saving or caching the temporal feature representation output by the LSTM; connecting to the database or query interface of the knowledge graph; using an appropriate query language (such as SPARQL, Cypher, etc.) to write query statements to extract static attributes related to the gas equipment, such as equipment model, manufacturer, and installation date; similarly, write query statements to extract relational information related to the equipment, such as the connection relationship and maintenance record of the equipment; executing the query statements, and saving the extracted static attributes and relational information as structured data (such as CSV, JSON, etc.).

[0174] The temporal features output by the long short-term memory network are concatenated with the static features and relational features extracted from the knowledge graph to form a fused feature vector, which includes:

[0175] Read the time series feature data output by the LSTM model from a saved file or cache. This data is a two-dimensional array (or tensor) in which each row represents a feature vector for a time step, so that the dimension of the time series features is compatible with subsequent processing. Read data from a file containing static attributes and relationship information previously extracted from the knowledge graph. This data can be in CSV, JSON, or other structured formats. Static attributes include device model, manufacturer, installation date, etc., while relationship information covers connection relationships between devices, maintenance records, etc. Convert this data into a format suitable for subsequent processing, such as a two-dimensional array or data frame.

[0176] Match the data points between time series features, static attributes, and relationship information one by one, which usually involves matching based on device ID or other unique identifiers. If there are unmatched data points, appropriate processing is required, such as discarding, interpolating, or filling missing values; preprocess static attributes and relationship information, such as encoding categorical variables, normalizing numerical variables, etc., to ensure that they are compatible with the time series features in terms of numerical range; choose concatenation on the feature dimension as the splicing method, which means directly connecting the vectors of time series features, static attributes, and relationship features end to end; for each device or data point, connect its corresponding time series feature vector, static attribute vector, and relationship feature vector according to the determined splicing method, which can be achieved through array operations in the programming language or related library functions (such as NumPy's concatenate function) to ensure that the spliced ​​feature vectors have consistent dimensions and formats.

[0177] Apply K-means to cluster the fused feature vectors to find groups with similar device status and behavior patterns, including: reading the fused feature vector data from a previously saved file or database to ensure that the data format is correct and the dimensions of each feature vector are consistent; determining the K value, which represents the expected number of clusters, that is, the number of groups to be divided into. A suitable K value can be determined by prior knowledge, business needs, or data exploration, or some methods (such as the Elbow Method) can be used to assist in determining the optimal K value; selecting a distance metric, such as Euclidean distance; using Python's scikit-learn library, which provides an implementation of K-means clustering, creating a KMeans object, setting the selected K value and distance metric (if the library supports custom distance metrics), calling the fit method, clustering the fused feature vector as input data, and using the labels_ attribute to obtain the group label to which each data point belongs. These labels are an integer array representing the group number to which each data point is assigned.

[0178] The isolation forest algorithm is used to identify feature vectors that are different from the normal pattern. Different feature vectors represent potential risks. Specifically, it includes: loading the fused feature vectors, training the isolation forest model, which can be done by using an anomaly detection library (such as IsolationForest in scikit-learn), inputting the fused feature vectors into the trained isolation forest model to obtain the anomaly score of each data point, setting a threshold, and marking the data points with anomaly scores higher than the threshold as potential risk points.

[0179] Based on the static attributes and relationship information related to gas equipment, groups with similar equipment status and behavior patterns, and potential risks, hidden patterns related to gas equipment risks are determined, including: summarizing and analyzing the discovered equipment group information with similar status, identified potential risk point information, and static attributes and relationship information extracted from the knowledge graph; through comparative analysis, exploring the differences and commonalities between different groups and between potential risk points and normal points; combining domain knowledge and expert judgment to identify hidden patterns related to gas equipment risks, such as common failure modes of specific models of equipment, manufacturer-related quality issues, or risk trends in specific installation environments.

[0180] In an embodiment of the present invention, by fusing the time series features output by LSTM with the static attributes and relationship information in the knowledge graph, the state of the gas equipment can be described more comprehensively. The time series features capture the dynamic behavior of the equipment, while the static attributes and relationship information provide the background and basic information of the equipment. This comprehensive description helps to more accurately evaluate the comprehensive state of the equipment. By performing cluster analysis (such as K-means) by fusing features, groups of equipment with similar states and behavior patterns can be identified. This helps to discover potential common problems or risks, so that preventive measures can be taken in advance to reduce the possibility of failures. The application of anomaly detection technologies such as the isolation forest algorithm can accurately identify feature vectors that are significantly different from normal patterns, which may represent potential risks. Comprehensive analysis of the fused feature vectors helps to discover hidden patterns related to gas equipment risks, which reveal potential causes of equipment performance degradation, precursors to failures, or risk trends related to specific equipment attributes.

[0181] In one embodiment of the present invention, based on the hidden patterns related to the gas equipment risk, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction result, including:

[0182] Based on historical data, a risk level label is assigned to each hidden pattern feature vector, which specifically includes: exporting historical data from the monitoring system of the gas equipment, including the equipment's operation log, sensor readings, environmental parameters (such as temperature, humidity), etc., integrating information from different data sources to ensure the integrity and consistency of the data, marking known risk events in the historical data, such as equipment failures, safety accidents, etc., and recording the time and conditions of their occurrence; cleaning the data, removing duplicate records, outliers and noise, and filling or interpolating missing data to ensure data continuity; normalizing or standardizing the data to eliminate dimensional differences between different features; analyzing historical data to identify features related to equipment risks, such as the stability of equipment operation, volatility of sensor readings, etc.; extracting hidden pattern feature vectors, which includes using statistical methods, time series analysis or machine learning techniques (such as principal component analysis PCA, autoencoders) to reduce dimensionality or extract key features; determining the classification criteria for risk levels based on the risk events marked in the historical data.

[0183] For example, "low risk" can be defined as stable equipment operation without obvious abnormalities; "medium risk" can be defined as equipment with minor abnormalities that do not affect normal operation; and "high risk" can be defined as equipment with obvious or impending failures. Analyze known risk events in historical data and extract features related to these events. These features include specific sensor reading anomalies, patterns of changes in equipment operating status, etc. Based on the extracted risk event features, formulate a set of rule matching logic. This set of logic should be able to determine the risk level to which the hidden pattern feature vector belongs based on its specific performance. The rule can be a simple threshold judgment, such as a sensor reading exceeding a certain value is judged as high risk; or it can be a complex logical combination, such as a certain risk level is judged when multiple features meet specific conditions at the same time. For each hidden pattern feature vector, apply the formulated rule matching logic to make judgments one by one. According to the matching results, assign the corresponding risk level label to the feature vector. For example, if a feature vector meets the high risk rule, it is marked as "high risk". Use a portion of historical data with known risk levels as a validation set to verify the formulated rules. Check whether the rule matching results are consistent with the actual risk level. If it is found that the rules have misjudgments or omissions, the rules need to be adjusted and optimized to improve the accuracy of the match. As new data is continuously collected, new risk events or characteristic patterns may be discovered. At this time, the rule base needs to be updated to adapt to new situations. Regularly review and evaluate the effectiveness of the rules to ensure that they can continue to accurately assign risk level labels to hidden pattern feature vectors.

[0184] Set the parameters of the SVM model, including the kernel function, penalty coefficient C, and kernel function parameters, including: analyzing the characteristics of the data, such as the linear separability of the data, the dimension of the features, etc.; if the data is linearly separable in the original space, you can choose a linear kernel function; if the data has a nonlinear relationship, you can consider using a polynomial kernel function or a radial basis function (RBF) kernel. The RBF kernel is usually suitable for processing nonlinear relationships in high-dimensional data; initialize a penalty coefficient C value, for example, starting from a small value (such as C=1); use cross-validation (such as k-fold cross-validation) to evaluate the performance of the SVM model under different C values. This involves dividing the data set into k parts, using k-1 parts as training data each time, and the remaining 1 part as validation data, repeating k times to obtain the average performance; adjusting the value of C based on the validation results. If the model is underfitting (i.e., both the training error and the validation error are large), you can try to increase the value of C; if it is overfitting (i.e., the training error is small but the validation error is large), you can try to reduce the value of C, and repeat the above process until a suitable C value is found so that the model can achieve good performance on both the training set and the validation set. For the selected kernel function (such as RBF kernel), its parameters, such as γ (gamma), need to be set. γ defines the range of influence of the data point. A larger γ value means that only close data points will have an impact on the classification decision. Similarly, cross-validation is used to evaluate the performance of the model under different γ values. Initialize a γ value and adjust it based on the validation results. You can try multiple different γ values ​​to find the best parameter combination. When adjusting the parameters, you can also consider using optimization methods such as Grid Search to search for the optimal combination of multiple parameters at the same time. This can help find suitable SVM model parameters more efficiently.

[0185] The SVM model is trained using a hidden pattern feature set with risk level labels, specifically including: dividing the hidden pattern feature set with risk level labels into a training set and a test set (or validation set), using 70%-80% of the data as the training set and the remaining data as the test set; using the data and labels of the training set to train the SVM model. This involves solving an optimization problem to find a decision boundary that can maximize the separation of data points with different risk levels.

[0186] By dividing the training set and validation set, the performance and generalization ability of the SVM model are evaluated to obtain the validation results, including: using the test set to evaluate the performance of the trained SVM model. Calculate the accuracy, recall, F1 score and other indicators of the model on the test set to evaluate the classification performance of the model. In addition to using the test set for evaluation, cross-validation (such as k-fold cross-validation) can also be used to evaluate the generalization ability of the model.

[0187] Adjust the SVM model parameters according to the verification results to obtain the final SVM model, including: adjusting the parameters of the SVM model according to the test set and cross-validation results, including changing the kernel function, adjusting the penalty coefficient C and kernel function parameters, etc.; after each parameter adjustment, the model needs to be retrained and its performance evaluated until a set of parameter combinations that can achieve the best performance is found;

[0188] For new gas equipment data, extracting its hidden pattern feature vector, specifically including: collecting new gas equipment data, which should have similar features and label distributions as the training data; using the same feature extraction method as the training data to extract the hidden pattern feature vector from the new data;

[0189] The hidden pattern feature vector is input into the final SVM model to predict the risk level to obtain the final risk level prediction result, which specifically includes: ensuring that the extracted hidden pattern feature vector is consistent with the format and dimension of the training data, and a preprocessing step is required; the processed hidden pattern feature vector is input into the trained SVM model to predict the risk level. The model will output the risk level label of each data point.

[0190] In an embodiment of the present invention, by training an SVM model to identify hidden patterns related to gas equipment risks, an accurate assessment of equipment risks can be achieved. This data-based prediction method is more objective and reliable than traditional subjective evaluation, and helps to timely discover potential high-risk situations. The SVM model can provide early warnings before risk events actually occur. By monitoring the real-time data of the equipment and extracting hidden pattern features, the model can predict future risk levels, allowing preventive measures to be taken to avoid accidents or reduce their impact. Accurate risk level predictions can help operators allocate resources more reasonably, such as giving priority to inspecting or maintaining equipment predicted to have high risk levels. The risk level prediction results provided by the SVM model can serve as an important input for decision support. Managers can formulate more intelligent strategies based on these prediction results, such as formulating maintenance plans, updating equipment or replacing parts. As a boundary-based classifier, the decision boundary of the support vector machine is usually intuitive and easy to understand. This helps to explain why the model makes a specific prediction, thereby enhancing the user's trust in the model's prediction results. The SVM model performs well in processing high-dimensional data and nonlinear relationships. Gas equipment data may contain a variety of complex hidden patterns and nonlinear relationships, and SVM can effectively capture these features and make accurate classification predictions. By continually collecting new data and retraining the model, the SVM can adapt to changes in the behavior patterns of gas appliances.

[0191] In one embodiment of the present invention, the final risk level prediction result and the original data are remotely transmitted to a cloud server or a user terminal via a wireless network, including:

[0192] Initialize a particle swarm based on the network environment and transmission requirements. Each particle represents a data transmission scheme, including the transmission path, data compression rate, and transmission protocol parameters. Specifically, evaluate the current network environment, including key indicators such as network bandwidth, latency, and packet loss rate. At the same time, clarify the transmission requirements, such as data transmission speed requirements, data integrity requirements, etc. According to the complexity of the problem and the availability of computing resources, set the size of the particle swarm, that is, the number of particles. Randomly assign initial values ​​to each particle, which represent different data transmission schemes. Specifically, it can include:

[0193] Select the network path from the sender to the receiver, which can be a direct path or an indirect path through an intermediate node. Determine the degree of data compression before transmission to balance transmission speed and data quality. Set relevant parameters of the transport layer protocol, such as TCP window size, timeout retransmission time, etc. Save the initialization state of each particle as the initial solution set of the particle swarm.

[0194] Define a fitness function to evaluate the performance of each transmission scheme;

[0195] According to the individual final solution and the global final solution, the speed and position of each particle are updated. Through multiple iterations, a data transmission scheme suitable for the current network environment and transmission requirements is obtained, which may include: determining the number of iterations of the particle swarm optimization algorithm, that is, the maximum number of rounds of the algorithm; assigning an initial speed and position to each particle. In the context of the data transmission scheme, the position represents the parameter setting of the transmission scheme, and the speed represents the adjustment range of these parameters in the search space; in each round of iteration, the following operations are performed:

[0196] Evaluate the performance of each particle according to the fitness function, and update the individual optimal solution (i.e., the best historical position) of each particle; find the particle with the best fitness value among all particles as the global optimal solution; and use the particle swarm optimization update formula to adjust the speed and position of each particle according to the individual optimal solution and the global optimal solution. This usually involves some random factors to facilitate the exploration of the search space.

[0197] After the iteration, the optimized transmission scheme corresponding to the global final solution is selected, which may include: terminating the iteration process when a preset number of iterations is reached or other stopping criteria (such as fitness value convergence) are met; after the iteration, the particle with the best fitness value is selected from all particles, and the transmission scheme represented by the particle is the optimized transmission scheme.

[0198] Using the optimized transmission scheme, the final risk level prediction results and original data are remotely transmitted to the cloud server or user terminal via the wireless network, which may include: configuring the network path, data compression rate and transmission protocol parameters according to the optimized transmission scheme; compressing the original data according to the configured data compression rate, and packaging it into a format suitable for network transmission; using the configured transmission protocol and parameters, the compressed and packaged data and the risk level prediction results are sent to the cloud server or user terminal via the wireless network; during the data transmission process, the transmission status is monitored and necessary log information is recorded for subsequent analysis and troubleshooting.

[0199] In an embodiment of the present invention, the data transmission scheme found by the particle swarm optimization algorithm can fully consider the network environment and transmission requirements, so as to select the most appropriate transmission path, data compression rate and transmission protocol parameters. This can significantly reduce the delay of data transmission, improve the transmission speed, and enable the risk level prediction results and the original data to reach the cloud server or user terminal more quickly. The optimized transmission scheme can select a lower data compression rate and a more efficient transmission protocol under the premise of ensuring the transmission quality, thereby reducing the resource consumption during the data transmission process. This helps to reduce the occupancy of network bandwidth and the energy consumption during the transmission process, thereby reducing the overall transmission cost. Since the network environment may change over time, the particle swarm optimization algorithm can dynamically adjust the transmission scheme to adapt to these changes. This flexibility enables the system to maintain efficient data transmission performance under different network conditions. The optimized transmission scheme usually selects a more stable transmission path and protocol, thereby reducing the packet loss rate and error rate during the data transmission process. This helps to improve the reliability of the entire system and ensure that the risk level prediction results and the original data can be accurately transmitted to the target location.

[0200] Furthermore, the fitness function The calculation formula is:

[0201] ;

[0202] in, Indicates the normalized value of actual transmission time; Indicates the amount of data successfully transferred; Indicates the total amount of data transferred; Indicates the normalized value of the bandwidth used; Indicates the actual transmission rate; Indicates the maximum transmission rate of the network.

[0203] In the embodiment of the present invention, each transmission scheme is recorded , , , , and ; Determine a reasonable transmission time range, which can be set based on the actual network environment and application requirements, denoted as [T min , T max ]. Among them, T min represents the shortest possible transmission time theoretically, and T max It represents the longest acceptable transmission time. Next, the actual transmission time T is measured. x In order to convert this actual value to a uniform scale, normalization is performed using the following steps:

[0204] From T max Subtract the actual transmission time T from x The result is the "remaining time" of the acceptable upper limit of the actual transmission time distance. Divide the above "remaining time" by (T max -T min ), that is, the entire acceptable time range. The obtained ratio T n It's T x The normalized value of T n The closer the value is to 1, the shorter the transmission time is and the better the performance is.

[0205] The actual bandwidth used is B u Divide by the maximum bandwidth of the network B max , the obtained ratio B n It's B u The normalized value of B, which represents the ratio of the used bandwidth to the maximum bandwidth. n The closer the value is to 1, the higher the bandwidth utilization rate is. Substitute the normalized value into the normalized fitness function Calculate the performance of different transmission schemes; value, higher A higher value indicates a better transmission solution. Normalization ensures that indicators of different dimensions and value ranges can be fairly compared and weighted in the fitness function; by adjusting the normalization method and the weights in the fitness function, it can easily adapt to different optimization goals and network environments; the fitness function comprehensively considers multiple key transmission performance indicators, so that the advantages and disadvantages of the transmission solution can be more comprehensively evaluated; a high fitness value directly points to a better transmission solution, providing a clear search direction for the particle swarm optimization algorithm.

[0206] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote collection system for gas equipment operation data, characterized in that: include: A graph construction module is used to perform preliminary feature extraction on preprocessed data using a convolutional neural network model to capture local features in gas equipment operation data, wherein the preprocessed data is preprocessed gas equipment operation data, including pressure, temperature, flow rate and leakage; extract deep features through a deep residual network based on local features; combine the deep features with equipment entities, attributes, relationships and timestamp information to construct a preliminary knowledge graph; jointly annotate entities and relationships in the preliminary knowledge graph using a conditional random field model, optimize the preliminary knowledge graph using a graph convolutional network, and update nodes through iteration; The risk prediction module is connected to the graph construction module and receives the preliminary knowledge graph as input data; the long short-term memory network uses the time series data in the preliminary knowledge graph as sequence input data, captures the law of the change of the gas equipment state over time by learning the long-term dependency in the sequence, and generates key features containing time series information; the time series feature representation output by the long short-term memory network is integrated with the equipment attributes and relationships in the knowledge graph to form a feature vector reflecting the comprehensive state of the gas equipment; Comprehensively analyze the feature vectors to obtain hidden patterns related to gas equipment risks; Based on the hidden patterns related to gas equipment risks, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction result; A remote transmission module, connected to the risk prediction module, for remotely transmitting the final risk level prediction result and raw data to a cloud server or a user terminal via a wireless network; The user interaction module is connected to the remote transmission module and is used to provide a visual interface to the user to display the real-time operating status of the gas equipment, the final risk level prediction results and historical data.

2. A remote collection system for gas equipment operation data according to claim 1, characterized in that: The convolutional neural network model is used to perform preliminary feature extraction on the preprocessed data to capture the local features in the gas equipment operation data, including: Determine the basic architecture of the convolutional neural network model, including convolutional layers, pooling layers, and fully connected layers; Define the parameters of the genetic algorithm, including population size, crossover rate, mutation rate, and number of generations; Encode the hyperparameters of the convolutional neural network model as genes of the genetic algorithm to form an initial population. The hyperparameters include learning rate, batch size, and number of convolution kernels. The preprocessed data is divided into training set, validation set and test set. For each generation of population, the following steps are performed: Decode each individual in the population into the hyperparameter settings of a convolutional neural network model; Train the convolutional neural network model using the training set and the current individual hyperparameter settings. In each training cycle, forward propagation, loss calculation, backpropagation and parameter update are performed; After each training cycle, the validation set is used to evaluate the performance of the convolutional neural network model and the accuracy is recorded; According to the accuracy rate on the validation set, the corresponding individuals are selected to enter the next generation population. Two individuals in the population are randomly selected, and some of their genes are exchanged to generate new individuals. The new individuals are used to replace some individuals in the original population to form the next generation population, until the preset evolutionary generation is reached and the genetic algorithm cycle is stopped. Select the corresponding individuals from the last generation of population as the final hyperparameters, and retrain the convolutional neural network model using the final hyperparameter settings to obtain the trained convolutional neural network model; The convolutional layer of the trained convolutional neural network model is used as a feature extractor, the input data is forward propagated through the convolutional layer, and the output of the convolutional layer is obtained as the extracted local feature.

3. A remote collection system for gas equipment operation data according to claim 2, characterized in that: Accuracy The calculation formula is: ； in, Yes Category The true positives, i.e., the categories The number of samples correctly predicted as positive; Yes Category The true negative of The number of samples correctly predicted as negative class; Yes Category The false positive of the category The number of negative samples that are incorrectly predicted as positive; Yes Category The false negatives of The number of positive samples that are incorrectly predicted as negative; is the total number of categories; is the mean of the Gaussian distribution, indicating a certain central location; is the standard deviation; Yes Category The number of samples; is the index value of the category.

4. A remote collection system for gas equipment operation data according to claim 3, characterized in that: According to local features, deep features are extracted through deep residual networks; Combine deep features with device entities, attributes, relationships, and timestamp information to build a preliminary knowledge graph, including: The deep residual network is trained using local features, and the network parameters are optimized through back propagation to obtain the trained deep residual network; Apply the trained deep residual network to the data of all local features to extract deep features; Integrate the deep features with the corresponding device entity information, attributes, relationships and timestamp information, so that the feature data and metadata correspond one-to-one to obtain the integrated data; Based on the integrated data, a preliminary equipment knowledge graph is constructed, in which nodes represent equipment entities, edges represent the relationships between entities, attribute information is attached to the corresponding nodes, and timestamp information is used to describe the dynamic changes of entities and relationships.

5. A remote collection system for gas equipment operation data according to claim 4, characterized in that: The conditional random field model is used to jointly annotate entities and relationships in the preliminary knowledge graph, and the graph convolutional network is used to optimize the preliminary knowledge graph, and the nodes are updated through iteration, including: Extract features for the conditional random field model from the preliminary knowledge graph, including part of speech, context information, and entity type; Use the trained conditional random field model to jointly annotate entities and relationships on the preliminary knowledge graph and identify the relationships between entities in the preliminary knowledge graph. The knowledge graph annotated by the conditional random field model is converted into graph structure data, where nodes represent entities and edges represent relationships. A feature matrix is ​​constructed for each node, which includes the attributes of the node and the information of adjacent nodes. Use graph convolutional networks to optimize the graph and capture the deep connections and features between nodes through multi-layer graph convolution operations; According to the deep associations and features between nodes, the entities and relationships in the graph are adjusted and optimized to obtain the optimization results; based on the optimization results, the status and attributes of each node in the graph are updated; The graph convolutional neural network is repeatedly used for multiple iterations. Each iteration adjusts the graph according to the results of the previous one until the preset number of iterations is reached. After the iteration is completed, the graph is finally optimized and adjusted.

6. A remote collection system for gas equipment operation data according to claim 5, characterized in that: The LSTM network uses the time series data in the preliminary knowledge graph as sequence input data. By learning the long-term dependencies in the sequence, it captures the law of changes in the state of the gas equipment over time and generates key features containing time series information, including: Extract time-series related data from the preliminary knowledge graph, including device status records and sensor readings arranged in chronological order; Read the data related to the time series and set a corresponding time window size; divide the time series data according to the set time window, form a sequence of data in each window, and standardize each sequence to obtain the processed time series data sequence; Determine the number of LSTM network layers, set the number of hidden units in each LSTM network layer, and determine the input and output dimensions of the network. The input dimensions match the dimensions of the processed time series data sequence. Use the deep learning framework to create an LSTM model, convert the processed time series data sequence into the tensor format required by the LSTM model, and divide the processed time series data sequence into training set, validation set, and test set; Use the training set data to train the LSTM model. After each training round, use the validation set to evaluate the performance of the LSTM model. Adjust the learning rate and optimizer parameters based on the performance of the validation set. Use the back-propagation algorithm to update the weights and biases of the LSTM network to obtain the trained LSTM model. The processed time series data sequence is input into the trained LSTM model, and the trained LSTM model outputs the hidden state of each time step. The hidden state contains the long-term dependencies and key features in the time series data.

7. A remote collection system for gas equipment operation data according to claim 6, characterized in that: The temporal feature representation output by the long short-term memory network is integrated with the device attributes and relationships in the knowledge graph to form a feature vector that reflects the comprehensive status of the gas equipment. The feature vectors are comprehensively analyzed to obtain hidden patterns related to gas equipment risks, including: Obtain the temporal feature representation of the LSTM network output, and extract static attributes and relationship information related to gas equipment from the knowledge graph, where static attributes include equipment model, manufacturer, and installation date; relationship information includes connection relationship with the equipment and maintenance records; The temporal features output by the long short-term memory network are concatenated with the static features and relational features extracted from the knowledge graph to form a fused feature vector; K-means is used to cluster the fused feature vectors to find groups with similar device status and behavior patterns; The isolation forest algorithm is used to identify feature vectors that are different from the normal pattern. Different feature vectors represent potential risks. Hidden patterns associated with gas equipment risks are identified based on static attributes and relationship information related to gas equipment, groups with similar equipment status and behavior patterns, and potential risks.

8. A remote collection system for gas equipment operation data according to claim 7, characterized in that: Based on the hidden patterns related to gas equipment risks, a multi-classification support vector machine is used to predict the risk level to obtain the final risk level prediction results, including: Assign a risk level label to each latent pattern feature vector based on historical data; Set the parameters of the SVM model, including the kernel function, penalty coefficient C, and kernel function parameters; Train the SVM model using the latent pattern feature set with risk level labels; By dividing the training set and the validation set, the performance and generalization ability of the SVM model are evaluated to obtain the validation results; Adjust the SVM model parameters according to the verification results to obtain the final SVM model; For new gas equipment data, extract its hidden pattern feature vector; The hidden pattern feature vector is input into the final SVM model to predict the risk level to obtain the final risk level prediction result.

9. A remote collection system for gas equipment operation data according to claim 8, characterized in that: The final risk level prediction results and raw data are remotely transmitted to the cloud server or user terminal via wireless network, including: According to the network environment and transmission requirements, a particle swarm is initialized, and each particle represents a data transmission scheme, including transmission path, data compression rate and transmission protocol parameters; Define a fitness function to evaluate the performance of each transmission scheme; Update the speed and position of each particle according to the individual final solution and the global final solution, and obtain a data transmission scheme suitable for the current network environment and transmission requirements through multiple iterations; After the iteration is completed, the optimized transmission scheme corresponding to the global final solution is selected; Using the optimized transmission scheme, the final risk level prediction results and original data are remotely transmitted to the cloud server or user terminal via the wireless network.

10. A remote collection system for gas equipment operation data according to claim 9, characterized in that: Fitness function The calculation formula is: ； in, Indicates the normalized value of actual transmission time; Indicates the amount of data successfully transferred; Indicates the total amount of data transferred; Indicates the normalized value of the bandwidth used; Indicates the actual transmission rate; Indicates the maximum transmission rate of the network.

Images