Industrial process optimization control method and system based on cloud abnormal data self-feedback

Abstract

The invention discloses an industrial process optimization control method and system based on cloud abnormal data self-feedback, and relates to the technical field of intelligent control and operation and maintenance. Comprising the following steps: receiving multi-dimensional feature vector data; constructing a feature space distribution baseline based on historical normal data, adaptively generating an abnormal threshold, and outputting abnormal event data; by taking the abnormal event data as a sample, dynamically generating and maintaining abnormal association graph data containing quantitative influence weights; traversing a spectrum identification propagation path by taking an abnormal event as a trigger, reconstructing and generating control strategy data in a mode of matching a node comprehensive influence score with a preset threshold value and selecting a corresponding instruction from an instruction library, and issuing and executing the control strategy data; and according to the effect evaluation data after strategy execution, feeding back and adjusting the update sensitivity of the baseline and the atlas. According to the invention, accurate traceability, active intervention and adaptive optimization control of industrial process abnormity are realized, and the intelligent level and robustness of the control system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and operation and maintenance technology, specifically to an industrial process optimization control method and system based on cloud-based abnormal data self-feedback. Background Technology

[0002] Industrial process optimization control is a key technology for improving manufacturing efficiency and product quality. Its core lies in analyzing production data and adjusting process parameters in real time to maintain stable system operation. In recent years, with the development of the Internet of Things (IoT) and cloud computing technologies, cloud-based data monitoring and control methods have gradually become mainstream, making it possible to centrally optimize large-scale, cross-process production.

[0003] In existing technologies, typical optimization methods often rely on setting fixed thresholds for key process parameters to trigger anomaly alarms, or on using statistical process control (SPC) models to perform local closed-loop adjustments on single equipment or processes. These methods can effectively identify deviations and implement control in stable production environments with weak coupling between variables, exhibiting a certain degree of real-time performance and reliability. However, for complex continuous industrial processes, there are strong couplings and nonlinear dynamic relationships between internal equipment, parameters, and processes. An anomaly in a local unit can propagate along a specific path, triggering a chain reaction that leads to widespread system fluctuations or production interruptions. Existing technologies lack the ability to model and reason about the mechanisms of such cross-level and cross-time-series anomaly propagation. Their control responses are often localized and delayed, unable to accurately pinpoint the root cause of the anomaly, and even less able to proactively adjust the control strategies of downstream key nodes to suppress propagation in the early stages of anomaly occurrence. This makes it difficult for control systems to cope with the systemic risks caused by anomaly propagation, limiting further improvements in optimization effectiveness. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an industrial process optimization control method and system based on cloud-based anomaly data self-feedback.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention discloses an industrial process optimization control method based on cloud-based anomaly data self-feedback, comprising the following steps:

[0007] Receive multi-dimensional feature vector data; the multi-dimensional feature vector data is generated by edge devices extracting features from raw industrial data.

[0008] Based on the historical normal multidimensional feature vector data, a feature space distribution baseline is constructed. By calculating the local density of the multidimensional feature vector data received in real time in the feature space distribution baseline, anomaly thresholds are adaptively generated, and anomaly event data with confidence is output.

[0009] Using the aforementioned abnormal event data as training samples, the intensity of causal influence between industrial entity nodes is iteratively calculated by introducing a Bayesian update rule with a time decay factor, and abnormal correlation graph data is dynamically generated and maintained; the abnormal correlation graph data includes quantified influence weights between nodes.

[0010] The propagation path is identified by traversing the abnormal correlation graph data, and the quantified influence weights are reconstructed based on the graph structure to generate control strategy data;

[0011] The control strategy data is sent to the execution end;

[0012] The system receives response data after execution, generates effect evaluation data by comparing changes in performance indicators, and adjusts the update sensitivity of the feature space distribution baseline and the anomaly correlation map data based on the effect evaluation data.

[0013] Secondly, this invention discloses an industrial process optimization control system based on cloud-based anomaly data self-feedback. The aforementioned industrial process optimization control method based on cloud-based anomaly data self-feedback includes:

[0014] The acquisition module is used to receive multidimensional feature vector data; the multidimensional feature vector data is generated by edge devices extracting features from raw industrial data.

[0015] The cloud-based anomaly detection module is used to construct a feature space distribution baseline based on the historically normal multidimensional feature vector data, adaptively generate anomaly thresholds by calculating the local density of the real-time received multidimensional feature vector data in the feature space distribution baseline, and output anomaly event data with confidence.

[0016] The graph evolution module is used to use the abnormal event data as training samples, and through a Bayesian update rule that introduces a time decay factor, iteratively calculate the causal influence strength between industrial entity nodes, dynamically generate and maintain abnormal association graph data; the abnormal association graph data includes quantified influence weights between nodes.

[0017] The strategy generation module is used to traverse the abnormal correlation graph data to identify the propagation path, and reconstruct the quantified influence weights based on the graph structure to generate control strategy data.

[0018] The strategy execution module is used to send the control strategy data to the execution end;

[0019] The feedback learning module is used to receive response data after execution, generate effect evaluation data by comparing changes in performance indicators, and adjust the update sensitivity of the feature space distribution baseline and the anomaly correlation map data based on the effect evaluation data.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention reduces the reliance on fixed thresholds, making the anomaly detection process less sensitive to normal data distribution drift caused by equipment aging, raw material batch changes, or production load adjustments. This effectively reduces false alarms and missed alarms while maintaining sensitivity, thus improving the reliability of anomaly event data.

[0022] 2. This invention can not only identify isolated anomalies, but also understand the propagation path and pattern of anomalies in the spatiotemporal dimension, thus providing data support for diagnosing problems from the root cause and overcoming the problem of one-sided and delayed control response caused by ignoring the correlation of anomalies in traditional methods.

[0023] 3. This invention suppresses the current abnormal node while weakening its momentum to propagate downstream in advance, thereby improving the overall anti-interference capability and operational robustness of the production process and realizing the transformation from passive alarm to active defense. Attached Figure Description

[0024] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0025] Figure 1 This is a flowchart of the steps of the present invention;

[0026] Figure 2 This is a diagram illustrating the steps for generating abnormal event data according to the present invention.

[0027] Figure 3 This is a system module connection diagram of the present invention;

[0028] Figure 4 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] In existing technologies, industrial process optimization control often relies on fixed threshold alarms or single-point closed-loop regulation, which is insufficient to address the problem of anomaly propagation across processes. Traditional methods often lag behind the spread of fluctuations as anomalies propagate along the production line, causing local anomalies to evolve into systemic failures. Existing systems lack the ability to dynamically model anomaly propagation paths, especially under complex operating conditions with tightly coupled variables, leading to collaborative failures of single-point control strategies and hindering the achievement of global optimization control.

[0031] To address the aforementioned issues, the study discovered a causal correlation between the time series and propagation paths of anomalous events, and implemented propagation tracking by establishing a dynamically evolving anomaly correlation graph. The research revealed that density-based anomaly detection is highly adaptable to changes in operating conditions, while the time-decaying correlation graph effectively distinguishes between random fluctuations and genuine propagation. Therefore, a method for dynamically updating the graph weights based on confidence events was proposed. Furthermore, through closed-loop verification, the strategy execution effect and the graph update mechanism were integrated into a feedback adjustment, constructing a self-evolving optimization control system.

[0032] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Example 1:

[0034] like Figure 1 As shown, the industrial process optimization and control method based on cloud-based anomaly data self-feedback includes the following steps:

[0035] Receives multidimensional feature vector data; the multidimensional feature vector data is generated by edge devices extracting features from raw industrial data.

[0036] A feature space distribution baseline is constructed based on historical normal multidimensional feature vector data. By calculating the local density of the real-time received multidimensional feature vector data in the feature space distribution baseline, anomaly thresholds are adaptively generated, and anomaly event data with confidence scores are output.

[0037] Using anomalous event data as training samples, the Bayesian update rule with a time decay factor is introduced to iteratively calculate the causal influence strength between industrial entity nodes, dynamically generate and maintain anomalous correlation graph data; the anomalous correlation graph data includes quantified influence weights between nodes.

[0038] Traverse the anomaly correlation graph data to identify the propagation path, and reconstruct the quantified impact weights based on the graph structure to generate control strategy data;

[0039] Send control strategy data to the execution end;

[0040] It receives response data after execution, generates effect evaluation data by comparing changes in performance indicators, and adjusts the update sensitivity of feature space distribution baseline and anomaly correlation map data based on the effect evaluation data.

[0041] In its specific implementation, this application describes an edge-side system comprised of data acquisition gateways and various sensors deployed in industrial settings. Edge devices, including sensors and actuators, are responsible for real-time acquisition of raw industrial data such as temperature, pressure, and flow rate. This raw data undergoes initial processing at the edge gateway, which is responsible for data preprocessing and uploading. The gateway uses built-in preprocessing algorithms for noise reduction and time-series alignment, and its computing unit performs sliding window analysis to extract statistical features, including mean, standard deviation, and trend terms. This generates multi-dimensional feature vector data representing the equipment and process status, which is then uploaded to the cloud system via the industrial network.

[0042] After receiving feature vector data from different edge devices, the cloud system's data storage and management module (typically composed of a distributed database and file system) is responsible for archiving this time-series data. The cloud-based data processing module retrieves historically stored multi-dimensional feature vector data under normal operating conditions. Within this feature space, a kernel density estimation algorithm is used to construct a feature space distribution baseline characterizing the system's health status. When new real-time feature vector data arrives, the module calculates the sparsity of data points within a preset neighborhood radius of the baseline to quantify its anomaly probability. The system dynamically adjusts the anomaly detection threshold based on the overall distribution of recent data sparsity and integrates the calculation results with the data quality identifiers reported from the edges, ultimately outputting anomaly event data with confidence scores. This process enables anomaly detection to adapt to changes in operating conditions, improving reliability.

[0043] Subsequently, the cloud-based graph analysis module uses high-confidence anomalous event data as training samples and applies a Bayesian update rule incorporating a time decay factor. Specifically, when the module detects an anomaly in an industrial entity node (such as a reactor or a conveyor belt), it calculates the conditional probability of other nodes subsequently exhibiting anomalies within a preset time window. The Bayesian formula is used to update the causal influence strength estimate from the pioneer node to subsequent nodes, while the introduction of the time decay factor ensures that the weight of historical evidence decreases over time, allowing the graph to reflect the latest associations in the system. Based on this, the system dynamically generates and maintains an anomaly association graph with industrial entities as nodes and causal influence strength and propagation delay as edges. This anomaly association graph intuitively reveals the potential propagation paths and impact ranges of anomalies in the production system.

[0044] Based on the aforementioned anomaly correlation graph, the strategy generation module automatically traverses the graph, starting with newly generated anomaly event data, to simulate and identify all possible downstream propagation paths. The module calculates the comprehensive impact score of each node based on the graph's topology and the quantified impact weights carried on the edges, thereby reallocating the resource weights of the control system. For example, for nodes located at the root of propagation or on the critical path, the system generates instructions to enhance their control strength or monitoring frequency, thus forming targeted control strategy data. This control strategy data is sent to the execution end via a secure channel, where it is received and executed by the edge programmable logic controller (PLC) or distributed control system (DCS).

[0045] Through the above technical solutions, this application enables industrial process control systems to not only adaptively detect anomalies, but also to deeply understand the dynamic propagation mechanism of anomalies, and generate forward-looking collaborative control strategies accordingly, thereby improving the overall robustness and operational efficiency of complex industrial processes in the face of internal disturbances and anomaly propagation.

[0046] This application further proposes that the specific steps for receiving multidimensional feature vector data include:

[0047] During the data acquisition phase, this application continuously collects raw industrial data at a sampling frequency of no less than 1Hz using edge hardware devices such as temperature, pressure, and flow sensors and PLC controllers deployed at the industrial site. This raw industrial data is transmitted to the edge gateway via an industrial bus (such as Modbus or PROFIBUS).

[0048] The edge gateway uses a built-in embedded processor to perform real-time preprocessing of the collected raw industrial data. First, a sliding window process is used, with the window length adjustable from 60 to 600 seconds depending on the process characteristics; for example, a 300-second window is used in chemical reactions, and a 60-second window in machining processes. Within each window, the system calculates multi-dimensional features, including time-domain statistics (mean, standard deviation, skewness), frequency-domain features (FFT main frequency amplitude), and trend terms (linear fitting slope), collectively forming a 128-dimensional feature vector. By extracting statistical features including the mean, standard deviation, and trend terms, multi-dimensional feature vector data is generated. These feature dimensions are determined through feature importance analysis, comprehensively characterizing the equipment's operating status.

[0049] To ensure data reliability, the system adds a quality identifier to each multidimensional feature vector data. This quality identifier includes at least a data integrity indicator. The quality identifier uses 16-bit binary encoding, where the first 4 bits represent the data integrity indicator, reflecting data integrity (based on CRC32 checksum); the middle 6 bits indicate the equipment operating conditions at the time of data acquisition (such as load rate and operating mode); and the last 6 bits record the signal-to-noise ratio (SNR) (calculated by the energy ratio before and after wavelet denoising). For example, when the SNR is below 20dB, the corresponding bit segment will be marked as low quality.

[0050] Based on a preset upload strategy, multi-dimensional feature vector data with quality indicators is uploaded. The data upload strategy is dynamically adjusted according to the quality indicators of the feature vectors. When the quality indicator shows that the data integrity is higher than 95% and the signal-to-noise ratio is greater than 25dB, an immediate upload mode is adopted; when the quality indicator shows that the data quality is at a medium level, it enters the buffer queue and is uploaded in batches every minute; when the network bandwidth is detected to be lower than 10Mbps, a data compression algorithm (such as PCA-based feature dimensionality reduction) is automatically started to compress the 128-dimensional features to 32 dimensions before uploading. This hierarchical upload mechanism effectively reduces network bandwidth pressure while ensuring data timeliness.

[0051] After the control strategy is executed, the system collects process response data again via edge sensors and uploads it to the cloud. The feedback learning module generates quantitative performance evaluation data by comparing changes in key performance indicators before and after strategy execution. This performance evaluation data is used to adjust the reconstruction cycle of the feature space distribution baseline and the update sensitivity of the anomaly correlation map data. If the strategy is effective, it accelerates model evolution; if it is ineffective, it tends to be conservative and may trigger a rollback, thus forming a closed loop of continuous improvement.

[0052] Key performance indicators include: ① Product qualification rate (unit: %), ② Unit energy consumption (unit: kW·h / ton), ③ System fluctuation range (unit: ℃ / min); the improvement threshold is set at 15% (i.e., an improvement of ≥15% in the indicators is considered effective).

[0053] Through the above implementation methods, this application achieves a reliable conversion from raw data to high-quality feature vectors, ensuring that the system can obtain stable and reliable feature inputs in different industrial scenarios. At the same time, it balances the contradiction between real-time performance and resource consumption through an adaptive upload strategy.

[0054] like Figure 2 As shown, this application further proposes specific steps for adaptively generating anomaly thresholds and outputting anomaly event data with confidence levels, including:

[0055] In the feature space, kernel density estimation is performed on multidimensional feature vector data under historical normal operating conditions to establish a feature space distribution baseline. During the baseline establishment phase, approximately 2.5 million sets of historical multidimensional feature vector data from 30 consecutive normal production days are retrieved as training samples. A baseline distribution model is constructed using the Gaussian kernel density estimation method, and its probability density function is... Represented as:

[0056]

[0057] Where x represents the real-time feature vector to be evaluated;

[0058] This is a historical normal sample;

[0059] d is the feature dimension, d=128;

[0060] h is the bandwidth parameter optimized using the Silverman rule, h=0.5;

[0061] n is the number of training samples.

[0062] The baseline distribution model is trained on a cloud GPU cluster using batch gradient descent. The model is considered to have converged when the rate of change of the likelihood function is less than 0.001 for 10 consecutive iterations.

[0063] For each real-time received multidimensional feature vector data, the sparsity of data points within a preset neighborhood radius in the feature space distribution baseline is calculated to generate real-time sparsity data. The neighborhood radius r = 2.5~5.0 (adaptively adjusted according to the feature dimension); preferably, the neighborhood radius r = 2.5, determined through cross-validation, to ensure that approximately 5% of the nearest neighbor samples are included in the 128-dimensional feature space. The real-time sparsity S(x) is obtained by calculating the inverse density of sample points within the neighborhood, as shown in the following formula:

[0064]

[0065] in, This represents the actual number of samples within the neighborhood.

[0066] This represents the total number of baseline samples;

[0067] Let r be the volume of a hypersphere with radius r in 128-dimensional space. , where d is the feature dimension and Γ is the gamma function.

[0068] when The higher the value, the more isolated the point is in the feature space, and the greater the probability of an anomaly.

[0069] The system acquires historical sparsity data within a preset first time window. Based on the statistical distribution of this historical sparsity data, it dynamically adjusts the outlier threshold T to generate adaptive threshold data. This dynamic adjustment of the outlier threshold is achieved using a sliding time window mechanism. The statistical distribution is fitted to a Gaussian distribution of the historical sparsity data, with the upper limit of the 95% confidence interval used as the initial threshold, and dynamically corrected using the operating condition fluctuation coefficient (±10%). The system maintains a preset first time window of 24 hours and continuously records historical sparsity data within this window. The mean sparsity within the window is calculated every 15 minutes. and standard deviation and update the anomaly threshold. The calculation formula is as follows:

[0070]

[0071] Where k is an adjustment coefficient, which adaptively changes between 1.5 and 3.0 according to the system operating status. For example, in chemical processes, the typical value is 1.8-2.5.

[0072] When the production load change rate is detected to exceed 10%, the system automatically increases the k value to above 2.5 to reduce false alarms; under stable operating conditions, the k value is maintained at around 1.8 to maintain detection sensitivity.

[0073] The real-time sparse data is compared with the adaptive threshold data: if the real-time sparse data is higher than the adaptive threshold data, an initial anomaly label is generated.

[0074] After generating initial anomaly markers, the system performs confidence level fusion calculation. The initial anomaly markers and quality indicators are fused, and anomaly event data with confidence levels is generated according to preset confidence level calculation rules.

[0075] Confidence level C is determined by the sparsity anomaly score. Together with the data quality metric Q, it is determined by the following formula:

[0076]

[0077] The weighting coefficient can be set as follows: This ratio was determined through experimental verification. The data quality metric Q is derived from the quality indicator in the edge preprocessing stage. When it indicates that the data integrity is below 90% or the signal-to-noise ratio is less than 20dB, the Q value decreases accordingly. Finally, only anomalies with a confidence level C exceeding 0.75 are submitted to the subsequent processing flow.

[0078] This application achieves accurate identification and quantitative assessment of anomalies in industrial processes. Feature space modeling based on kernel density estimation can accurately describe the complex distribution pattern of normal operating conditions, while the confidence mechanism combined with data quality effectively distinguishes between real anomalies and data acquisition problems. The system can maintain stable detection performance under different operating conditions, avoiding the misjudgment problem caused by traditional fixed threshold methods when operating conditions change.

[0079] This application further proposes that, by introducing a Bayesian update rule with a time decay factor, the specific steps for iteratively calculating the causal influence strength between industrial entity nodes include:

[0080] When an anomaly event with a confidence level C exceeding 0.75 is detected at the first node (denoted as node A) among the computational industrial entity nodes, the system immediately initiates the correlation analysis process. Within a preset second time window, the system calculates the conditional probability of the second node exhibiting anomalies. Taking an anomaly in the temperature of the upstream reactor as an example, the system will monitor whether downstream nodes (including pressure vessels, heat exchangers, etc.) subsequently exhibit anomalies within a preset second time window (typically set to 10-30 minutes based on process propagation characteristics, such as 25 minutes for polymerization reactions).

[0081] The Bayesian formula is used to update the causal influence strength estimate from the first node to the second node, and a time decay factor is introduced into the Bayesian formula to reduce the weight of historical evidence.

[0082] Only when the intensity of the updated causal influence exceeds a preset probability threshold will the corresponding directed edge be established or strengthened in the abnormal correlation graph data.

[0083] The system calculates the causal influence strength between nodes using Bayesian update rules, and its core formula is:

[0084]

[0085] This represents the prior probability of the strength of the causal influence from the first node (node ​​A) to the second node (node ​​B);

[0086]

[0087]

[0088] Time decay factor Using exponential decay:

[0089]

[0090] In the formula, The decay constant represents the time difference between the current time and the time when historical evidence occurred. The process time was set to 720 minutes (i.e. 12 hours) based on the process characteristics. This reduced the weight of historical evidence from a week ago to about 0.01, ensuring that the graph could reflect the latest state of the system.

[0091] In practice, the system maintains a priori probability matrix, with initial values ​​set based on the process flow diagram and expert experience. For example, for directly connected process equipment, the initial values ​​of the priori probability matrix are set based on the process flow diagram, with an initial probability of 0.3 for directly related nodes and 0.1 for indirectly related nodes. Each time new evidence of anomaly propagation is observed, the system updates the corresponding causal probability using the above formula. Only when the updated causal influence strength exceeds a preset probability threshold of 0.6 will the system establish or strengthen the corresponding directed edge in the anomaly association graph. Simultaneously, the system records the latest update time of each edge to ensure the correct application of the time decay factor.

[0092] The graph maintenance process also includes regular structural optimization. Regularly triggered instant updates for anomalous events are combined with a global decay maintenance system every 12 hours. The system performs a global scan every 24 hours, removing edges with a causal influence strength below 0.2 and marking edges with a strength between 0.2 and 0.4 for observation. This mechanism effectively prevents excessive complexity in the graph structure and ensures inference efficiency.

[0093] Through the above technical solution, this application can continuously optimize the understanding of the intrinsic connections within the process during operation, providing a reliable topological foundation for subsequent root cause analysis and control strategy generation. Compared to static process flow diagrams, this dynamically evolving correlation map better reflects the dynamic coupling relationships in the actual production process, improving the accuracy and timeliness of anomaly diagnosis.

[0094] This application further proposes that, during the dynamic generation and maintenance of abnormal correlation graph data, a complete graph data lifecycle management mechanism has been established. The maintenance program automatically performs a global traversal every 24 hours, systematically evaluating and updating the causal influence strength of all edges in the graph. The maintenance cycle is determined based on the production rhythm of typical industrial processes, which can reflect system state changes in a timely manner while avoiding excessive consumption of computing resources.

[0095] The specific steps for maintaining abnormal correlation graph data include:

[0096] Periodically traverse all edges in the abnormal association graph data;

[0097] Apply a decay operation to the causal influence strength of each edge in the anomalous correlation graph data;

[0098] When the intensity of causal influence is lower than the preset failure threshold, the corresponding edge is removed from the abnormal association graph data.

[0099] During the traversal, the system employs a piecewise function strategy for the decay operation applied to each edge. For edges actively updated within the last 7 days (i.e., edges supported by at least one causal evidence), the decay coefficient is set to 0.95; for edges updated between 7 and 30 days, the decay coefficient is adjusted to 0.85; and for edges not updated for more than 30 days, a decay coefficient of 0.7 is used. This differentiated decay strategy ensures the stability of recently important associations while accelerating the natural elimination of outdated associations. The specific calculation of the decay operation is as follows: the causal influence strength of each edge is multiplied by the corresponding decay coefficient, and the last update timestamp of that edge is recorded.

[0100] The failure threshold was set considering the actual characteristics of the industrial process. By analyzing the stability of anomaly propagation patterns in historical data, the failure threshold was determined to be 0.15. This value was determined based on the following consideration: when the causal influence strength is below this threshold, the correlation effect of the corresponding anomaly propagation relationship in the actual production process is weaker than the influence level of random noise. Before performing a removal operation, the system marks the relevant edge as "to be observed" and continuously monitors it for the next three maintenance cycles. If no new evidence emerges during this period, the edge is ultimately removed from the graph, and the removal record is stored in the graph change log for auditing purposes.

[0101] The maintenance process also includes a consistency check. The system checks for orphaned nodes (nodes with both in-degree and out-degree being zero), which are typically created due to edge removal. For nodes that remain orphaned for three consecutive maintenance cycles, the system sends a confirmation request to the operations personnel and decides whether to retain the node based on the feedback. This mechanism effectively prevents the meaningless expansion of the graph structure, ensuring the simplicity and usability of the graph.

[0102] By implementing the aforementioned maintenance mechanism, this application ensures that the graph always reflects the latest production process correlation characteristics through a dynamic maintenance strategy, providing a reliable foundation for graph-based anomaly diagnosis and control decisions. Compared to static correlation models, this graph maintenance mechanism with self-optimization capabilities can better adapt to the dynamic changes in complex industrial environments, ensuring the long-term stability and accuracy of the system.

[0103] This application further proposes that the specific steps for reconstructing the quantified influence weights based on the graph structure to generate control strategy data include:

[0104] Starting from the node corresponding to the abnormal event data, a depth-first search algorithm is used to traverse the entire path in the abnormal correlation graph data to identify all downstream propagation paths. The traversal depth is set to 3-5 node levels according to the process characteristics. This range is determined based on the hop count statistics of abnormal propagation in typical industrial processes and can cover more than 95% of actual propagation scenarios.

[0105] Based on the causal influence strength along the downstream propagation path, the comprehensive influence score of each node in the anomaly correlation graph data is calculated. The calculation process for the comprehensive influence score of each node employs an improved PageRank algorithm, which not only considers the edge weights of directly connected nodes but also introduces a path decay factor to reflect the effect of multi-hop propagation. Specifically, the comprehensive influence score of node v... Calculated using the following formula:

[0106]

[0107] Where U represents the set of all upstream abnormal nodes;

[0108] The strength of the causal influence from node u to v;

[0109] d(u,v) is the shortest path distance between the two nodes;

[0110] The attenuation factor is set according to the attenuation characteristics of process propagation, preferably 0.6-0.8. It can accurately reflect the intensity attenuation law of anomalies during propagation through historical data verification.

[0111] It is the initial impact value of the source node, determined based on the confidence level of the abnormal event.

[0112] The system matches the overall impact score with multiple preset priority thresholds, and selects the corresponding control parameter adjustment command from a preset control command library based on the matching results. The calculated overall impact score is compared with preset three-level priority thresholds. These thresholds are set based on process safety requirements and equipment operating specifications: the first-level threshold of 0.8 corresponds to critical protection actions, the second-level threshold of 0.6 corresponds to important adjustment operations, and the third-level threshold of 0.4 corresponds to general optimization measures. Based on the threshold range in which the score falls, the system selects the corresponding control parameter adjustment command from the preset control command library. The control command library was jointly established by process experts and maintenance engineers during the system commissioning phase and contains standardized control strategy templates for different equipment types and anomaly modes.

[0113] The system combines all selected control parameter adjustment commands to generate control strategy data. During the command combination phase, a conflict detection and resolution mechanism is employed to ensure compatibility between different control commands. When a command conflict is detected, the system prioritizes the control requirements of the nodes with higher impact scores, while ensuring that fundamental process constraints are not violated. The final generated control strategy data is encapsulated in a standardized JSON format, containing complete version information, execution condition descriptions, and expected performance evaluations.

[0114] Through the aforementioned control strategy generation mechanism, this application can dynamically adjust the control response according to the real-time situation of abnormal propagation. Compared with traditional fixed control strategies, it has better adaptability and global optimization effect, providing a reliable guarantee for the safe and stable operation of complex industrial processes.

[0115] This application further proposes that the specific steps for matching the comprehensive influence score with a preset priority threshold and selecting the corresponding control parameter adjustment instruction from a preset control instruction library based on the matching result include:

[0116] During the control command matching phase, the system achieves precise policy allocation based on a pre-established threshold-command mapping relationship.

[0117] If the overall impact score of a node exceeds a first preset threshold, a first instruction is selected from the control instruction library to increase the proportional gain of the corresponding control loop. The first preset threshold is set to 0.8, a value derived from the triggering requirements for critical equipment protection actions in process safety regulations. When the overall impact score of a node exceeds this threshold, it indicates that the node is in the core path of abnormal propagation and has a high risk level. At this time, the system calls the first instruction from the control instruction library. This instruction template requires increasing the proportional gain of the corresponding control loop by 30-50% from its original value. The specific adjustment range is determined based on the linear interpolation of the impact score. For example, for a temperature control loop, the proportional gain is adjusted from the original 2.0 to 2.8; for a pressure control loop, the proportional gain is adjusted from 1.5 to 2.1. This gain enhancement strategy can strengthen the control system's ability to suppress abnormal fluctuations and accelerate the system's recovery to a stable state.

[0118] If a node is located on the propagation path and its overall impact score exceeds the second preset threshold, a second instruction to shorten the node's data sampling period is selected from the control instruction library. For nodes located on the abnormal propagation path and whose overall impact score exceeds the second preset threshold of 0.6, the system activates a monitoring enhancement mechanism. The second preset threshold is determined based on the need for early warning of anomalies and can effectively identify potential risk points that require key monitoring. The system selects a second instruction from the instruction library to shorten the data sampling period of the corresponding node from the usual 5 seconds to 1 second, and simultaneously changes the node's data upload mode from periodic upload to change-triggered upload. This upgrade in monitoring strategy enables the system to capture early signs of abnormal propagation more promptly, providing more timely data support for subsequent control decisions.

[0119] If the overall impact score of a node is lower than the third preset threshold, a third instruction is selected from the control instruction library to maintain or reduce the proportional gain of the node's control loop. When the overall impact score of a node is lower than the third preset threshold by 0.4, it indicates that the node is less affected by abnormal propagation. In this case, the system uses the third instruction to maintain the proportional gain of the corresponding control loop at the baseline level or appropriately reduce it by 5-10%, while restoring the sampling period to the normal setting. This differentiated processing method avoids excessive consumption of control resources and ensures that the system concentrates its main control capabilities on critical links. All threshold parameters are determined through process risk assessment during the system initialization phase and dynamically calibrated during operation based on actual results.

[0120] The control command library is built upon an equipment characteristic database and a process knowledge base. Each command template includes complete parameter setting specifications and execution condition descriptions. When selecting a command to execute, the system simultaneously verifies the current status of the equipment and process constraints to ensure the feasibility and safety of the control command. For example, before increasing the proportional gain, the system confirms whether the equipment is in automatic control mode and whether the relevant sensors are functioning properly.

[0121] Through the precise matching mechanism of the aforementioned control commands, this application achieves adaptive control based on anomaly propagation. The threshold-based hierarchical strategy ensures the allocation of control resources to critical nodes, while the standardized management of the command library guarantees the standardization and traceability of control actions. This hierarchical control response mechanism enables the system to optimize the allocation efficiency of control resources and improve the overall intelligence and operational efficiency of the control system while ensuring process safety.

[0122] This application further proposes that the steps for distributing control policy data to the execution end include:

[0123] Add version identifiers and execution condition constraints to the generated control policy data;

[0124] Control strategy data is sent to the corresponding edge controller via a secure channel; the edge controller executes the control strategy from the PLC / DCS.

[0125] The control command is executed after the edge controller verifies the integrity and validity of the control strategy data.

[0126] During the control strategy issuance process, the system first adds complete metadata identifiers to the generated control strategy data. Version identifiers use the format "major version.minor version.revision number," where the major version corresponds to significant changes in the control strategy, the minor version records adjustments to strategy parameters, and the revision number indicates emergency fixes and updates. Execution constraints are embedded in the strategy data package in structured data format, including equipment status prerequisites (e.g., requiring the actuator to be in automatic mode), process parameter boundaries (e.g., temperature setpoint variation not exceeding ±10%), time validity windows (e.g., the maximum execution time of the strategy is 30 minutes), and rollback trigger conditions (e.g., immediate termination if critical parameters exceed limits). These constraints are derived from the equipment operation manual and process safety procedures to ensure that control commands are executed within safe limits.

[0127] Policy data transmission employs a secure TLS 1.3-based channel, with two-way certificate authentication enabling identity verification between the cloud and edge controller. Data packets are encrypted using the AES-256 algorithm before transmission and appended with a SHA-256 digital signature to prevent tampering. The transmission protocol is designed for asynchronous acknowledgment; after sending policy data, the cloud requires the edge controller to return a receipt acknowledgment within 500 milliseconds. If no acknowledgment is received within this timeframe, the data is retransmitted via a backup communication link, with a maximum of three retransmissions. This mechanism ensures reliable delivery of policy data in industrial network environments.

[0128] Upon receiving policy data, the edge gateway first performs integrity and validity verification. Integrity verification includes checking the hash value and digital signature of the data packet to ensure that the data has not been tampered with during transmission. Validity verification checks whether the policy version is higher than the current execution version, whether the execution conditions meet the on-site device status, and whether the parameter settings are within the device's allowable range. During the verification process, the controller collects device status data in real time and compares it with the policy execution conditions. Only when all verification conditions are met will the controller load the policy instruction into the execution queue.

[0129] During the execution phase, a gradual parameter adjustment strategy is adopted, setting a reasonable rate of change for key parameters. For example, the temperature setpoint is adjusted gradually to approach the target value at a rate not exceeding 2°C per minute, avoiding any impact on the process. During execution, the controller continuously monitors equipment feedback and process parameters. Once an anomaly is detected or the rollback condition is met, the strategy execution is immediately terminated and the system is restored to a safe state, while the execution results are fed back to the cloud system.

[0130] Through the aforementioned secure delivery and execution mechanisms, this application achieves closed-loop control from cloud-based decision-making to edge execution. Version management and execution constraints ensure the standardization and security of the control strategy, encrypted transmission and integrity verification guarantee the reliability of data exchange, while gradual adjustment and anomaly rollback mechanisms ensure the smoothness of the control process. This end-to-end secure control architecture enables cloud-based intelligent decision-making to be securely and reliably implemented in the industrial field, leveraging the advantages of cloud analytics while ensuring the safe and stable operation of the industrial control system.

[0131] This application further proposes that the specific steps for adjusting the update sensitivity of the feature spatial distribution baseline and anomaly correlation map data based on the effect evaluation data include:

[0132] Calculate the rate of change of key performance indicators before and after the implementation of the control strategy data, and use it as a score for the strategy effectiveness.

[0133] During the system feedback learning phase, the data processing for effectiveness evaluation employed a multi-indicator fusion calculation method. Key performance indicators included equipment operational stability, energy consumption efficiency, and product quality pass rate, with each indicator assigned different weights based on process importance. Strategy effectiveness scoring was then performed. Calculated using the following formula:

[0134]

[0135] in, and These represent the values ​​of the i-th performance indicator before and after the strategy implementation, respectively.

[0136] The weighting coefficients for the corresponding indicators are set based on production management standards. For example, in continuous chemical processes, the weight for operational stability is set to 0.5, energy efficiency to 0.3, and product quality to 0.2. This weighted calculation method can comprehensively and objectively reflect the overall effect of the control strategy.

[0137] n represents the total number of indicators.

[0138] The system sets a preset improvement threshold of 0.15 and a deterioration threshold of -0.08. These thresholds are determined by analyzing the statistical distribution of effective optimization and failure cases in historical data.

[0139] When the strategy effectiveness score exceeds the preset improvement threshold, the update frequency of the feature space distribution baseline and anomaly correlation map data is increased. The system shortens the update cycle of the feature space distribution baseline from 24 hours to 8 hours and the update cycle of the anomaly correlation map from 12 hours to 4 hours. The increase in update frequency is achieved by increasing the data sampling density of model training, while keeping the bandwidth parameter of kernel density estimation unchanged to ensure the stability of model updates.

[0140] When the strategy performance score falls below a preset deterioration threshold, the update frequency is reduced and a strategy rollback mechanism is triggered. First, the update cycles for the feature space distribution baseline and the anomaly correlation map are extended to 48 hours and 24 hours, respectively, to reduce the model's sensitivity to recent low-quality data. Simultaneously, the strategy rollback mechanism is triggered, which includes three steps: immediately stopping the execution of the current control strategy and restoring the control parameters to the previous stable version; analyzing the cause of the strategy failure to identify whether it was due to model bias or sudden changes in operating conditions; and adding the failed cases to the training sample library for improving future strategy generation algorithms.

[0141] During the model update sensitivity adjustment process, the system employs a smooth transition strategy. Changes in update frequency are not instantaneous but are gradually adjusted in increments of 25%, with each adjustment cycle lasting 6 hours, to avoid system oscillations caused by sudden changes in sensitivity. Simultaneously, the system records all historical sensitivity adjustments, establishing an adjustment effect evaluation file for optimizing future sensitivity control parameters.

[0142] Through the aforementioned adaptive learning mechanism, this application ensures the accuracy of the optimization direction through quantitative scoring-based performance evaluation, achieves reasonable allocation of system resources through differentiated response strategies triggered by tiered thresholds, and maintains the stability of the learning process through a smooth transition mechanism. This adaptive learning architecture enables the system to continuously optimize its perception capabilities and decision-making level during continuous operation, gradually improving the control quality of complex industrial processes and achieving a transformation from single-step optimization to continuous progress.

[0143] The following is an example of an industrial process optimization control method and system based on cloud-based anomaly data self-feedback:

[0144] A continuous reaction process line in a fine chemical company comprises three key units: reactor R101 (main reaction), heat exchanger E201 (material cooling), and distillation column T301 (product separation), connected in series with pump sets via pipelines. The system needs to monitor parameters such as temperature, pressure, and flow rate in real time to address the propagation of abnormalities caused by catalyst activity decay and fluctuations in raw material composition.

[0145] The edge sensing layer deploys temperature sensors (measurement range -50~300℃, accuracy ±0.1℃, sampling frequency 1Hz), pressure transmitters (0~2MPa, accuracy 0.2%FS), and industrial-grade ARM Cortex-A53 edge gateways. It performs sliding window processing (window length 300 seconds, step size 60 seconds) on the raw data, extracts statistical features such as mean, standard deviation, and trend terms to generate a 64-dimensional feature vector, adds a quality label with data integrity ≥95%, and then uploads it to the cloud.

[0146] The cloud-based anomaly detection module calls upon 30 days of normal operating data (2.5 million feature vectors), constructs a probability density model using a kernel density estimation algorithm, and calculates the sparsity within a neighborhood radius r=2.5 for the real-time received feature vectors, based on the mean of historical sparsity data within a 24-hour window. with standard deviation Dynamically adjust the abnormal threshold to (Threshold 0.69 when k=2.0), when the real-time sparsity S(x)=0.82 exceeds the threshold, the fused data quality is identified using the confidence formula. Generate anomalous event data with a confidence level of 0.87.

[0147] The graph evolution module takes this anomalous event as a starting point, calculates the conditional probability of downstream node anomalousness within a 30-minute time window, and uses a Bayesian update rule combined with a time decay factor. (τ=720 minutes) Update the causal influence strength, increasing the causal probability of R101→E201 from the initial 0.3 to 0.65. After exceeding the preset threshold of 0.6, a directed edge is established in the anomaly correlation graph. The strategy generation module identifies the propagation path R101→E201→T301 through depth-first search, calculates the comprehensive influence score of each node based on the improved PageRank algorithm, where node I(E201) = 0.65 × 0.7^1 × 0.87 = 0.39 for E201, matches the third threshold of 0.4, selects the instruction to reduce the proportional gain of the control loop by 10%, adds the version identifier "V1.2.1" to the control strategy data, and sends it to the PLC controller of E201 through the TLS 1.3 encrypted channel, executing the adjustment of the proportional gain from 2.0 to 1.8. The feedback learning module collects the data after execution and calculates the strategy effect score. After exceeding the improvement threshold of 0.15, the feature spatial distribution baseline update cycle is shortened from 24 hours to 12 hours, and the attenuation factor τ of the abnormal correlation map is reduced from 720 minutes to 600 minutes.

[0148] In actual operation, when R101 experiences temperature anomalies due to decreased catalyst activity, the system completes anomaly detection within 5 minutes. By identifying the propagation path in advance through dynamic correlation graphs, the pre-control strategy for E201 ensures that the temperature fluctuation of downstream T301 does not exceed the allowable range of the process. After feedback optimization, the response time to anomalies caused by subsequent raw material fluctuations is shortened to 3 minutes, significantly reducing the number of unplanned shutdowns per month and effectively ensuring production continuity.

[0149] This embodiment fully presents the closed-loop process from anomaly detection and propagation modeling to active control, achieving the practical effect of suppressing anomaly propagation and improving system robustness in complex industrial processes.

[0150] Example 2:

[0151] like Figure 3 and Figure 4 As shown, the industrial process optimization control system based on cloud-based anomaly data self-feedback, using the aforementioned industrial process optimization control method based on cloud-based anomaly data self-feedback, includes:

[0152] The acquisition module is used to receive multidimensional feature vector data; the multidimensional feature vector data is generated by edge devices extracting features from raw industrial data.

[0153] The cloud-based anomaly detection module is used to construct a feature space distribution baseline based on historical normal multidimensional feature vector data. By calculating the local density of the real-time received multidimensional feature vector data in the feature space distribution baseline, it adaptively generates anomaly thresholds and outputs anomaly event data with confidence scores.

[0154] The graph evolution module is used to iteratively calculate the causal influence strength between industrial entity nodes by introducing a Bayesian update rule with a time decay factor, using abnormal event data as training samples, and dynamically generating and maintaining abnormal association graph data; the abnormal association graph data includes quantified influence weights between nodes.

[0155] The strategy generation module is used to identify the propagation path by traversing the abnormal correlation graph data, using abnormal event data as the trigger point, and reconstructing the quantitative impact weights based on the graph structure to generate control strategy data.

[0156] The strategy execution module is used to send control strategy data to the execution end;

[0157] The feedback learning module is used to receive response data after execution, generate effect evaluation data by comparing changes in performance indicators, and adjust the update sensitivity of feature space distribution baseline and anomaly correlation map data based on the effect evaluation data.

[0158] The system's hardware architecture consists of an edge sensing layer, a cloud processing layer, and an execution control layer. The edge sensing layer is deployed in the industrial field and includes various sensors (temperature, pressure, flow sensors, etc.) and edge gateways (using industrial-grade ARM processors). The cloud processing layer is deployed on a cloud server cluster, equipped with GPU accelerator cards for model training. The execution control layer comprises field-mounted programmable logic controllers (PLCs) and distributed control systems (DCS). Data communication between these layers occurs via industrial Ethernet and 4G / 5G networks.

[0159] The system's workflow begins with data acquisition at the edge sensing layer. Sensors continuously collect raw industrial data at a sampling frequency of 1Hz and transmit it to the edge gateway via Modbus / TCP protocol. The gateway's built-in preprocessing program performs sliding window processing on the raw data (window length 300 seconds, step size 10 seconds), calculating statistical features such as mean, standard deviation, and trend term in real time, generating 128-dimensional feature vector data. After adding quality labels containing data integrity indicators to the feature vectors, the gateway selects to upload them to the cloud immediately or in batches, depending on network conditions.

[0160] Upon receiving feature vector data, the cloud-based anomaly detection module first retrieves sample data from the historical database under normal operating conditions to construct a baseline for the feature space distribution. This module runs on a cloud server equipped with a GPU and uses a kernel density estimation algorithm to establish a probability density model for the 128-dimensional feature space. For feature vectors input in real-time, the module calculates their local density in the feature space and generates anomaly event data with confidence levels through a dynamic threshold mechanism. When an anomaly event with a confidence level exceeding 0.75 is detected, the system immediately triggers subsequent processing steps.

[0161] The graph evolution module takes anomaly events detected by the front end as input and dynamically updates the anomaly association graph using a time-decay-based Bayesian inference algorithm. This module maintains a graph structure containing all process equipment nodes, with each edge recording the causal influence strength and last update time between nodes. The module periodically performs decay operations, downweighting or removing outdated associations to ensure the graph always reflects the latest process status. Graph data is stored in a cloud-based graph database, supporting efficient path lookup and impact propagation analysis.

[0162] Upon receiving an anomaly event, the strategy generation module immediately initiates a graph traversal algorithm to identify all possible propagation paths of the anomaly. Based on the causal influence strength along the path, the module calculates the comprehensive influence score of each node and selects the corresponding control parameter adjustment instruction from the control instruction library by matching it with a preset threshold. The generated control strategy data contains complete version information and execution constraints and is sent to the execution control layer via an encrypted channel.

[0163] Upon receiving the control strategy, the PLC / DCS at the execution control layer first verifies the integrity of the data and the compliance of the execution conditions. If everything is correct, it executes the control command. During execution, the equipment status is continuously monitored, and a rollback mechanism is immediately activated upon detecting any anomalies. Execution results and process data are fed back to the cloud system via the edge gateway.

[0164] The feedback learning module collects the effect data after the strategy is executed, calculates the strategy effect score through a multi-index fusion algorithm, and dynamically adjusts the update sensitivity of the feature space distribution baseline and the anomaly correlation map based on the score results, thus completing the entire optimization loop.

[0165] Taking a distillation column system in a chemical production process as an example. After system deployment, on a certain day, the edge sensor detected a fluctuation in the column bottom temperature. After feature extraction, a feature vector containing parameters such as the average temperature and fluctuation frequency was generated. The cloud-based anomaly detection module discovered a local density anomaly in this vector in the feature space and generated a high-confidence anomaly event. The graph evolution module identified, based on historical data, that this anomaly might affect parameters such as the column top temperature and reflux ratio, and immediately updated the anomaly correlation graph. After traversing the graph, the strategy generation module determined that the column top temperature controller needed priority adjustment, selected an instruction to increase the proportional gain from the instruction library, and generated a control strategy. The strategy execution module sent the control instruction to the column top temperature PLC controller, and the controller executed the parameter adjustment after verification. After the adjustment, the system monitored an improvement in product purity and a decrease in energy consumption. The feedback learning module calculated the strategy effectiveness score to be 0.18, exceeding the improvement threshold, so the model update cycle was shortened to accelerate the system learning process.

[0166] Through the above implementation methods, this application achieves a complete closed loop from anomaly detection to control optimization. The hardware-coordinated architecture ensures the system's real-time performance and reliability, the hierarchical processing mechanism optimizes the utilization efficiency of computing resources, and the adaptive learning capability enables the system to continuously improve control quality. This system architecture, integrating perception, analysis, decision-making, and execution, provides a complete technical solution for the intelligent optimization of complex industrial processes, improving production efficiency and product quality while ensuring system stability.

[0167] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An industrial process optimization control method based on cloud-based anomaly data self-feedback, characterized in that: Includes the following steps: Receive multidimensional feature vector data; The multidimensional feature vector data is generated by edge devices extracting features from raw industrial data; Based on the historical normal multidimensional feature vector data, a feature space distribution baseline is constructed. By calculating the local density of the multidimensional feature vector data received in real time in the feature space distribution baseline, anomaly thresholds are adaptively generated, and anomaly event data with confidence is output. Using the aforementioned abnormal event data as training samples, the intensity of causal influence between industrial entity nodes is iteratively calculated by introducing a Bayesian update rule with a time decay factor, and abnormal correlation graph data is dynamically generated and maintained; the abnormal correlation graph data includes quantified influence weights between nodes. The propagation path is identified by traversing the abnormal correlation graph data, and the quantified influence weights are reconstructed based on the graph structure to generate control strategy data; The control strategy data is sent to the execution end; The system receives response data after execution, generates effect evaluation data by comparing changes in performance indicators, and adjusts the update sensitivity of the feature space distribution baseline and the anomaly correlation map data based on the effect evaluation data.

2. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 1, characterized in that: The specific steps for receiving multidimensional feature vector data include: The collected raw industrial data is processed using a sliding window method to extract statistical features, including the mean, standard deviation, and trend term, and to generate the multidimensional feature vector data. Add a quality identifier to the multidimensional feature vector data, the quality identifier including at least a data integrity indicator; Based on the preset upload strategy, upload multidimensional feature vector data with the quality identifier.

3. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 2, characterized in that: The specific steps for adaptively generating anomaly thresholds and outputting anomaly event data with confidence scores include: In the feature space, the kernel density of the multidimensional feature vector data under historical normal operating conditions is estimated to establish the feature space distribution baseline. For each real-time received multidimensional feature vector data, the sparsity of data points within a preset neighborhood radius in the feature space distribution baseline is calculated to generate real-time sparse data. Obtain historical sparse data within a preset first time window, and dynamically adjust the anomaly threshold based on the statistical distribution of the historical sparse data to generate adaptive threshold data; The real-time sparse data is compared with the adaptive threshold data: if the real-time sparse data is higher than the adaptive threshold data, a preliminary anomaly marker is generated; By combining the initial anomaly marker and the quality identifier, and using a preset confidence calculation rule, the anomaly event data with confidence level is generated.

4. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 1, characterized in that: The specific steps for iteratively calculating the causal influence strength between industrial entity nodes by introducing a Bayesian update rule with a time decay factor include: When abnormal event data is detected at the first node among the computing industrial entity nodes, the conditional probability of the second node being abnormal is calculated within a preset second time window. The causal influence strength estimate from the first node to the second node is updated using a Bayesian formula, and the time decay factor is introduced into the Bayesian formula to reduce the weight of historical evidence. Only when the intensity of the updated causal influence exceeds a preset probability threshold, will the corresponding directed edge be established or strengthened in the abnormal correlation graph data.

5. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 4, characterized in that: The process of dynamically generating and maintaining abnormal correlation graph data includes the following specific steps: Periodically traverse all edges in the abnormal association graph data; A decay operation is applied to the causal influence strength of each edge in the anomalous correlation graph data; When the intensity of the causal influence is lower than a preset failure threshold, the corresponding edge is removed from the abnormal correlation graph data.

6. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 1, characterized in that: The specific steps for reconstructing the quantified influence weights based on the graph structure to generate control strategy data include: Starting from the node corresponding to the abnormal event data, the abnormal association graph data is traversed to identify all downstream propagation paths; Based on the causal influence intensity along the downstream propagation path, calculate the comprehensive influence score of each node in the abnormal correlation graph data; The comprehensive impact score is matched with multiple preset priority thresholds, and the corresponding control parameter adjustment instruction is selected from the preset control instruction library according to the matching result. The control strategy data is generated by combining all selected control parameter adjustment instructions.

7. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 6, characterized in that: The specific steps for matching the comprehensive impact score with a preset priority threshold and selecting the corresponding control parameter adjustment instruction from a preset control instruction library based on the matching result include: If the overall impact score of the node exceeds the first preset threshold, then a first instruction for increasing the proportional gain of the control loop corresponding to the node is selected from the control instruction library; If the node is located on the propagation path and the overall influence score exceeds the second preset threshold, then a second instruction for shortening the data sampling period of the node is selected from the control instruction library; If the overall impact score of the node is lower than the third preset threshold, then a third instruction for maintaining or reducing the proportional gain of the node's control loop is selected from the control instruction library.

8. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 1, characterized in that: The steps for sending control policy data to the execution end include: Add version identifiers and execution condition constraints to the generated control strategy data; The control strategy data is sent to the corresponding edge controller via a secure channel; After verifying the integrity and validity of the control strategy data, the edge controller executes the control command.

9. The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in claim 1, characterized in that: The specific steps for adjusting the feature space distribution baseline and the update sensitivity of the anomaly correlation map data based on the effect evaluation data include: Calculate the rate of change of key performance indicators before and after the implementation of the control strategy data, and use it as the strategy effectiveness score; When the strategy effectiveness score exceeds the preset improvement threshold, the update frequency of the feature space distribution baseline and the abnormal correlation map data is increased. When the strategy performance score falls below a preset deterioration threshold, the update frequency is reduced and a strategy rollback mechanism is triggered.

10. An industrial process optimization control system based on cloud-based anomaly data self-feedback, characterized in that: The industrial process optimization control method based on cloud-based anomaly data self-feedback as described in any one of claims 1 to 9 includes: The acquisition module is used to receive multidimensional feature vector data; the multidimensional feature vector data is generated by edge devices extracting features from raw industrial data. The cloud-based anomaly detection module is used to construct a feature space distribution baseline based on the historically normal multidimensional feature vector data, adaptively generate anomaly thresholds by calculating the local density of the real-time received multidimensional feature vector data in the feature space distribution baseline, and output anomaly event data with confidence. The graph evolution module is used to use the abnormal event data as training samples, and through a Bayesian update rule that introduces a time decay factor, iteratively calculate the causal influence strength between industrial entity nodes, dynamically generate and maintain abnormal association graph data; the abnormal association graph data includes quantified influence weights between nodes. The strategy generation module is used to traverse the abnormal correlation graph data to identify the propagation path, and reconstruct the quantified influence weights based on the graph structure to generate control strategy data. The strategy execution module is used to send the control strategy data to the execution terminal; The feedback learning module is used to receive response data after execution, generate effect evaluation data by comparing changes in performance indicators, and adjust the update sensitivity of the feature space distribution baseline and the anomaly correlation map data based on the effect evaluation data.

Images