Progressive multi-layer federated learning method and system for chemical process safety monitoring

Abstract

This invention relates to the field of intelligent monitoring technology in the chemical industry, and discloses a progressive multi-layer federated learning method and system for safety monitoring of chemical processes. The method includes: constructing a federated learning system with a multi-layer Internet of Things (IoT) and completing initial configuration; collecting chemical process safety parameters through edge monitoring nodes, and training local models in stages based on data quality, uploading the parameters of local models that meet performance requirements to the corresponding edge aggregation nodes; weighted aggregation by the edge aggregation nodes based on the contribution of each local model, uploading the parameters of the edge aggregation model that meet performance requirements to the global central node; the global central node dynamically aggregating the data based on differences in data distribution to obtain a global aggregation model; deploying the models at each layer to the respective layer nodes to carry out multi-level linked safety monitoring, and adaptively iteratively updating the models at each layer based on monitoring feedback. This invention can realize real-time detection, early warning, and overall control of various abnormal parameters in the production process.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology in the chemical industry, and in particular to a progressive multi-layer federated learning method and system for safety monitoring of chemical processes. Background Technology

[0002] In modern chemical production processes, system operation involves complex reaction kinetics and strongly coupled multivariate processes, spanning multiple key stages such as reaction, separation, purification, and storage and transportation. These processes are highly sensitive to and require strict control over core process parameters such as temperature, pressure, flow rate, gas concentration, liquid level, and pH value. These parameters exhibit significant nonlinear coupling relationships and dynamic correlations. Against this backdrop, even minor deviations in local parameters can disrupt the original thermodynamic or kinetic equilibrium, triggering chain reactions and leading to serious safety accidents such as explosions, leaks of toxic and harmful gases, fires, or equipment corrosion. This not only causes significant economic losses but also poses a significant threat to human life and the ecological environment. Furthermore, from a data perspective, chemical process monitoring data exhibits significant characteristics such as multi-source heterogeneity, strong temporal sequence, and inconsistent distribution. On the one hand, data collected from different production units, operating conditions, and monitoring points show significant differences in statistical distribution, characteristic space, and dimensional scale. On the other hand, due to differences in process routes, equipment types, and operating strategies among chemical enterprises, their data distribution exhibits significant cross-regional, non-independent, and identically distributed characteristics. Meanwhile, monitoring data is usually stored in various production units or within enterprises, making it difficult to share directly in a centralized manner.

[0003] Under these conditions, traditional centralized modeling methods face two key bottlenecks: First, aggregating scattered data to a central server for unified modeling not only increases the risk of data leakage and tampering but also fails to meet the confidentiality requirements of chemical enterprises for process data. Second, ignoring differences in data distribution during unified modeling can lead to insufficient generalization ability of the model across different production scenarios, making it difficult to achieve stable and reliable anomaly detection and safety early warning. Therefore, from a system modeling perspective, there is an urgent need for a technical solution that can achieve collaborative modeling in a distributed data environment while simultaneously ensuring data privacy protection and heterogeneous adaptability. In this context, federated learning, as a distributed learning paradigm that does not require sharing of raw data and achieves multi-node collaborative optimization only through model parameter interaction, provides a new technical path to solve the above problems. By training models locally at each data source and then aggregating them at the upper layer, multi-source monitoring information can be effectively integrated while ensuring data security, thereby improving the overall performance of chemical process safety monitoring models.

[0004] However, existing technologies still fall short of meeting the complex needs of chemical process monitoring in practical applications, mainly in the following aspects: First, centralized modeling methods pose significant data security risks. Chemical production data typically contains highly sensitive information such as process formulations, equipment operating status, and control strategies. Traditional methods require uploading raw datasets to a central server for unified modeling. This process not only increases the risk of data leakage and tampering, but also makes it difficult to achieve cross-regional or cross-enterprise data sharing and collaborative modeling in practical applications due to enterprises' confidentiality requirements for core data.

[0005] Secondly, the monitoring data exhibits a significant heterogeneous distribution. Data collected from different equipment types, operating conditions, and monitoring locations show significant differences in feature space, statistical distribution, and dimensional scale. Traditional methods are usually modeled based on the assumption of identical distribution, which limits the generalization ability of the model when applied across different scenarios and easily leads to false alarms or false negatives.

[0006] Secondly, sensor data quality issues hinder the improvement of monitoring performance. Affected by factors such as equipment aging, environmental disturbances, and electromagnetic interference, the actual collected data generally suffers from problems such as noise pollution, missing data, and abnormal fluctuations. If directly used for model training, it will significantly reduce model stability and prediction accuracy, making it difficult to meet the high reliability requirements of safety monitoring.

[0007] Finally, existing federated learning methods are not well adapted to chemical engineering scenarios. Most existing methods adopt a single-layer federated architecture of "terminal-center" without fully considering the hierarchical structure of "device-edge-cloud" in the industrial IoT environment. At the same time, they lack a hierarchical and progressive optimization mechanism for heterogeneous data and do not effectively combine data quality assessment and preprocessing strategies, making it difficult to ensure both real-time performance and model accuracy and robustness. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies that combine chemical process safety monitoring with federated learning, such as architectural mismatch, low model accuracy, and lack of iterative capabilities, this invention proposes a progressive multi-layer federated learning method and system for chemical process safety monitoring. Through hierarchical architecture design, quality adaptive local training, differentiated aggregation strategy, and closed-loop iterative mechanism, it can achieve accurate monitoring of chemical process safety, privacy protection, and continuous model optimization, providing reliable technical support for the safety management of chemical enterprises.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A progressive multi-layer federated learning method for safety monitoring of chemical processes includes: Construct a federated learning system with a multi-layered Internet of Things and complete the initial configuration. The federated learning system includes edge monitoring nodes, edge aggregation nodes and a global center node that are connected in sequence. Chemical process safety parameters are collected through edge monitoring nodes, and local models are trained in stages based on data quality. The parameters of the local models that meet the performance requirements are then uploaded to the corresponding edge aggregation nodes. Edge aggregation nodes perform weighted aggregation based on the contribution of each local model, and upload the edge aggregation model parameters that meet the performance requirements to the global central node. The global central node dynamically aggregates the data based on the differences in data distribution to obtain the global aggregation model. The models at each layer are deployed to the nodes at each layer to carry out multi-level linkage security monitoring, and the models at each layer are adaptively iterated and updated based on the monitoring feedback.

[0010] Furthermore, the construction of a federated learning system with a multi-layered Internet of Things and the completion of initial configuration include: A three-layer architecture is established, consisting of edge monitoring nodes, edge aggregation nodes, and global center nodes. The edge monitoring nodes correspond to various monitoring points in the chemical industry, the edge aggregation nodes correspond to chemical workshops or regional control centers, and the global center nodes correspond to the overall control center of the chemical enterprise. Configure a basic model adapted to chemical safety monitoring for each layer node, initialize model weights and set relevant training hyperparameters, and configure encrypted communication protocols between layers; complete connectivity testing and adaptation debugging for each layer node to ensure normal data acquisition, parameter interaction and unified management functions.

[0011] Furthermore, the collection of chemical process safety parameters through edge monitoring nodes includes: Edge monitoring nodes collect key safety parameters in chemical production in real time, including temperature, pressure, and medium concentration; the collected raw data is preprocessed, including outlier removal, missing value imputation, and data standardization, to obtain preprocessed monitoring data.

[0012] Furthermore, prior to the phased training of the local model based on data quality, the following steps are included: Construct a data quality assessment index system, including data completeness, data accuracy, data timeliness, and data consistency; calculate the comprehensive data quality score using a weighted summation method; and classify the data into high-quality data, medium-quality data, and low-quality data based on the comprehensive score.

[0013] Furthermore, the staged training of the local model based on data quality includes: The adaptive learning training is performed in three stages. In the first stage, high-quality data is used to train until the loss converges. In the second stage, medium-quality data is added and the learning rate is reduced to fine-tune the model. In the third stage, low-quality data is introduced and the learning rate is further reduced to train until the performance is stable. After training, the local model performance is verified by reserving a test set. If the performance meets the standard, the local model parameters are uploaded. If the performance does not meet the standard, the model is retrained.

[0014] Furthermore, the edge aggregation nodes perform weighted aggregation based on the contribution of each local model, including: The edge aggregation node receives the local model parameters and performance indicators of all edge monitoring nodes within its jurisdiction; combining the local model performance, data quality, and node importance, it uses a game theory algorithm to calculate the contribution of each local model and normalizes it to obtain the contribution weight. The local model parameters are aggregated using a federated average algorithm based on contribution weights, and a momentum term is introduced to smooth parameter updates and avoid oscillations. The performance of the edge aggregation model is verified. If the performance meets the criteria, the parameters are uploaded. If the performance does not meet the criteria, the weights are adjusted and the aggregation is repeated.

[0015] Furthermore, the global central node dynamically aggregates data based on distribution differences to obtain a global aggregation model, including: The global central node receives model parameters from each edge aggregation node and quantifies the heterogeneity of dataset distribution in each edge layer; attention weights are introduced to adapt each edge aggregation model based on the distribution differences. By combining attention weights with knowledge of chemical process mechanisms, the parameters of each edge model are aggregated and fine-tuned to obtain a global aggregated model. The overall performance of the global aggregated model is verified, including the early warning response speed index. If the index is met, the model is deployed; if the index is not met, the aggregation strategy is adjusted and the model is retrained.

[0016] Furthermore, the deployment of each layer model to each layer node to carry out multi-level linkage security monitoring includes: The global central node distributes global aggregation model parameters to edge aggregation nodes. The edge aggregation nodes merge their own aggregation models with the global aggregation model to generate a collaborative model and distribute it to the edge monitoring nodes. The models at each layer are deployed accordingly to achieve local early warning, regional verification, and global control. Edge monitoring nodes combine real-time parameters and models to calculate security risk values ​​and trigger early warnings according to risk levels; edge aggregation nodes and global central nodes sequentially review and assess the risks, issue emergency commands, and record monitoring, early warning, and handling-related logs.

[0017] Furthermore, the adaptive iterative update of each layer of the model based on monitoring feedback includes: Each layer of nodes synchronously collects feedback information, which includes model performance, monitoring data, and application effects; the feedback information is analyzed to identify model problems, which include changes in data quality, distribution shifts, parameter aging, and process adjustments. For different model problems, corresponding strategies are adopted to adjust the training ratio, attention coefficient, model parameters or hyperparameters to complete model iteration; the performance of the updated model is verified, and the optimization effect is evaluated by the performance improvement rate. If it is effective, the original model is replaced; if it is ineffective, the strategy is adjusted and optimization is repeated.

[0018] A progressive multi-layer federated learning system for safety monitoring of chemical processes includes edge monitoring nodes, edge aggregation nodes, and a global central node that are connected in sequence. The edge monitoring node is configured to collect safety parameters of chemical processes, train a local model in stages based on data quality, upload the local model parameters that meet the performance requirements to the corresponding edge aggregation node, and perform local real-time safety monitoring and early warning. The edge aggregation node is configured to perform weighted aggregation based on the contribution of each local model, upload the edge aggregation model parameters that meet the performance requirements to the global central node, and perform regional-level security monitoring, review and coordination. The global central node is configured to receive edge aggregation model parameters uploaded by each edge aggregation node, analyze the differences in data distribution at each edge layer and dynamically aggregate the data based on the distribution differences to obtain a global aggregation model, verify the performance of the global model and then send the model parameters to each edge aggregation node, while simultaneously executing global security control and issuing emergency commands, and adaptively iterating and updating the model based on the monitoring information fed back by each layer node.

[0019] The beneficial effects of this invention are as follows: 1. This invention constructs a hierarchical IoT federated learning architecture consisting of "edge monitoring nodes - edge aggregation nodes - global central nodes", which is precisely adapted to the needs of distributed monitoring of chemical processes. It achieves an organic combination of data privacy protection and distributed training, avoids the risk of data leakage in centralized training, and reduces the pressure of data transmission.

[0020] 2. This invention employs progressive local training based on adaptive curriculum learning according to data quality. It fully considers the quality differences of chemical monitoring data, conducts training in stages, effectively improves the accuracy and generalization ability of the local model, and solves the problem of insufficient model performance caused by ignoring data quality in existing methods.

[0021] 3. This invention performs weighted aggregation based on game theory contribution evaluation at the edge layer and dynamic collaborative aggregation based on data distribution differences at the global layer. It takes into account both the contribution differences of each local model and the heterogeneity of data distribution, thereby improving the adaptability and accuracy of the global model and enabling it to better adapt to the complex scenarios of chemical processes.

[0022] 4. This invention constructs a complete model deployment, multi-level linkage monitoring and adaptive iterative optimization mechanism, realizing a closed loop of the entire process from model training, aggregation, deployment to optimization. It can continuously adapt to the dynamic changes of chemical processes, maintain efficient and accurate safety monitoring capabilities in the long term, and provide reliable technical support for the safety management and control of chemical processes. Attached Figure Description

[0023] Figure 1This is a flowchart of a progressive multi-layer federated learning method for safety monitoring of chemical processes, as described in Embodiment 1 of the present invention.

[0024] Figure 2 This is a schematic diagram of a progressive multi-layer federated learning method for safety monitoring of chemical processes, as described in Embodiment 1 of the present invention. Detailed Implementation

[0025] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] Example 1 See Figure 1 and Figure 2 This embodiment provides a progressive multi-layer federated learning method and system for chemical process safety monitoring. The multi-layer federated learning system includes a terminal perception layer, an edge aggregation layer, and a global server layer. Secure data and model parameters are transmitted between these layers via an encrypted communication protocol. The multi-layer federated learning method includes the following steps: S1. Construct a hierarchical IoT federated learning system and complete the initial configuration; S2. Perform progressive local model training with adaptive curriculum learning based on data quality; S3. Perform weighted aggregation of the local model at the edge layer based on contribution evaluation; S4. Dynamically and collaboratively aggregate the model at the global level, taking into account differences in data distribution. S5. Deploy the model to nodes at each layer and carry out multi-level linkage security monitoring and decision-making; S6. Adaptively iteratively update and continuously optimize the model based on monitoring feedback.

[0027] Furthermore, step S1 of this embodiment includes the following sub-steps: S11. Construct a three-layer IoT federated learning architecture consisting of "edge monitoring nodes - edge aggregation nodes - global center nodes". The edge monitoring nodes correspond to various monitoring points in chemical production (such as reactors, pipelines, storage tanks, etc.) and are used to collect key safety parameters of the chemical process and perform local model training. The edge aggregation nodes correspond to chemical production workshops or regional control centers and are used to aggregate the local models of edge monitoring nodes within their jurisdiction. The global center nodes correspond to the overall control center of the chemical enterprise and are used to perform global collaborative aggregation of the aggregated models of each edge aggregation node.

[0028] S12. Initialize the model parameters of each layer node. Based on the requirements of chemical process safety monitoring, select an appropriate basic model, configure initial model weights for edge monitoring nodes, edge aggregation nodes, and global center nodes, set hyperparameters such as learning rate, number of iterations, and convergence threshold for model training, and configure encryption protocols for data transmission to ensure privacy and security during data transmission.

[0029] S13. Complete the connectivity test and adaptability debugging of each layer node to ensure that the edge monitoring node can collect data normally and communicate with the edge aggregation node, the edge aggregation node can receive local model parameters normally and interact with the global center node, and the global center node can realize unified management and control of each edge aggregation node to ensure the stable operation of the entire hierarchical IoT federated learning system.

[0030] Furthermore, step S2 of this embodiment includes the following sub-steps: S21. The edge monitoring node collects key safety parameters of the chemical process in real time through the data acquisition module. The collected raw data is preprocessed, including outlier removal (using the 3σ criterion to remove data that exceeds the normal range), missing value filling (using linear interpolation to fill missing data), and data standardization (normalizing the data to the [0,1] interval) to obtain preprocessed monitoring data.

[0031] S22. Conduct a quality assessment of the preprocessed monitoring data, construct a data quality assessment index system, including data completeness, data accuracy, data timeliness, and data consistency, and calculate the data quality score using a weighted summation method. The specific formula is as follows:

[0032] in, The overall score for data quality is 100 points. For the first The weight of each evaluation indicator ( Preferably, , , , ); Corresponding data integrity score, Corresponding data accuracy score, Corresponding data timeliness score, Corresponding data consistency scores are assigned, and the specific scoring methods for each indicator are tailored to the chemical monitoring scenario; based on the scoring results, the data is categorized into high-quality data (…). Medium quality data Low-quality data ),in This indicates that the data quality meets the standards and can be directly used for model training. Further verification is required before use. It should be used with caution and the proportion of training should be controlled.

[0033] S23. Implement progressive local training for adaptive curriculum learning based on data quality scores, and formulate a three-stage training curriculum: The first stage uses high-quality data for training, with a learning rate of... ( The first stage uses the learning rate to control the model parameter update speed (with a value range adapted to the chemical monitoring data training scenario), and training continues until the loss value converges. The second stage involves training with medium-quality data, adjusting the learning rate accordingly. ( The second stage learning rate, lower than the first stage, is used to fine-tune model parameters, further optimizing the model; the third stage introduces low-quality data for training, reducing the learning rate to [a lower value]. ( This is the third-stage learning rate, the lowest learning rate, used to stabilize model performance. Training continues until the loss value stabilizes, improving the model's generalization ability.

[0034] S24. After local training is completed, the edge monitoring node performs performance verification on the local model using a reserved test dataset (20% of the preprocessed data is randomly selected as the test set). If the performance indicators meet the preset requirements (preferably, prediction accuracy ≥ 95%, warning recall ≥ 90%, false positive rate ≤ 5%), the local model parameters are uploaded to the corresponding edge aggregation node; otherwise, the process returns to S23 to retrain locally until the performance requirements are met.

[0035] Furthermore, step S3 of this embodiment includes the following sub-steps: S31. The edge aggregation node receives local model parameters and corresponding performance indicators uploaded by all edge monitoring nodes within its jurisdiction, and evaluates the contribution of each local model through the contribution evaluation module.

[0036] S32. The contribution of each local model is calculated using the Shapley value method in game theory. A contribution calculation model is constructed by combining the performance of the local models, data quality scores, and the importance of monitoring nodes. After weight correction and normalization, the contribution weights of each local model are obtained. The higher the contribution, the greater the weight, ensuring that the local model corresponding to high-performance, high-quality data plays a greater role in the aggregation process.

[0037] S33. Based on the contribution weight, the parameters of each local model are weighted and aggregated, using a federated average algorithm and introducing a momentum term. Smoothing parameters avoids parameter oscillations during aggregation and ensures the stability of the aggregation model parameters. The core aggregation formula is as follows:

[0038] in, For the first Parameters of the edge layer model after wheel aggregation For the first The edge layer model parameters after round aggregation (the initial values ​​are the initial model weights set in step S13). This is the momentum term (used to smooth parameter updates and avoid parameter oscillations during aggregation). The number of edge monitoring nodes within the jurisdiction. For the first Contribution weights of each local model ( ), For the first The first edge monitoring node The parameters of the model after local training.

[0039] S34. Validate the performance of the edge layer aggregation model using the test dataset reserved by the edge aggregation node. If the performance indicators meet the preset requirements (preferably, prediction accuracy ≥ 96%, warning recall ≥ 92%, false alarm rate ≤ 4%), then upload the edge layer aggregation model parameters to the global central node; if not, adjust the contribution weights and re-aggregate until the performance requirements are met.

[0040] Furthermore, step S4 of this embodiment includes the following sub-steps: S41. The global central node receives relevant data uploaded by each edge aggregation node, analyzes the distribution differences of datasets in each edge layer through the dynamic adaptation module, and uses KL divergence to quantify the heterogeneity of data distribution, with a preset KL divergence threshold. ( Threshold for determining distribution differences This indicates small distribution differences. (This indicates a large distribution difference), and at the same time, the variance of each edge layer dataset is calculated to help quantify the distribution difference (the larger the variance, the more dispersed the data distribution and the stronger the heterogeneity).

[0041] S42. Adjust the global aggregation strategy based on data distribution differences, and introduce attention weights. To adapt to edge layer aggregation models with varying data distributions, higher attention weights are assigned to model parameters with significant distribution differences, reducing the impact of data heterogeneity on the global model. The core formula for calculating attention weights is as follows:

[0042] in, This is the attention coefficient (used to adjust the effect of distribution differences on attention weights; the larger the value, the more significant the effect of distribution differences on the weights). The average distribution of all edge layer datasets (obtained by weighted averaging of the distributions of each edge layer dataset). For the first The KL divergence between the distribution of each edge layer dataset and the average distribution (the larger the divergence value, the greater the distribution difference). For the first Attention weights of edge layer aggregation models ( ), This represents the number of edge aggregation nodes.

[0043] S43. Perform global-level dynamic collaborative aggregation, combining attention weights and knowledge of chemical process mechanisms to aggregate and fine-tune the aggregation model parameters of each edge layer, ensuring that the global model fits the actual mechanism of chemical production. The core aggregation formula is as follows:

[0044] in, Based on the parameters of the global aggregation model, For the first Attention weights of an edge layer aggregation model The number of edge aggregation nodes, For the first The edge layer aggregation model parameters are uploaded by each edge aggregation node; after fine-tuning, the global aggregation model parameters are obtained. .

[0045] S44. Validate the performance of the global aggregation model by using a global test dataset. In addition to the core performance indicators, add the early warning response speed indicator. If the preset requirements are met (preferably, prediction accuracy ≥ 97%, early warning recall ≥ 93%, false alarm rate ≤ 3%, and early warning response speed ≤ 100ms), then proceed to the model deployment stage; if not, adjust the aggregation strategy and perform global aggregation again until the performance requirements are met.

[0046] Furthermore, step S5 of this embodiment includes the following sub-steps: S51. The global central node sends the global aggregation model parameters to each edge aggregation node. Each edge aggregation node merges the edge layer aggregation model with the global model to generate an edge collaborative model, which is then sent to each edge monitoring node to achieve model synchronization of each layer node.

[0047] S52. Deploy each layer of the model to the corresponding node to achieve layered monitoring functions such as local real-time early warning, regional overall review, and global control.

[0048] S53. Conduct multi-level coordinated safety monitoring, with edge monitoring nodes collecting key parameters of the chemical process in real time, and calculating safety risk values ​​through local models and edge collaborative models. The core formula is as follows:

[0049] in, For the first Security risk value of each edge monitoring node (the higher the value, the higher the security risk). This refers to the quantity of key safety parameters in the chemical industry (such as temperature, pressure, and concentration of toxic gases). For the first Risk weights for key safety parameters (set according to the safety importance of the parameter, such as toxic gas concentration having a higher weight than liquid level). For the first The first edge monitoring node collected the data. Real-time values ​​of key security parameters, For the first Safety standard values ​​for key safety parameters (set according to chemical production safety regulations). For the first The relative deviation of each parameter is used to quantify the degree to which the parameter deviates from the standard value; based on the risk value, the corresponding level of early warning is triggered, and the edge aggregation node and the global central node are reviewed and assessed in turn, and emergency response instructions are issued to achieve multi-level linkage of "local early warning - regional review - global control".

[0050] S54. Establish a monitoring log to record monitoring data, early warning information, model inference results, emergency response instructions and execution status of each node in real time, so as to provide data support for subsequent model iteration and optimization.

[0051] Furthermore, step S6 of this embodiment includes the following sub-steps: S61. Each layer of nodes synchronously collects monitoring feedback information, including feedback on model performance, monitoring data, and practical applications.

[0052] S62. Analyze the feedback information to identify problems in the model and their causes, such as changes in data quality and data distribution shifts.

[0053] S63. Perform adaptive iterative updates based on the problem analysis results, adopting corresponding iterative strategies for different problem types. The core adjustment methods are as follows: 1) Changes in data quality: Adjusting the local training ratio ;in, To adjust the training ratio of data at each quality level ( (Corresponding to high, medium, and low data quality levels) The initial training ratio for data at each quality level. This represents the average quality score for the current data across all quality levels. This represents the maximum value of the average score across all current quality levels.

[0054] 2) Data distribution offset: Adjusting the attention coefficient ,in, The adjusted attention coefficient. The attention coefficient before adjustment (initial value is 5). This represents the maximum value of the KL divergence of each edge layer dataset. This represents the maximum value of the KL divergence of each edge layer dataset in the previous round.

[0055] 3) Model parameter aging: Introduce parameter decay coefficients to update model parameters. ( );in, For the updated number One local model parameter, For the aging One local model parameter, This is the parameter attenuation coefficient (which controls the degree of attenuation of the old parameter). The first result obtained from the new training Local model parameters.

[0056] 4) Process Adjustment: Adjust model hyperparameters, such as the learning rate. ;in, The learning rate after process adjustment. This is the initial learning rate before process adjustments. This is the process adjustment coefficient (set according to the magnitude of process changes, with a value range of 0.1~0.3; the greater the process change, the higher the coefficient). The larger the value.

[0057] S64. After iterative updates, the performance of the updated model is verified using the performance improvement rate. The core formula for evaluating the effectiveness of iterations is as follows:

[0058] in, The performance improvement rate after model iteration. The overall performance score of the model after iterative updates (obtained by weighting the model's core performance indicators). The overall performance score of the model before the iterative update. This indicates that the iterative optimization is effective and the original model can be replaced; if If each individual performance indicator meets the preset requirements, the original model is replaced; if not, the iteration strategy is adjusted and optimization is repeated to form a continuous iteration mechanism.

[0059] Example 2 This embodiment is based on embodiment 1: This embodiment provides a progressive multi-layer federated learning method for safety monitoring of chemical processes, which will now be described in detail in the context of the entire lithium extraction process from salt lakes.

[0060] It should be noted that this embodiment involves all key process units and utility units, including old brine pretreatment, salt fields and brine transportation system, adsorption lithium extraction unit, nanofiltration / reverse osmosis / electrodialysis membrane refining system, MVR evaporation and concentration system, lithium precipitation reaction system, lithium carbonate / lithium chloride synthesis system, hazardous chemical storage area, circulating water station, air compressor station, substation, and plant-wide pipe racks. Monitoring targets include all safety-related parameters across dimensions, such as temperature, pressure, liquid level, flow rate, pH, conductivity, lithium concentration, calcium and magnesium ion concentration, corrosive gas concentration, motor temperature, pump vibration, valve on / off status, membrane inlet and outlet pressure difference, current and voltage, pipeline corrosion and thinning, media leakage, and gas monitoring.

[0061] This embodiment constructs a three-layer federated learning architecture of "terminal perception layer - edge aggregation layer - global server layer" to achieve real-time anomaly detection, risk classification and early warning, and multi-level linkage response throughout the entire production process while strictly protecting the privacy of process data. The specific implementation steps are as follows: Step S1: Build a hierarchical IoT federated learning system and complete the initial configuration.

[0062] 1) This embodiment constructs a three-layer IoT federated learning architecture consisting of "edge monitoring nodes - edge aggregation nodes - global central nodes". Edge monitoring nodes correspond to 168 on-site monitoring points within the salt lake lithium salt smart factory, deployed in key locations such as adsorption tower groups, elution towers, membrane treatment workshops, electrodialysis membrane stacks, MVR evaporators, steam compressors, lithium precipitation reactors, transfer tanks, brine pumping stations, acid and alkali storage tank areas, hazardous chemical warehouses, high-voltage power distribution rooms, pipe gallery monitoring points, and wastewater treatment areas. They are used to collect key safety parameters of the chemical process and perform local model training. Edge aggregation nodes are divided into 7 regional aggregation nodes according to the process flow, corresponding to the salt field brine transportation area, adsorption lithium extraction area, membrane refining and electrodialysis area, MVR evaporation and concentration area, lithium precipitation and lithium carbonate synthesis area, lithium chloride production area, and utilities and storage area. They are used to aggregate the local models of the edge monitoring nodes within their respective areas. The global central node is deployed in the central control center of the smart factory, used for global collaborative aggregation and safety management of the aggregated models of each edge aggregation node.

[0063] 2) Initialize the model parameters for each layer node. Based on the requirements of chemical process safety monitoring, an LSTM+Transformer hybrid temporal anomaly detection model is selected as the base model. The initial model weights for each layer node are uniformly assigned a value of 0.01, the initial learning rate for local training is set to 0.01, the number of local iterations per round is set to 50, and the convergence threshold of the loss function is set to 0.001. The number of global aggregation iterations is set to 20, the edge aggregation momentum term is set to 0.9, and the initial value of the global attention coefficient is set to 5. Simultaneously, the AES-256 encryption protocol is configured to ensure the security of model parameter transmission; no raw data or process feature data is transmitted throughout the process. According to the safety specifications for lithium extraction from salt lakes, the following safety standard values ​​are preset for key parameters: adsorption tower working pressure ≤ 1.2 MPa, membrane system inlet and outlet pressure difference ≤ 0.85 MPa, MVR evaporation temperature ≤ 110℃, lithium precipitation reactor pH controlled at 10.0±0.5, corrosive and harmful gas concentration ≤ 10 ppm, storage tank liquid level ≤ 85%, motor temperature ≤ 75℃, pump vibration intensity ≤ 4.5 mm / s, and electrodialysis membrane stack current ≤ 150 A.

[0064] 3) Complete connectivity testing and compatibility debugging for each layer of nodes to ensure accurate sensor data acquisition from edge monitoring nodes, with no drift or frequent packet loss; communication latency between edge monitoring nodes and edge aggregation nodes, and between edge aggregation nodes and the global center node, is consistently ≤50ms; model loading for each node is normal, computational resource usage is reasonable, and there are no errors or crashes; the encrypted communication module is effective, and parameter transmission cannot be eavesdropped on, tampered with, or replayed. The system as a whole operates stably and is ready to initiate progressive multi-layer federated learning training.

[0065] Step S2: Build a hierarchical IoT federated learning system and complete the initial configuration.

[0066] 1) Edge monitoring nodes collect key safety parameters of the entire lithium salt production process in real time at a frequency of 1 time / second, including adsorption tower liquid level, adsorption pressure, elution flow rate, elution temperature; nanofiltration / reverse osmosis system feed water flow rate, product water flow rate, membrane inlet and outlet pressure difference, membrane flux, and recovery rate; electrodialysis system membrane stack voltage, electrode chamber current, concentrate and dilute water flow rate, and temperature difference; MVR system evaporation temperature, vacuum degree, steam flow rate, and compressor current; lithium precipitation reactor stirring speed, reaction temperature, pH value, and feed flow rate; pump bearing temperature, vibration amplitude, and inlet and outlet pressure; pipeline pressure, corrosive gas concentration, leak monitoring value, and ambient temperature and humidity, etc.; preprocessing is performed on the raw data: the 3σ criterion is used to remove abnormal values ​​such as spike pulses and sudden jumps caused by sensor failure and electromagnetic interference; linear interpolation is used to fill in missing data caused by short-term communication interruption and sensor sleep; and Min-Max normalization is used to uniformly map parameters of different dimensions and magnitudes to the [0,1] interval to eliminate the influence of dimensional differences on model training.

[0067] 2) Conduct a multi-dimensional data quality assessment on the preprocessed monitoring data, construct an indicator system that includes data integrity, data accuracy, data timeliness, and data consistency, and calculate the comprehensive score according to the formula of this invention: Taking the core monitoring nodes in the membrane treatment workshop as an example, its data integrity... ,accuracy Timeliness ,consistency Calculated This data was deemed high-quality. Based on the scoring results, the data was divided into three levels: It is divided into high-quality data (core points such as membrane stack, evaporator, reactor, online analyzer, etc.); Data is categorized into medium-quality data (pipe gallery, outdoor pump room, intermittently operating equipment). The data is divided into low-quality data (open-air salt field locations, weak signal areas, and old sensor locations).

[0068] 3) Implement a three-stage adaptive learning process based on data quality levels to achieve progressive local model training. The first stage uses only high-quality data ( Conduct basic training, learning rate The training continues until the loss function changes less than the threshold for 35 consecutive rounds and fully converges; in the second stage, medium-quality data is added to the training set. The learning rate was lowered to The model underwent 25 rounds of fine-tuning; in the third stage, a small amount of low-quality data was introduced for generalization enhancement, and the learning rate was further reduced to... The model was trained for 40 rounds to improve its robustness in complex environments, weak signals, and interference conditions.

[0069] 4) After local training is completed, a 20% test set randomly selected from the preprocessed data is used for independent performance verification. The evaluation indicators are strictly set according to the invention requirements: prediction accuracy ≥ 95%, anomaly warning recall rate ≥ 90%, and false positive rate ≤ 5%. Model parameters that meet the indicator requirements are encrypted and uploaded to their respective edge aggregation nodes; nodes that do not meet the requirements are automatically returned to S23 to readjust the learning rate, data ratio, and training epochs until the performance requirements are met.

[0070] Step S3: Build a hierarchical IoT federated learning system and complete the initial configuration.

[0071] 1) The edge aggregation node receives encrypted model parameters, performance indicators and data quality scores uploaded by all edge monitoring nodes within its jurisdiction. It first performs parameter integrity verification, format verification and decryption verification, and removes invalid models with abnormal transmission, missing parameters and incorrect format to ensure that all models participating in the aggregation are valid and usable.

[0072] 2) The Shapley value method was used to calculate the contribution of each local model, considering three key factors: the model's prediction accuracy on the test set, the overall data quality score, and the importance of the monitoring points in process safety. Points for core equipment (membrane stack, reactor, evaporator, hazardous chemical storage) received higher weights, while those for general pipe racks and auxiliary equipment received relatively lower weights. After weight adjustment and normalization, the contribution weights of each local model were obtained. This allows models with better performance, higher data quality, and greater security to occupy a larger proportion in edge aggregation.

[0073] 3) Perform weighted aggregation based on contribution weights, and introduce a momentum term. To smooth the parameter update process and avoid parameter oscillations and divergence, the core aggregation formula given in this invention is followed: Perform multiple rounds of iterative aggregation until the parameters of the edge layer model tend to stabilize and the loss function is reduced to a minimum.

[0074] 4) After edge aggregation is completed, performance is verified using a regional test set maintained independently by the edge nodes. The requirements are: prediction accuracy ≥ 96%, warning recall ≥ 92%, and false positive rate ≤ 4%. If the criteria are met, the edge aggregation model is encrypted and uploaded to the global central node; if not, the contribution weights are readjusted, the aggregation rounds are optimized, or a local model is supplemented and retrained, and aggregation is re-executed until the criteria are met.

[0075] Step S4: Build a hierarchical IoT federated learning system and complete the initial configuration.

[0076] 1) The global central node receives edge models, data distribution statistics, and regional performance reports uploaded by all 7 edge aggregation nodes. It uses KL divergence to quantify the data distribution differences between different edge regions and sets a distribution difference threshold. A KL divergence value less than 0.2 indicates similar data distribution, while a value greater than or equal to 0.2 indicates significant distribution differences. The variance of each region's dataset is also calculated; a larger variance indicates more dispersed data and stronger heterogeneity. In the brine lake lithium extraction scenario, the data distribution in the adsorption lithium extraction region and the membrane refining region is similar, while the membrane system, evaporation system, and lithium deposition system exhibit significant distributional heterogeneity due to differences in process mechanisms.

[0077] 2) Calculate the attention weights according to the formula of this invention based on the distribution differences:

[0078] in Higher weights are assigned to marginal models with greater distribution differences, reducing the impact of heterogeneity on the global model.

[0079] 3) Substitute into the global aggregation formula The basic global model is obtained, and then the model parameters are fine-tuned by combining the knowledge of the specific process mechanism of lithium extraction from salt lakes. This enhances the ability to extract features of high-risk modes such as abnormal membrane pressure difference, electrodialysis overflow, evaporation temperature drift, lithium precipitation pH deviation, corrosive leakage, and over-temperature and over-pressure. This makes the global model not only data-driven but also more in line with the constraints of chemical mechanism.

[0080] 4) Use a plant-wide integrated test set for global model validation. In addition to accuracy, recall, and false alarm rate, add an early warning response speed indicator, requiring the following: prediction accuracy ≥ 97%, early warning recall ≥ 93%, false alarm rate ≤ 3%, and early warning response speed ≤ 100ms. If all indicators meet the standards, proceed to the model deployment phase; if not, return to adjust the attention coefficient, aggregation strategy, or mechanistic constraint strength, and re-execute global aggregation until the high-reliability monitoring requirements of the industrial site are met.

[0081] Step S5: Deploy the model to nodes at each layer and carry out multi-level linkage security monitoring and decision-making.

[0082] 1) The global central node will encrypt and distribute the final global aggregation model to all 7 edge aggregation nodes. Each edge node will fuse the features of the global model and the local edge aggregation model to generate an edge collaborative model that is more in line with the regional working conditions. Then, it will distribute the model to all terminal sensing nodes under its jurisdiction to complete the synchronization of the global model and the unification of parameters, ensuring that the whole plant implements a unified safety judgment standard and early warning logic.

[0083] 2) Deploy the global model, edge collaboration model, and local model to the corresponding node hardware environment. The terminal node is responsible for local real-time inference and on-site early warning, the edge node is responsible for regional risk review and collaborative judgment, and the global node is responsible for monitoring the security situation of the entire plant, summarizing and displaying risks, and coordinating emergency commands, forming a three-level security monitoring architecture of "real-time terminal early warning - edge area review - unified global control".

[0084] 3) Initiate multi-level linkage security monitoring, and the terminal node calculates the security risk value in real time according to the following formula. :

[0085] in For the number of key safety parameters, For the first Risk weights for each parameter This is a real-time value. This represents the safety standard value. A higher risk value indicates a more significant deviation of the parameter from the safe range and a higher degree of danger. When The system immediately triggers local audible and visual warnings, while simultaneously uploading data and warning information to edge nodes. Edge nodes perform comprehensive regional risk calculations and validity checks to rule out sensor malfunctions and false triggers. Global nodes aggregate information from across the entire region, conduct cross-regional correlation analysis, and issue handling instructions such as adjustment, load reduction, shutdown, isolation, purging, and emergency interlocking to achieve a closed-loop emergency response.

[0086] 4) The system automatically establishes a full lifecycle monitoring log, which records the monitoring data, collection time, parameter deviation, risk value, early warning level, model inference results, review conclusions, emergency instructions, execution status, recovery time and other full-process information of each node in real time. The log is fully retained for subsequent model iteration and optimization, accident tracing, process analysis and compliance audit, forming a traceable, reviewable and optimizable monitoring closed loop.

[0087] Step S6: Adaptively iteratively update and continuously optimize the model based on monitoring feedback.

[0088] 1) After the system has been running continuously for a period of time, each node synchronously collects monitoring feedback information, including real-time performance indicators of the model, long-term trend of data quality changes, records of working condition switching and process adjustment, early warning accuracy, false alarm / missed alarm events, equipment aging status, sensor drift, personnel handling feedback, etc., forming a comprehensive basis for iterative optimization.

[0089] 2) To address the issue of data quality changes, the data quality change iteration strategy preset in step S63 is adopted to adjust the local training ratio of data at each quality level, ensuring that the training data quality is accurately matched with the model training requirements. The initial training ratio is preset to 60% high-quality data, 30% medium-quality data, and 10% low-quality data. This initial ratio is set based on the historical quality distribution of lithium extraction production monitoring data in salt lakes, the model training effect, and the accuracy requirements of safety monitoring. It aims to prioritize the training ratio of high-quality data while making reasonable use of medium and low-quality data to improve the model's generalization ability.

[0090] Based on the current data quality assessment results, the average scores for each quality level are as follows: (High-quality data) (Medium quality data) (Low-quality data) The maximum average score of all current quality levels. (i.e., the average score of high-quality data); substitute the above parameters into the training ratio adjustment formula corresponding to changes in data quality. Substitute the values ​​one by one to calculate the adjusted training ratio for each quality level of data: For high-quality data, The initial training ratio of 60% is maintained because it has the highest average score and stable data quality, and can continue to be used as the core data for model training; for medium-quality data, The proportion has decreased slightly from the initial 30% because the average score of medium-quality data is higher than that of high-quality data. Appropriately reducing the training proportion can reduce its adverse impact on model accuracy; for low-quality data, The proportion has decreased significantly from the initial 10%. This is because the average score of this type of data is only 55, and the data integrity and accuracy are poor. Significantly reducing the training proportion can avoid introducing too much noise from low-quality data, which would lead to an increase in the model's false alarm rate.

[0091] After the calculation is completed, each edge monitoring node strictly follows the adjusted training ratio (60% high quality, 26% medium quality, and 7% low quality) to retrain its local model. During the training process, the three-stage course learning strategy of step S23 is used, and the learning rate and training rounds of each stage are adjusted accordingly to ensure that the training process meets the needs after the change in data quality. At the same time, the edge aggregation node adjusts the contribution evaluation weights, focusing on reducing the contribution weight of the local model corresponding to low quality data (from the original average of 0.08 to 0.05), and increasing the weight of the local model corresponding to high quality and medium quality data (the model weight corresponding to high quality data is increased from the original 0.14 to 0.16). Then, the weighted aggregation algorithm of step S33 is used to perform 10 rounds of edge layer aggregation again to ensure that the edge aggregation model can adapt to the scenario after the change in data quality.

[0092] 3) After completing the local model retraining and edge layer aggregation, a comprehensive performance verification of the iteratively updated model is performed. The verification scope covers the edge aggregation model and the global aggregation model. The verification dataset uses the latest collected monitoring data and the global test dataset (integrating the latest monitoring data of all workshops in the lithium extraction plant of the salt lake and historical data of the past month, totaling 3000 sets). The verification indicators strictly follow the performance requirements preset in steps S34 and S44.

[0093] The results showed that the false positive rate of the edge aggregation model decreased from 4.5% before the iteration to 3.8%, meeting the preset requirement of ≤4% for the false positive rate of the edge aggregation model. At the same time, its prediction accuracy increased from 96.5% to 97.2%, and the warning recall rate increased from 92.1% to 93.5%. Both core indicators were better than the preset standards, indicating that the data quality adjustment and contribution weight optimization have achieved the expected results.

[0094] For the global model, calculate its overall performance score. The score is calculated by weighting prediction accuracy (weight 0.3), warning recall (weight 0.3), false positive rate (weight 0.2), and warning response speed (weight 0.2). After iteration, the global model achieved a prediction accuracy of 97.8%, a warning recall of 94.2%, a false positive rate of 2.7%, and a warning response speed of 82ms, all of which met the preset requirements. The overall performance score of the global model before the iteration update is as follows: ,Will and Substitute into the performance improvement formula Calculation .

[0095] because Furthermore, all individual performance indicators of the edge aggregation model and the global aggregation model meet the corresponding preset requirements and comply with the iteration validity judgment criteria of step S64. Therefore, the original model is officially replaced, and the updated edge collaboration model and global aggregation model are redeployed to the corresponding nodes to continue the safety monitoring of the entire enterprise's lithium extraction process in the salt lake. At the same time, the training ratio, contribution weight, and model performance change data adjusted in this iteration are recorded in detail in the monitoring log to provide an iterative reference for possible data quality changes, model parameter aging, and other issues in the future.

[0096] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0097] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A progressive multi-layer federated learning method for safety monitoring of chemical processes, characterized in that, include: Construct a federated learning system with a multi-layered Internet of Things and complete the initial configuration. The federated learning system includes edge monitoring nodes, edge aggregation nodes and a global center node that are connected in sequence. Chemical process safety parameters are collected through edge monitoring nodes, and local models are trained in stages based on data quality. The parameters of the local models that meet the performance requirements are then uploaded to the corresponding edge aggregation nodes. Edge aggregation nodes perform weighted aggregation based on the contribution of each local model, and upload the edge aggregation model parameters that meet the performance requirements to the global central node. The global central node dynamically aggregates the data based on the differences in data distribution to obtain the global aggregation model. The models at each layer are deployed to the nodes at each layer to carry out multi-level linkage security monitoring, and the models at each layer are adaptively iterated and updated based on the monitoring feedback.

2. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, The construction of a federated learning system with a multi-layered Internet of Things (IoT) and the completion of initial configuration include: A three-layer architecture is established, consisting of edge monitoring nodes, edge aggregation nodes, and global center nodes. The edge monitoring nodes correspond to various monitoring points in the chemical industry, the edge aggregation nodes correspond to chemical workshops or regional control centers, and the global center nodes correspond to the overall control center of the chemical enterprise. Configure a basic model adapted to chemical safety monitoring for each layer node, initialize model weights and set relevant training hyperparameters, and configure encrypted communication protocols between layers; complete connectivity testing and adaptation debugging for each layer node to ensure normal data acquisition, parameter interaction and unified management functions.

3. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, The collection of chemical process safety parameters through edge monitoring nodes includes: Edge monitoring nodes collect key safety parameters in chemical production in real time, including temperature, pressure, and medium concentration; the collected raw data is preprocessed, including outlier removal, missing value imputation, and data standardization, to obtain preprocessed monitoring data.

4. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, Before the phased training of the local model based on data quality, the following steps are included: Construct a data quality assessment index system, including data completeness, data accuracy, data timeliness, and data consistency; calculate the comprehensive data quality score using a weighted summation method; and classify the data into high-quality data, medium-quality data, and low-quality data based on the comprehensive score.

5. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 4, characterized in that, The phased training of the local model based on data quality includes: The adaptive learning training is performed in three stages. In the first stage, high-quality data is used to train until the loss converges. In the second stage, medium-quality data is added and the learning rate is reduced to fine-tune the model. In the third stage, low-quality data is introduced and the learning rate is further reduced to train until the performance is stable. After training, the local model performance is verified by reserving a test set. If the performance meets the standard, the local model parameters are uploaded. If the performance does not meet the standard, the model is retrained.

6. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, The edge aggregation nodes perform weighted aggregation based on the contribution of each local model, including: The edge aggregation node receives the local model parameters and performance indicators of all edge monitoring nodes within its jurisdiction; combining the local model performance, data quality, and node importance, it uses a game theory algorithm to calculate the contribution of each local model and normalizes it to obtain the contribution weight. Adopt the federated averaging algorithm to weighted aggregate the local model parameters based on the contribution degree weights, and introduce a momentum term to smooth the parameter update to avoid oscillations; verify the performance of the edge aggregation model, upload the parameters after reaching the standard, and adjust the weights and re-aggregate if not up to the standard.

7. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, The global central node dynamically aggregates to obtain a global aggregation model in combination with the data distribution differences, including: The global central node receives the model parameters of each edge aggregation node, quantifies the distribution heterogeneity of each edge layer dataset; introduces attention weights according to the distribution differences to adapt to each edge aggregation model; Combines the attention weights with the chemical process mechanism knowledge, aggregates and fine-tunes the model parameters of each edge to obtain a global aggregation model; verifies the comprehensive performance of the global aggregation model, including the early warning response speed index, and enters the deployment stage after reaching the standard, and adjusts the aggregation strategy and retrains if not up to the standard.

8. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, Deploying each layer of the model to each layer of nodes to carry out multi-level linked safety monitoring, including: The global central node sends the global aggregation model parameters to the edge aggregation nodes. The edge aggregation nodes fuse their own aggregation models with the global aggregation model to generate a collaborative model and send it to the edge monitoring nodes; deploy each layer of the model correspondingly to achieve local early warning, regional review and global control; The edge monitoring nodes combine the real-time parameters with the model to calculate the safety risk value, and trigger an early warning according to the risk level; the edge aggregation nodes and the global central node review and judge in turn, issue emergency instructions, and record the relevant logs of monitoring, early warning and disposal at the same time.

9. The progressive multi-layer federated learning method for chemical process safety monitoring according to claim 1, characterized in that, Adapting and iteratively updating each layer of the model according to the monitoring feedback, including: Each layer of nodes synchronously collects feedback information, and the feedback information includes model performance, monitoring data and application effects; analyzes the feedback information to identify model problems, and the model problems include data quality changes, distribution shifts, parameter aging and process adjustments; Adopt corresponding strategies to adjust the training ratio, attention coefficient, model parameters or hyperparameters for different model problems to complete the model iteration; verify the performance of the updated model, evaluate the optimization effect through the performance improvement rate, replace the original model if effective, and adjust the strategy and repeat the optimization if ineffective.

10. A progressive multi-layer federated learning system for safety monitoring of chemical processes, characterized in that, Including edge monitoring nodes, edge aggregation nodes and a global central node that are sequentially communicatively connected; The edge monitoring node is configured to collect chemical process safety parameters, train the local model in stages based on the data quality, upload the local model parameters that meet the performance requirements to the corresponding edge aggregation node, and at the same time perform local real-time safety monitoring and early warning; The edge aggregation node is configured to perform weighted aggregation based on the contribution degree of each local model, upload the edge aggregation model parameters that meet the performance requirements to the global central node, and at the same time perform regional-level safety monitoring review and overall planning; The global central node is configured to receive the edge aggregation model parameters uploaded by each edge aggregation node, analyze the data distribution differences of each edge layer, and dynamically aggregate to obtain a global aggregation model based on the distribution differences. After verifying the performance of the global model, the model parameters are sent to each edge aggregation node, and at the same time perform global safety control and emergency instruction issuance, and adaptively iteratively update the model according to the monitoring information fed back by each layer of nodes.

Images