Chemical process safety risk early warning system based on edge calculation

Abstract

The invention discloses a chemical process safety risk early warning system based on edge calculation, and the system comprises an edge calculation module which is used for receiving real-time process data of a chemical device, and generating structured feature data; the ordered target statistical module is used for receiving the structured feature data and generating target statistical data; the improved CatBoost model module is used for receiving the target statistical data and generating a chemical process safety risk early warning prediction result; the edge lightweight model module is used for performing preprocessing on the trained improved CatBoost model to generate an edge deployment model version; the risk assessment module is used for generating chemical process safety risk early warning information; and the communication interface module is used for sending the safety risk early warning information of the chemical process to a field operation terminal. According to the invention, a closed-loop system with linkage of real-time data processing, autonomous risk prediction and on-site early warning on a chemical production site is realized.

Description

Technical Field

[0001] This invention relates to the fields of edge computing, industrial process safety and intelligent manufacturing, and in particular to a chemical process safety risk early warning system based on edge computing. Background Technology

[0002] In chemical production processes, due to the complexity of the processes, the large size of the equipment systems, and the variability of operating conditions, safety risks such as leaks, explosions, and poisoning can easily arise due to equipment failures, parameter fluctuations, and improper operation, seriously threatening personnel safety and company property. To ensure the safety of chemical processes, traditional methods mainly rely on manual inspections, fixed threshold alarms, and centralized data processing platforms. However, these methods suffer from problems such as slow response times, high processing pressure, and poor adaptability to different scenarios, making it difficult to meet the higher requirements for real-time, accurate, and intelligent safety risk assessment in today's intelligent manufacturing environment.

[0003] With the rapid development of artificial intelligence and big data technologies, machine learning models are widely used in chemical process risk identification and early warning scenarios. Among them, the CatBoost model, as a high-performance gradient boosting decision tree model, is gaining increasing attention in industrial scenarios due to its excellent ability to handle categorical features and its prediction accuracy. However, in actual deployment, the CatBoost model still suffers from problems such as high computational resource consumption, complex model structure, and low inference efficiency, making it unsuitable for direct application to edge devices. Furthermore, traditional CatBoost models do not fully consider the ordered nature of samples during categorical feature processing, leading to risks of target leakage and overfitting, and lack effective mechanisms for real-time updates and incremental learning. In addition, most existing models rely on centralized computing models, making distributed edge intelligent deployment impossible and limiting their real-time response capabilities in industrial settings.

[0004] Currently, some research has begun to explore edge computing architectures, enabling local inference and real-time early warning by deploying models to edge nodes. However, significant challenges remain: how to deploy efficiently compressed machine learning models on resource-constrained edge nodes, how to ensure the reliability and interpretability of model predictions, and how to achieve dynamic incremental updates of the model during operation. Especially in complex chemical industrial settings, building a stable, efficient, and scalable early warning system by combining process data streams with model inference mechanisms remains a challenging task.

[0005] Therefore, how to provide a chemical process safety risk early warning system based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a chemical process safety risk early warning system based on edge computing. This invention fully integrates edge computing architecture, CatBoost model improvement strategies, and ordered target statistical coding methods, and details how to achieve real-time acquisition, intelligent processing, model inference, and dynamic updating of chemical process data on edge computing nodes. By introducing an adaptive symmetric tree construction mechanism and edge lightweight model compression technology, efficient model deployment and inference efficiency are improved. Combined with an incremental training mechanism, the model continuously evolves with changes in on-site data, possessing advantages such as fast response speed, low resource consumption, high prediction accuracy, and adaptability to complex on-site conditions.

[0007] A chemical process safety risk early warning system based on edge computing according to an embodiment of the present invention includes:

[0008] The edge computing module is used to receive real-time process data from the chemical plant, cache the real-time process data as time slice data according to a preset sampling period, perform data preprocessing and feature construction operations on the time slice data, and generate structured feature data.

[0009] The ordered target statistics module is used to receive structured feature data, identify categorical feature fields in the structured feature data, perform ordered target statistical encoding on the categorical features, and generate target statistical data.

[0010] The improved CatBoost model module is used to receive target statistical data, perform adaptive symmetric tree construction and adaptive symmetric tree inference based on the target statistical data, and generate early warning prediction results for chemical process safety risks. The improved CatBoost model module includes an adaptive symmetric tree module and an ordered gradient boosting framework. The adaptive symmetric tree module includes a hierarchical split controller, a hierarchical optional symmetric module, and a local structure self-checking module.

[0011] The edge lightweight model module is used to perform pruning, quantization and structural compression on the trained improved CatBoost model to generate an edge deployment model version, and to perform incremental updates on the model parameters in the edge deployment model version based on continuously collected real-time process data.

[0012] The risk assessment module is used to compare the early warning and prediction results of chemical process safety risks with preset risk level thresholds and generate early warning information for chemical process safety risks.

[0013] The communication interface module is used to send early warning information on chemical process safety risks to the field operation terminal.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Step 1: The edge computing node receives real-time process data and caches the real-time process data as time slice data according to the preset sampling period; it performs data preprocessing and feature construction operations on the time slice data to obtain structured feature data;

[0016] Step 2: Input the structured feature data into the ordered target statistics module, perform ordered target statistical encoding on the categorical features in the structured feature data, and generate target statistical data.

[0017] Step 3: Call the adaptive symmetric tree module in the improved CatBoost model to perform adaptive symmetric tree construction on the target statistical data;

[0018] The hierarchical split controller determines the current set of nodes to be split and sends the set of nodes to be split to the hierarchical optional symmetric module to generate corresponding candidate symmetric splits and local split candidates of the nodes to be split;

[0019] Based on preset judgment conditions, a pure symmetric splitting is selected to generate a splitting layer structure. This splitting layer structure is then sent to the local structure self-inspection module, which then performs the inspection on the splitting layer structure.

[0020] If the test fails, a rollback operation is performed on this split layer structure; if the test passes, this split layer structure is confirmed as the valid structure of the current split layer.

[0021] Step 4: Perform adaptive symmetric tree inference. Calculate the leaf values ​​of the tree structure confirmed by the local structure self-checking module to obtain the output of a single adaptive symmetric tree. In the ordered gradient boosting framework, perform a weighted summation of the outputs of all adaptive symmetric trees to obtain the chemical process safety risk early warning prediction results.

[0022] Step 5: Deploy the improved CatBoost model on an edge computing node, perform pruning, quantization, and structural compression on the trained improved CatBoost model through the edge lightweight model module, and perform incremental updates on the improved CatBoost model version on the edge computing node based on continuously collected real-time process data;

[0023] Step 6: In the edge computing node, compare the prediction results of chemical process safety risk warning with the preset risk level threshold, generate the corresponding chemical process safety risk warning information, and send it to the field operation terminal.

[0024] Optionally, the real-time process data includes temperature, pressure, flow rate, liquid level, valve opening degree, and equipment vibration signal; data preprocessing and feature construction operations are performed on the time slice data, including filling in missing data, normalizing numerical data, and generating statistical features and rate of change features based on a sliding window.

[0025] Optionally, the improved CatBoost model includes an ordered target statistics module, an adaptive symmetric tree module, an ordered gradient boosting framework, and an edge lightweight model module;

[0026] Optionally, step two specifically involves:

[0027] The ordered target statistics module receives structured feature data in the edge computing node and identifies categorical feature fields in the structured feature data;

[0028] For each category feature, at least one sample sequence is generated by the ordered target statistics module, and samples in the structured feature data are read sequentially according to the order of the sample sequence.

[0029] The ordered target statistical value is obtained by performing statistical operations on the supervised learning target values ​​of historical samples that are earlier than the current sample in the sample sequence and have the same categorical feature values. The statistical operation is to average the supervised learning target values.

[0030] Based on the ordered target statistics, assign corresponding code values ​​to the categories in each categorical feature, and establish a mapping relationship between category values ​​and code values;

[0031] By using mapping relationships, the categorical feature fields in the structured feature data are replaced with the corresponding coded values, while retaining the original numerical feature fields, to generate the target statistical data.

[0032] Optionally, step three specifically includes:

[0033] Input the target statistical data into the adaptive symmetric tree module in the improved CatBoost model, and initialize the root node for each adaptive symmetric tree to be trained under the ordered gradient boosting framework;

[0034] Perform adaptive symmetric tree construction:

[0035] The hierarchical split controller determines the current set of nodes to be split at each splitting level of the current adaptive symmetric tree and sends the set of nodes to be split to the hierarchical optional symmetric module;

[0036] The hierarchical optional symmetric module generates candidate symmetric splits for the current splitting layer and local splitting candidates for the nodes to be split based on the target statistical data of the samples contained in each node in the set of nodes to be split and the supervised learning target value.

[0037] The layered optional symmetric module selects a pure symmetric split for the current split layer from among candidate symmetric splits and local split candidates according to preset judgment conditions, thus forming the structure of this split layer;

[0038] The split layer structure is sent to the local structure self-inspection module, which then detects the split layer structure based on the loss value of the parent node set before splitting, the loss value of the child nodes after splitting, and the sample number and label distribution of the child nodes.

[0039] If the test fails, the local structure self-test module performs a rollback operation on the current split layer structure, and the hierarchical split controller re-determines the set of nodes to be split in the current split layer.

[0040] When the test passes, the local structure self-test module confirms that the structure of this split layer is the valid structure of the current split layer, and uses the child nodes formed by the current split layer as the set of nodes to be split in the next split layer.

[0041] Repeat the adaptive symmetric tree construction process until the current adaptive symmetric tree meets the preset splitting stopping condition, thus completing the structure construction of the current adaptive symmetric tree;

[0042] And in the ordered gradient boosting framework, the adaptive symmetric tree construction process is performed sequentially on all adaptive symmetric trees.

[0043] Optionally, the execution of adaptive symmetric tree inference specifically includes:

[0044] Under the control of the ordered gradient boosting framework, the target statistical data is sequentially input into each adaptive symmetric tree that has completed the structure construction;

[0045] For each adaptive symmetric tree, based on the splitting conditions of each splitting layer in the adaptive symmetric tree, a path determination is performed on each sample record in the target statistical data, and each sample record is assigned to the corresponding leaf node according to the splitting path of the sample record in the adaptive symmetric tree;

[0046] For each leaf node in each adaptive symmetric tree, the leaf value of the leaf node is calculated based on the supervised learning target value of the training sample assigned to the leaf node and the update amount information determined by the ordered gradient boosting framework for the current iteration, and the output result of the adaptive symmetric tree is generated using the leaf value.

[0047] The outputs of all adaptive symmetric trees are weighted and accumulated sequentially according to the preset tree weight parameters in the ordered gradient boosting framework to obtain the chemical process safety risk early warning prediction results for the target statistical data.

[0048] Optionally, step five specifically includes:

[0049] After training the improved CatBoost model in the training environment, export the model parameters and adaptive symmetric tree structure of the improved CatBoost model, generate deployment model data, and distribute the deployment model data to the edge computing nodes;

[0050] In the edge computing node, the deployment model data is loaded through the edge lightweight model module, and some adaptive symmetric trees, split layers and leaf nodes in the deployment model data are deleted according to the preset pruning rules to obtain the pruned model data.

[0051] In the edge lightweight model module, the model parameters in the pruning model data are quantized, converting the floating-point split threshold, leaf value and tree weight parameters into a low-bit-width fixed-point numerical representation to generate quantized model data.

[0052] In the edge lightweight model module, structural compression processing is performed on the quantized model data. The adaptive symmetric tree structure, split condition index and leaf value in the quantized model data are re-encoded according to the preset data structure to generate the edge deployment model version.

[0053] An incremental training dataset is formed based on continuously collected real-time process data in the edge computing node. The leaf value parameters and tree weight parameters in the edge deployment model version are updated by the edge lightweight model module using the incremental training dataset, thus obtaining the updated edge deployment model version.

[0054] Optionally, step six specifically includes:

[0055] In the edge computing node, the edge deployment model version is updated, and adaptive symmetric tree inference is performed on the target statistical data to obtain the early warning and prediction results of chemical process safety risks.

[0056] In the edge computing node, a preset risk level threshold set is called, and the chemical process safety risk early warning prediction result is compared with each threshold in the preset risk level threshold set one by one to determine the risk level range in which the chemical process safety risk early warning prediction result is located.

[0057] Based on the risk level range of the chemical process safety risk early warning and prediction results, a risk level label is generated for the chemical process safety risk early warning and prediction results.

[0058] Chemical process safety risk early warning information is constructed in edge computing nodes. The chemical process safety risk early warning information includes risk level identifier, chemical process safety risk early warning prediction result value, real-time process data timestamp corresponding to the chemical process safety risk early warning prediction result, equipment identifier and process unit identifier.

[0059] The chemical process safety risk warning information is sent to the field operation terminal through the communication interface of the edge computing node.

[0060] The beneficial effects of this invention are:

[0061] This invention relates to a chemical process safety risk early warning system based on edge computing. Firstly, it establishes a closed-loop chain for the acquisition, modeling, prediction, and updating of on-site data from chemical plants by constructing an ordered target statistics module, an adaptive symmetric tree module, and a lightweight edge model module. The system utilizes edge computing nodes to acquire and structure real-time process data, and leverages an improved CatBoost model to achieve efficient ordered feature encoding and tree model training, effectively improving the modeling effect on categorical data and ensuring the stability and accuracy of predictions.

[0062] An adaptive symmetric tree mechanism is introduced during model construction, employing a hierarchical optional symmetry strategy and a local structure self-checking mechanism. This makes the model structure more flexible and controllable, improving the rationality of the split structure and the model's generalization ability. Simultaneously, by constructing and reasoning the tree layer by layer within an ordered gradient boosting framework, the model can better adapt to the nonlinear characteristics and complex variable relationships in chemical processes, significantly improving the response speed and prediction accuracy of risk warnings.

[0063] This invention also includes a lightweight edge model module that performs pruning, quantization, and structural compression on the model, significantly reducing the operational load on edge computing devices. Simultaneously, the system supports an incremental update mechanism based on real-time data acquisition, adjusting only the leaf value parameters and tree weight parameters in the deployed model without rebuilding the entire tree structure, greatly improving the efficiency of model updates and the practicality of edge deployment. Finally, by combining risk level thresholds to determine the early warning results, accurate risk information is pushed to on-site operation terminals, enhancing the system's interpretability and decision support capabilities in practical applications.

[0064] Therefore, this invention, through hardware and software co-optimization, ensures the feasibility of deployment in edge environments while taking into account model performance and update flexibility. It has significant advantages such as rapid response, high prediction accuracy, low resource consumption, and strong maintainability, meeting the safety assurance requirements under high-risk conditions in chemical processes. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a schematic diagram of the overall structure of a chemical process safety risk early warning system based on edge computing proposed in this invention.

[0067] Figure 2 This is an overall flowchart of a chemical process safety risk early warning method based on edge computing proposed in this invention;

[0068] Figure 3This is a schematic diagram of the improved CatBoost model structure of a chemical process safety risk early warning system based on edge computing proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1-3 A chemical process safety risk early warning system based on edge computing, comprising:

[0071] The edge computing module is used to receive real-time process data from the chemical plant, cache the real-time process data as time slice data according to a preset sampling period, perform data preprocessing and feature construction operations on the time slice data, and generate structured feature data.

[0072] The ordered target statistics module, deployed on the edge computing module, is used to receive structured feature data, identify categorical feature fields in the structured feature data, perform ordered target statistical encoding on the categorical features, and generate target statistical data.

[0073] The improved CatBoost model module is used to receive target statistical data, perform adaptive symmetric tree construction and adaptive symmetric tree inference based on the target statistical data, and generate early warning prediction results for chemical process safety risks. The improved CatBoost model module includes an adaptive symmetric tree module and an ordered gradient boosting framework. The adaptive symmetric tree module includes a hierarchical split controller, a hierarchical optional symmetric module, and a local structure self-checking module.

[0074] The edge lightweight model module, deployed in the edge computing module, is used to perform pruning, quantization and structural compression on the trained improved CatBoost model to generate an edge deployment model version, and to perform incremental updates on the model parameters in the edge deployment model version based on continuously collected real-time process data.

[0075] The risk assessment module, deployed on the edge computing module, is used to compare the early warning and prediction results of chemical process safety risks with preset risk level thresholds to generate early warning information of chemical process safety risks.

[0076] The communication interface module is used to send early warning information on chemical process safety risks to the field operation terminal.

[0077] In this embodiment, the modules are connected through the following method:

[0078] Step 1: The edge computing node receives real-time process data and caches the real-time process data as time slice data according to the preset sampling period; it performs data preprocessing and feature construction operations on the time slice data to obtain structured feature data;

[0079] Step 2: Input the structured feature data into the ordered target statistics module, perform ordered target statistical encoding on the categorical features in the structured feature data, and generate target statistical data.

[0080] Step 3: Call the adaptive symmetric tree module in the improved CatBoost model to perform adaptive symmetric tree construction on the target statistical data;

[0081] The hierarchical split controller determines the current set of nodes to be split and sends the set of nodes to be split to the hierarchical optional symmetric module to generate corresponding candidate symmetric splits and local split candidates of the nodes to be split;

[0082] Based on preset judgment conditions, a pure symmetric splitting is selected to generate a splitting layer structure. This splitting layer structure is then sent to the local structure self-inspection module, which then performs the inspection on the splitting layer structure.

[0083] If the test fails, a rollback operation is performed on this split layer structure; if the test passes, this split layer structure is confirmed as the valid structure of the current split layer.

[0084] Step 4: Calculate the leaf values ​​of the tree structure confirmed by the local structure self-checking module to obtain the output of a single adaptive symmetric tree. Then, in the ordered gradient boosting framework, sum the weighted outputs of all adaptive symmetric trees to obtain the chemical process safety risk early warning prediction results.

[0085] Step 5: Deploy the improved CatBoost model on an edge computing node, perform pruning, quantization, and structural compression on the trained improved CatBoost model through the edge lightweight model module, and perform incremental updates on the improved CatBoost model version on the edge computing node based on continuously collected real-time process data;

[0086] Step 6: In the edge computing node, compare the prediction results of chemical process safety risk warning with the preset risk level threshold, generate the corresponding chemical process safety risk warning information, and send it to the field operation terminal.

[0087] In this embodiment, the real-time process data includes temperature, pressure, flow rate, liquid level, valve opening degree, and equipment vibration signal; data preprocessing and feature construction operations are performed on the time slice data, including filling in missing data, normalizing numerical data, and generating statistical features and rate of change features based on a sliding window.

[0088] In this embodiment, the improved CatBoost model includes an ordered target statistics module, an adaptive symmetric tree module, an ordered gradient boosting framework, and an edge lightweight model module; the adaptive symmetric tree module includes a hierarchical split controller, a hierarchical optional symmetry module, and a local structure self-checking module.

[0089] In this embodiment, step two specifically includes:

[0090] The ordered target statistics module receives structured feature data in the edge computing node and identifies categorical feature fields in the structured feature data;

[0091] For each category feature, at least one sample sequence is generated by the ordered target statistics module, and samples in the structured feature data are read sequentially according to the order of the sample sequence.

[0092] The ordered target statistical value is obtained by performing statistical operations on the supervised learning target values ​​of historical samples that are earlier than the current sample in the sample sequence and have the same categorical feature values. The statistical operation is to average the supervised learning target values.

[0093] The supervised learning objective value is the true label value corresponding to each sample in the training dataset;

[0094] Based on the ordered target statistics, assign corresponding code values ​​to the categories in each categorical feature, and establish a mapping relationship between category values ​​and code values;

[0095] By using mapping relationships, the categorical feature fields in the structured feature data are replaced with the corresponding coded values, while retaining the original numerical feature fields, to generate the target statistical data.

[0096] In this embodiment, step three specifically includes:

[0097] Input the target statistical data into the adaptive symmetric tree module in the improved CatBoost model, and initialize the root node for each adaptive symmetric tree to be trained under the ordered gradient boosting framework;

[0098] Perform adaptive symmetric tree construction:

[0099] The hierarchical split controller determines the current set of nodes to be split at each splitting level of the current adaptive symmetric tree and sends the set of nodes to be split to the hierarchical optional symmetric module;

[0100] The hierarchical optional symmetric module generates candidate symmetric splits for the current splitting layer and local splitting candidates for the nodes to be split based on the target statistical data of the samples contained in each node in the set of nodes to be split and the supervised learning target value.

[0101] The layered optional symmetric module selects a pure symmetric split for the current split layer from among candidate symmetric splits and local split candidates according to preset judgment conditions, thus forming the structure of this split layer;

[0102] The preset determination conditions include at least one of the following:

[0103] The information gain of the candidate symmetric split on all nodes to be split is not lower than the preset difference threshold of the information gain of the corresponding local split candidate.

[0104] The sample distribution differences of the set of nodes to be split meet the preset distribution consistency threshold;

[0105] The number of samples of the left and right child nodes generated by the candidate symmetric split is higher than the preset sample capacity threshold.

[0106] The split layer structure is sent to the local structure self-inspection module, which then detects the split layer structure based on the loss value of the parent node set before splitting, the loss value of the child nodes after splitting, and the sample number and label distribution of the child nodes.

[0107] The loss value of the parent node set before splitting is calculated by the degree of deviation between the supervised learning target value of all samples in the parent node set and the model's prediction value on the parent node, and the loss is determined by summing up the losses.

[0108] The loss value of the child nodes after splitting is calculated by the degree of deviation between the supervised learning target value of all samples in each child node and the model's predicted value on that child node, and the loss is determined by summing or averaging.

[0109] The detection includes a local structure self-inspection module comparing the change in the sum of the loss values ​​of the parent node set before splitting and the loss values ​​of the child nodes after splitting, the number of samples of each child node, and the distribution differences of the label distribution of each child node with the corresponding preset threshold conditions to determine whether the split layer structure passes the detection.

[0110] If the test fails, the local structure self-test module performs a rollback operation on the current split layer structure, and the hierarchical split controller re-determines the set of nodes to be split in the current split layer.

[0111] When the test passes, the local structure self-test module confirms that the structure of this split layer is the valid structure of the current split layer, and uses the child nodes formed by the current split layer as the set of nodes to be split in the next split layer.

[0112] Repeat the adaptive symmetric tree construction process until the current adaptive symmetric tree meets the preset splitting stopping condition, thus completing the structure construction of the current adaptive symmetric tree;

[0113] And in the ordered gradient boosting framework, the adaptive symmetric tree construction process is performed sequentially on all adaptive symmetric trees.

[0114] This invention introduces an adaptive symmetric tree construction mechanism into an improved CatBoost model, achieving hierarchical modeling and dynamic structural optimization of target statistical data. Specifically, after initializing the root node for each adaptive symmetric tree to be trained within the ordered gradient boosting framework, a hierarchical split controller determines the set of nodes to be split layer by layer. In the hierarchical optional symmetry module, candidate symmetric splits and local split candidates are generated based on the target statistical data of each node sample and the supervised learning target value. Then, a pure symmetric split is selected to form the current split layer structure according to preset judgment conditions. Subsequently, the local structure self-checking module detects the split layer structure by combining the loss value of the parent node set before splitting, the loss value of the child nodes after splitting, the number of child node samples, and the label distribution. Based on the detection results, a rollback or confirmation of a valid structure is performed, ensuring that the splitting operation of each layer undergoes structural rationality verification. By repeating the above process for all split layers until the preset splitting stopping condition is met, the current adaptive symmetric tree construction is completed, and all adaptive symmetric trees are sequentially constructed within the ordered gradient boosting framework. This invention achieves adaptive generation and optimized construction of tree models while ensuring structural stability and prediction reliability.

[0115] In this embodiment, step four specifically includes:

[0116] Under the control of the ordered gradient boosting framework, the target statistical data is sequentially input into each adaptive symmetric tree that has completed the structure construction;

[0117] For each adaptive symmetric tree, based on the splitting conditions of each splitting layer in the adaptive symmetric tree, a path determination is performed on each sample record in the target statistical data, and each sample record is assigned to the corresponding leaf node according to the splitting path of the sample record in the adaptive symmetric tree;

[0118] For each leaf node in each adaptive symmetric tree, the leaf value of the leaf node is calculated based on the supervised learning target value of the training sample assigned to the leaf node and the update amount information determined by the ordered gradient boosting framework for the current iteration, and the output result of the adaptive symmetric tree is generated using the leaf value.

[0119] The outputs of all adaptive symmetric trees are weighted and accumulated sequentially according to the preset tree weight parameters in the ordered gradient boosting framework to obtain the chemical process safety risk early warning prediction results for the target statistical data.

[0120] In this embodiment, step five specifically includes:

[0121] After training the improved CatBoost model in the training environment, export the model parameters and adaptive symmetric tree structure of the improved CatBoost model, generate deployment model data, and distribute the deployment model data to the edge computing nodes;

[0122] In the edge computing node, the deployment model data is loaded through the edge lightweight model module, and some adaptive symmetric trees, split layers and leaf nodes in the deployment model data are deleted according to the preset pruning rules to obtain the pruned model data.

[0123] In the edge lightweight model module, the model parameters in the pruning model data are quantized, converting the floating-point split threshold, leaf value and tree weight parameters into a low-bit-width fixed-point numerical representation to generate quantized model data.

[0124] In the edge lightweight model module, structural compression processing is performed on the quantized model data. The adaptive symmetric tree structure, split condition index and leaf value in the quantized model data are re-encoded according to the preset data structure to generate the edge deployment model version.

[0125] An incremental training dataset is formed based on continuously collected real-time process data in the edge computing node. The leaf value parameters and tree weight parameters in the edge deployment model version are updated by the edge lightweight model module using the incremental training dataset, thus obtaining the updated edge deployment model version.

[0126] The leaf value parameter is updated through the following steps:

[0127] In the edge computing nodes, each sample record in the incremental training dataset is assigned to its respective leaf node according to the splitting path of the adaptive symmetric tree;

[0128] For each leaf node, based on the supervised learning target value of the samples in the incremental training dataset assigned to that leaf node and the update amount information determined by the ordered gradient boosting framework for the current iteration, calculate the update amount value used to update the leaf value of the leaf node.

[0129] The updated value is combined with the original leaf value parameter in the edge deployment model version to generate the updated leaf value parameter.

[0130] The tree weight parameters are updated through the following steps:

[0131] Based on the output of the incremental training dataset on each adaptive symmetric tree, calculate the residual information of the adaptive symmetric tree on the incremental training dataset;

[0132] The update amount used to update the tree weight parameters is determined based on residual information and the pre-defined weight update rules in the ordered gradient boosting framework.

[0133] The updated value is combined with the original tree weight parameters in the edge deployment model version to generate the updated tree weight parameters.

[0134] In this embodiment, step six specifically includes:

[0135] In the edge computing node, the edge deployment model version is updated, and adaptive symmetric tree inference is performed on the target statistical data to obtain the early warning and prediction results of chemical process safety risks.

[0136] In the edge computing node, a preset risk level threshold set is called, and the chemical process safety risk early warning prediction result is compared with each threshold in the preset risk level threshold set one by one to determine the risk level range in which the chemical process safety risk early warning prediction result is located.

[0137] Based on the risk level range of the chemical process safety risk early warning and prediction results, a risk level label is generated for the chemical process safety risk early warning and prediction results.

[0138] Chemical process safety risk early warning information is constructed in edge computing nodes. The chemical process safety risk early warning information includes risk level identifier, chemical process safety risk early warning prediction result value, real-time process data timestamp corresponding to the chemical process safety risk early warning prediction result, equipment identifier and process unit identifier.

[0139] The chemical process safety risk warning information is sent to the field operation terminal through the communication interface of the edge computing node.

[0140] Example 1:

[0141] To verify the application effect of this invention in a real industrial scenario, the chemical process safety risk early warning system based on edge computing proposed in this invention was applied to the operation of a continuous reaction unit in a large chemical enterprise. The system performs real-time acquisition, processing, and early warning prediction of key operating parameters of the unit. This unit includes multiple key measuring points, such as the temperature, pressure, inlet and outlet flow rates, liquid level of the reactor, as well as the opening degree of key valves and pipeline vibration signals. Daily operating data changes frequently and exhibits strong nonlinear coupling relationships, making it impossible for traditional methods based on static thresholds to accurately identify complex risk scenarios.

[0142] The system is deployed on edge server nodes within the factory area, with a sampling period of 1 second, generating one time slice of data per minute, continuously collecting approximately 1.3 million real-time data samples over one month. First, the edge nodes perform missing data imputation and normalization on the raw data using a pre-defined data preprocessing workflow, and extract statistical features such as mean, range, and volatility, as well as constructed features such as trend change rate, based on a sliding window. Subsequently, ordered target statistical coding is performed on multiple categorical fields (such as equipment number and process stage) in the constructed structured feature data to generate target statistical data.

[0143] Target statistical data was input into the improved CatBoost model for training. During the model training phase, an adaptive symmetric tree construction mechanism was introduced. Candidate splitting structures were selected through a hierarchical split controller and a layered optional symmetry module. A local structure self-checking module performed rollback verification based on the loss value and sample label distribution to ensure that each splitting structure possessed global consistency and local robustness. The system ultimately trained 32 adaptive symmetric trees with an average tree depth of 7 layers. After model deployment, the edge lightweight model module performed pruning, quantization, and structural compression operations on the model, achieving a compression rate of 38% of the original model size, successfully deploying it to edge computing devices.

[0144] In actual operation and testing, by continuously collecting new process data and forming an incremental training dataset, the edge lightweight model module realizes local updates of leaf values ​​and tree weights once per hour. The system calls the updated model to perform inference prediction on the latest sampled data and generates risk level labels based on the set 5-level risk level thresholds. It successfully triggered early warnings in many early abnormal temperature rise and traffic fluctuation events, significantly improving risk response efficiency.

[0145] According to feedback from on-site operators, in the three months prior to deployment, there were a total of 13 false alarms and 2 missed alarms. After deploying the system of this invention, the number of false alarms decreased to 3, and the number of missed alarms was 0, with the average risk response time shortened by approximately 27%. Specific performance indicators are shown in Table 1.

[0146] Table 1. Performance Data Comparison Before and After System Deployment

[0147]

[0148] As shown in Table 1, after deployment at the edge, this invention can fully leverage the advantages of the model's lightweight design, high-precision prediction, and adaptive learning, demonstrating excellent safety risk identification capabilities and early warning response efficiency. This is significantly better than the performance of traditional static threshold and non-edge deployment modes, providing an effective solution for local intelligent safety risk control in the chemical industry.

[0149] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A chemical process safety risk early warning system based on edge computing, characterized in that, include: The edge computing module is used to receive real-time process data from the chemical plant, cache the real-time process data as time slice data according to a preset sampling period, perform data preprocessing and feature construction operations on the time slice data, and generate structured feature data. The ordered target statistics module is used to receive structured feature data, identify categorical feature fields in the structured feature data, perform ordered target statistical encoding on the categorical features, and generate target statistical data. The improved CatBoost model module is used to receive target statistical data, perform adaptive symmetric tree construction and adaptive symmetric tree inference based on the target statistical data, and generate early warning prediction results for chemical process safety risks. The improved CatBoost model module includes an adaptive symmetric tree module and an ordered gradient boosting framework. The adaptive symmetric tree module includes a hierarchical split controller, a hierarchical optional symmetric module, and a local structure self-checking module. The edge lightweight model module is used to perform pruning, quantization and structural compression on the trained improved CatBoost model to generate an edge deployment model version, and to perform incremental updates on the model parameters in the edge deployment model version based on continuously collected real-time process data. The risk assessment module is used to compare the early warning and prediction results of chemical process safety risks with preset risk level thresholds and generate early warning information for chemical process safety risks. The communication interface module is used to send early warning information on chemical process safety risks to the field operation terminal.

2. The chemical process safety risk early warning system based on edge computing according to claim 1, characterized in that, The modules are connected in the following way: Step 1: The edge computing node receives real-time process data and caches the real-time process data as time slice data according to the preset sampling period; it performs data preprocessing and feature construction operations on the time slice data to obtain structured feature data; Step 2: Input the structured feature data into the ordered target statistics module, perform ordered target statistical encoding on the categorical features in the structured feature data, and generate target statistical data. Step 3: Call the adaptive symmetric tree module in the improved CatBoost model to perform adaptive symmetric tree construction on the target statistical data; The hierarchical split controller determines the current set of nodes to be split and sends the set of nodes to be split to the hierarchical optional symmetric module to generate corresponding candidate symmetric splits and local split candidates of the nodes to be split; Based on preset judgment conditions, a pure symmetric splitting is selected to generate a splitting layer structure. This splitting layer structure is then sent to the local structure self-inspection module, which then performs the inspection on the splitting layer structure. If the test fails, a rollback operation is performed on this split layer structure; if the test passes, this split layer structure is confirmed as the valid structure of the current split layer. Step 4: Perform adaptive symmetric tree inference. Calculate the leaf values ​​of the tree structure confirmed by the local structure self-checking module to obtain the output of a single adaptive symmetric tree. In the ordered gradient boosting framework, perform a weighted summation of the outputs of all adaptive symmetric trees to obtain the chemical process safety risk early warning prediction results. Step 5: Deploy the improved CatBoost model on an edge computing node, perform pruning, quantization and structural compression on the trained improved CatBoost model through the edge lightweight model module, and perform incremental updates on the improved CatBoost model version on the edge computing node based on continuously collected real-time process data; Step 6: In the edge computing node, compare the prediction results of chemical process safety risk warning with the preset risk level threshold, generate the corresponding chemical process safety risk warning information, and send it to the field operation terminal.

3. The chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, The real-time process data includes temperature, pressure, flow rate, liquid level, valve opening, and equipment vibration signals; data preprocessing and feature construction operations are performed on the time slice data, including filling in missing data, normalizing numerical data, and generating statistical features and rate of change features based on a sliding window.

4. The chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, The improved CatBoost model includes an ordered target statistics module, an adaptive symmetric tree module, an ordered gradient boosting framework, and an edge lightweight model module.

5. A chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, Step two specifically involves: The ordered target statistics module receives structured feature data in the edge computing node and identifies categorical feature fields in the structured feature data; For each category feature, at least one sample sequence is generated by the ordered target statistics module, and samples in the structured feature data are read sequentially according to the order of the sample sequence. The ordered target statistical value is obtained by performing statistical operations on the supervised learning target values ​​of historical samples that are earlier than the current sample in the sample sequence and have the same categorical feature values. The statistical operation is to average the supervised learning target values. Based on the ordered target statistics, assign corresponding code values ​​to the categories in each categorical feature, and establish a mapping relationship between category values ​​and code values; By using mapping relationships, the categorical feature fields in the structured feature data are replaced with the corresponding coded values, while retaining the original numerical feature fields, to generate the target statistical data.

6. A chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, Step three specifically involves: Input the target statistical data into the adaptive symmetric tree module in the improved CatBoost model, and initialize the root node for each adaptive symmetric tree to be trained under the ordered gradient boosting framework; Perform adaptive symmetric tree construction: The hierarchical split controller determines the current set of nodes to be split at each splitting level of the current adaptive symmetric tree and sends the set of nodes to be split to the hierarchical optional symmetric module; The hierarchical optional symmetric module generates candidate symmetric splits for the current splitting layer and local splitting candidates for the nodes to be split based on the target statistical data of the samples contained in each node in the set of nodes to be split and the supervised learning target value. The layered optional symmetric module selects a pure symmetric split for the current split layer from among candidate symmetric splits and local split candidates according to preset judgment conditions, thus forming the structure of this split layer; The split layer structure is sent to the local structure self-inspection module, which then detects the split layer structure based on the loss value of the parent node set before splitting, the loss value of the child nodes after splitting, and the sample number and label distribution of the child nodes. If the test fails, the local structure self-test module performs a rollback operation on the current split layer structure, and the hierarchical split controller re-determines the set of nodes to be split in the current split layer. When the test is passed, the local structure self-test module confirms that the structure of this split layer is the valid structure of the current split layer, and uses the child nodes formed by the current split layer as the set of nodes to be split in the next split layer. Repeat the adaptive symmetric tree construction process until the current adaptive symmetric tree meets the preset splitting stopping condition, thus completing the structure construction of the current adaptive symmetric tree; And in the ordered gradient boosting framework, the adaptive symmetric tree construction process is performed sequentially for all adaptive symmetric trees.

7. A chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, The execution of adaptive symmetric tree reasoning specifically involves: Under the control of the ordered gradient boosting framework, the target statistical data is sequentially input into each adaptive symmetric tree that has completed the structure construction; For each adaptive symmetric tree, based on the splitting conditions of each splitting layer in the adaptive symmetric tree, a path determination is performed on each sample record in the target statistical data, and each sample record is assigned to the corresponding leaf node according to the splitting path of the sample record in the adaptive symmetric tree; For each leaf node in each adaptive symmetric tree, the leaf value of the leaf node is calculated based on the supervised learning target value of the training sample assigned to the leaf node and the update amount information determined by the ordered gradient boosting framework for the current iteration, and the output result of the adaptive symmetric tree is generated using the leaf value. The outputs of all adaptive symmetric trees are weighted and accumulated sequentially according to the preset tree weight parameters in the ordered gradient boosting framework to obtain the chemical process safety risk early warning prediction results for the target statistical data.

8. A chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, Step five specifically involves: After training the improved CatBoost model in the training environment, export the model parameters and adaptive symmetric tree structure of the improved CatBoost model, generate deployment model data, and distribute the deployment model data to the edge computing nodes; In the edge computing node, the deployment model data is loaded through the edge lightweight model module, and some adaptive symmetric trees, split layers and leaf nodes in the deployment model data are deleted according to the preset pruning rules to obtain the pruned model data. In the edge lightweight model module, the model parameters in the pruning model data are quantized, converting the floating-point split threshold, leaf value and tree weight parameters into a low-bit-width fixed-point numerical representation to generate quantized model data. In the edge lightweight model module, structural compression processing is performed on the quantized model data. The adaptive symmetric tree structure, split condition index and leaf value in the quantized model data are recombined and encoded according to the preset data structure to generate the edge deployment model version. An incremental training dataset is formed based on continuously collected real-time process data in the edge computing node. The leaf value parameters and tree weight parameters in the edge deployment model version are updated by the edge lightweight model module using the incremental training dataset, thus obtaining the updated edge deployment model version.

9. A chemical process safety risk early warning system based on edge computing according to claim 2, characterized in that, Step six specifically includes: In the edge computing node, the edge deployment model version is updated, and adaptive symmetric tree inference is performed on the target statistical data to obtain the early warning and prediction results of chemical process safety risks. In the edge computing node, a preset risk level threshold set is called, and the chemical process safety risk early warning prediction result is compared with each threshold in the preset risk level threshold set one by one to determine the risk level range in which the chemical process safety risk early warning prediction result is located. Based on the risk level range of the chemical process safety risk early warning and prediction results, a risk level label is generated for the chemical process safety risk early warning and prediction results. Chemical process safety risk early warning information is constructed in edge computing nodes. The chemical process safety risk early warning information includes risk level identifier, chemical process safety risk early warning prediction result value, real-time process data timestamp corresponding to the chemical process safety risk early warning prediction result, equipment identifier and process unit identifier. The chemical process safety risk warning information is sent to the field operation terminal through the communication interface of the edge computing node.

Images