Intelligent control method and system for feed production line

Abstract

The present application belongs to the technical field of intelligent control, and particularly relates to a kind of feed production line intelligent control method and system, its method includes: obtaining multi-source heterogeneous sensor data, initial basic credibility distribution is generated in combination with Gaussian mixture model and sliding window volatility;Time series confidence knowledge graph is constructed and condition credibility function is continuously updated;Inference derived evidence, use the mutual information value after time window alignment and feature aggregation to calculate confidence weight, weighted fusion evidence and high conflict credibility are attenuated and redistributed, global credibility distribution and conflict coefficient are obtained;Based on conflict coefficient adjustment asymmetric utility function matrix, calculate bayesian expected utility, select the optimal strategy and issue instructions.The present application improves the adaptability of credibility distribution, avoids conflict evidence fusion distortion, enhances the adaptive ability of decision to asymmetric risk, effectively guarantees the safety and stability of feed production.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology. More specifically, this invention relates to an intelligent control method and system for a feed production line. Background Technology

[0002] The feed production workshop environment is complex, with high dust levels, strong vibrations, and electromagnetic interference. This results in noise, drift, and abrupt changes in the multi-source heterogeneous data collected by sensors, leading to significant uncertainty. Traditional deterministic control algorithms struggle to perceive the true production status when faced with deeply coupled multi-parameter and randomly fluctuating conditions. When logical contradictions arise in the multi-source data, the control system is prone to misjudgments, causing frequent fluctuations or executing erroneous actions, leading to quality problems such as nutrient loss in the finished feed and excessive pellet powdering. Therefore, constructing a control framework that can adapt to dynamic and complex operating conditions and achieve highly reliable evidence fusion and intelligent risk decision-making is a crucial step in improving the level of intelligence in feed production.

[0003] Currently, existing technologies utilize real-time data collection from multiple fields, including equipment operating parameters and material characteristics, to predict faults based on multi-task learning models. Chinese patent application CN121069951A discloses an anomaly alarm method and system for a feed production control system. Using a pre-trained fault prediction model, the prediction confidence of fault type and fault point is calculated, where the confidence is positively correlated with the distance from the feature vector to the known pattern cluster center or decision boundary. Based on this technology, feature extraction and cluster analysis of multi-dimensional data can improve the accuracy and sensitivity of identifying production anomalies to a certain extent.

[0004] However, while the aforementioned technical solutions can improve the coverage of fault diagnosis to some extent, they still have significant shortcomings when applied to intelligent control of production lines. In the basic confidence assignment generation stage, existing technologies mostly rely on static distance metrics or single models, failing to consider the correlation between historical data distribution characteristics and real-time fluctuations, resulting in poor adaptability of confidence assignment to dynamic operating conditions. Existing knowledge graphs are mostly static structures, lacking a time-series update mechanism to characterize the intensity of influence transmission between nodes, and cannot achieve synchronous evolution of the reasoning process and real-time operating conditions. In the process of multi-source evidence fusion, existing technologies lack a weight adjustment mechanism based on the correlation of finished product quality indicators. When faced with highly conflicting evidence generated by divergence from multiple sources of sensors, the fusion results are prone to distortion or evidence deadlock. Decision models mostly use fixed symmetric utility functions, failing to consider the asymmetric risks caused by control errors in feed production, and cannot adjust risk preferences in real time according to the conflict coefficient, leading to an overly aggressive or conservative control strategy under abnormal operating conditions, making it difficult to meet the actual needs of high reliability and high stability operation of production lines. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as poor adaptability of reliability allocation, lack of temporal updates in knowledge graphs, easy distortion in the fusion of highly conflicting evidence, and difficulty in adapting control decisions to asymmetric risks, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an intelligent control method for a feed production line, comprising: S1, acquiring multi-source heterogeneous sensor data during the feed production process, setting key operating condition parameter thresholds as production state propositions, and generating initial basic confidence assignment functions for each production state proposition by combining a Gaussian mixture model constructed based on historical sensor data and a sliding window volatility based on real-time data; S2, constructing a time-series confidence knowledge graph with production operating condition parameters as nodes and preset correlation relationships between parameters as directed edges, and continuously updating the conditional confidence function in the time-series confidence knowledge graph representing the strength of influence transmission between nodes according to real-time data; S3, using the initial basic confidence assignment function as source node evidence, and updating the conditional confidence function in the time-series confidence knowledge graph according to the conditional confidence function. The confidence propagation inference is performed using the degree function to generate derived evidence. The confidence weight is calculated using the mutual information value between the historical data of the evidence source after time window alignment and feature aggregation and the quality indicators of the finished material batch. The source node evidence and derived evidence are fused using the weighted evidence combination rule to obtain the global confidence distribution and conflict coefficient. S4. The basic asymmetric utility function matrix is ​​preset. Based on the conflict coefficient, the basic asymmetric utility function matrix is ​​adjusted using the risk preference function to convert the global confidence distribution into a Bayesian decision probability distribution. Combining the aforementioned Bayesian decision probability distribution and the adjusted utility function matrix, the Bayesian expected utility is calculated for the candidate control strategy. The candidate control strategy with the largest Bayesian expected utility value is selected, and production control instructions are generated and issued.

[0007] This invention achieves synchronous evolution of the inference process and real-time operating conditions by fusing dynamic volatility to generate basic confidence assignments and continuously updating the conditional confidence function. It utilizes mutual information values ​​to calculate confidence weights and introduces a conflict redistribution mechanism to avoid distortion and deadlock in high-conflict sensor evidence fusion. Based on the dynamic conflict coefficient, it adjusts the asymmetric utility function matrix, enabling the decision-making algorithm to perceive and adapt to asymmetric risks. This effectively avoids overly aggressive or conservative control strategies under abnormal operating conditions, improving the accuracy, safety, and high stability of feed production line control in complex environments.

[0008] Preferably, the generation of the initial basic reliability assignment function includes: iteratively training the historical sensor data using the expectation-maximization algorithm to obtain the mean and covariance matrix of each Gaussian component of the Gaussian mixture model; calculating the standard deviation of the real-time data within the current preset time window as the sliding window volatility; inputting the current real-time data into the Gaussian mixture model to calculate the posterior probability of each Gaussian component; performing dimensionless processing on the sliding window volatility to obtain the dimensionless volatility; calculating the normal adjustment coefficient and the abnormal adjustment coefficient based on the dimensionless volatility; and, according to the normal or abnormal attributes of the corresponding production state proposition, multiplying the posterior probability of the Gaussian component of the corresponding normal state proposition by the normal adjustment coefficient, and multiplying the posterior probability of the Gaussian component of the corresponding abnormal state proposition by the abnormal adjustment coefficient, and then performing normalization processing to obtain the initial basic reliability assignment function of the corresponding production state proposition.

[0009] This invention introduces dimensionless volatility to dynamically adjust the posterior probability of Gaussian mixture model components, enabling the basic confidence assignment to simultaneously take into account the distribution characteristics of historical data of the equipment and the abnormal fluctuation state of the current real-time signal. This significantly improves the adaptive capability of the system in state perception under complex working conditions and avoids the misjudgment of working conditions that is easily caused by simply relying on static thresholds.

[0010] Preferably, the step of continuously updating the conditional confidence function representing the strength of influence transmission between nodes in the time-series confidence knowledge graph based on real-time data includes: acquiring the time series of sensor data of the source node and the target node within adjacent acquisition periods; calculating the time warping distance between the time series of sensor data of the source node and the time series of sensor data of the target node; using the time warping distance to calculate the influence attenuation coefficient between nodes; multiplying the preset basic conditional confidence function by the influence attenuation coefficient; and allocating the reduced confidence quality to the global unknown frame to obtain the updated conditional confidence function representing the strength of influence transmission between nodes.

[0011] This invention utilizes time-warped distance to accurately calculate the influence attenuation coefficient, characterizing the differences in transmission delay in the time response of multi-source heterogeneous sensor data. By adaptively discounting the conditional confidence and transferring lost quality to the unknown frame, it effectively reduces the erroneous interference caused by operating condition drift or physical transmission delay on confidence propagation inference.

[0012] Preferably, the step of performing confidence propagation reasoning based on conditional confidence functions in the temporal confidence knowledge graph to generate derived evidence for non-source nodes includes: traversing the temporal confidence knowledge graph and identifying all non-source nodes without assigned initial evidence as target nodes; for each target node, extracting all predecessor node evidence pointing to that node and their corresponding conditional confidence functions, wherein the predecessor node evidence includes the initial source node evidence of the predecessor node or the derived evidence generated in the previous round of confidence propagation reasoning; calculating the product of each predecessor node evidence and its corresponding conditional confidence function, marginalizing and summing the state space of the predecessor node to obtain single-path transmission evidence mapped to the state space of the target node; and aggregating all single-path transmission evidence pointing to the same target node using combination rules to generate derived evidence for non-source nodes.

[0013] Preferably, obtaining the global reliability distribution and conflict coefficient includes: calculating the conflict factor between each pair of source node evidence and derived evidence according to the weighted evidence combination rule, and using the sum of all conflict factors as the conflict coefficient; determining whether the conflict coefficient is greater than a preset conflict threshold, and if it is, determining that there is high conflict reliability; multiplying the evidence combination item determined to have high conflict reliability by a decay factor, the decay factor being composed of a negative exponential function of the inverse of the corresponding confidence weight; redistributing the reliability quality released after the conflict reliability decays to the unconflicted propositions according to the average reliability proportion of each remaining unconflicted proposition in the evidence before fusion, allocating the decayed residual conflict reliability quality to the unknown frame, and obtaining the global reliability distribution of the current production condition after normalization.

[0014] This invention employs a negative exponential decay factor based on the inverse of the confidence weight to specifically suppress high-conflict confidence, and proportionally redistributes the released confidence quality to unconflicted propositions and unknown frames. This mechanism breaks the decision-making deadlock caused by logical contradictions in multimodal sensor information, preserves objectively correct propositions in high-confidence source evidence to the greatest extent possible, and eliminates the destructive misleading effect of low-confidence conflicting evidence on the overall judgment result.

[0015] Preferably, adjusting the basic asymmetric utility function matrix includes: establishing a basic asymmetric utility function matrix with candidate control strategies as rows and production state propositions as columns, where matrix elements represent the baseline utility value of executing the corresponding strategy under a specific production state, wherein the control strategy matching the current production state corresponds to a positive baseline utility value, and the control strategy corresponding to the erroneous shutdown strategy, the missed fault reporting strategy, or the over-intervention strategy corresponds to a negative baseline utility value, and the absolute value of the negative baseline utility corresponding to the missed fault reporting strategy or the equipment damage strategy is greater than the absolute value of the negative baseline utility corresponding to ordinary malfunction; constructing a risk preference function, which is a monotonically decreasing exponential function with the conflict coefficient as the independent variable; multiplying the positive baseline utility value in the basic asymmetric utility function matrix by the output value of the risk preference function, and dividing the negative baseline utility value by the output value of the risk preference function to obtain the adjusted utility function matrix.

[0016] This invention constructs a monotonically decreasing exponential risk preference function with the conflict coefficient as the independent variable, achieving a proportional shrinkage of positive return expectations and a multiple amplification of negative fault penalties on the basic matrix. This bidirectional asymmetric dynamic adjustment mechanism enables the controller to automatically deduce control boundaries with extremely high safety margins when facing data environments with extremely high uncertainty and high conflict, significantly reducing the rate of serious equipment damage accidents caused by misjudgments in traditional logic fusion.

[0017] Preferably, the generation and issuance of production control instructions includes: performing probability transformation processing on the global confidence distribution, assigning the confidence quality allocated to the unknown frame or composite proposition to the corresponding single production state proposition, and obtaining a Bayesian decision probability distribution; for each candidate control strategy, multiplying and summing the state utility values ​​corresponding to each state in the adjusted utility function matrix of the strategy with the corresponding state probabilities in the Bayesian decision probability distribution, and obtaining the Bayesian expected utility of the candidate control strategy; sorting all candidate control strategies in descending order of their Bayesian expected utility, and selecting the candidate control strategy ranked first; generating production control instructions containing equipment action parameters and execution time based on the selected candidate control strategy, and issuing them to the actuators of the corresponding production equipment via industrial Ethernet.

[0018] Preferably, the acquisition of multi-source heterogeneous sensor data during the feed production process includes: using the OPCUA protocol library to connect to the programmable logic controller of the feed production line, collecting data from motor current sensors, equipment speed sensors, valve opening sensors, temperature sensors, humidity sensors, and vibration sensors in real time, and writing the collected data into a time-series database.

[0019] This invention clarifies the acquisition method and underlying storage structure of multi-source heterogeneous data. Through a unified communication protocol specification and time-series database architecture, it ensures the stable access and persistent storage of high-frequency, multi-dimensional sensor raw data streams, providing an absolutely reliable data foundation for subsequent high-intensity dynamic reliability calculations and spectral inference.

[0020] Preferably, the construction of a time-series confidence knowledge graph with production condition parameters as nodes and preset correlations between parameters as directed edges includes: sequentially adding the feeding auger speed as a graph node, adding the ring die pressure roller differential speed as a graph node, adding the steam condensate volume as a graph node, adding the conditioning temperature as a graph node, and adding the pelleting chamber blockage risk as a virtual graph node, and adding directed edges between nodes according to the control logic of the feed process flow to construct the main framework of the time-series confidence knowledge graph.

[0021] Secondly, the present invention provides an intelligent control system for a feed production line, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent control method for a feed production line is implemented.

[0022] By adopting the above technical solution, a computer program is generated from the above-mentioned intelligent control method for a feed production line and stored in a memory so that it can be loaded and executed by a processor. Terminal equipment can then be made based on the memory and processor for convenient use.

[0023] The beneficial effects of this invention are as follows: This invention constructs a highly adaptive uncertainty knowledge representation system for multi-source heterogeneous sensor data by combining a Gaussian mixture model based on historical data with the sliding window volatility of real-time signals. This effectively overcomes the shortcomings of traditional confidence allocation in adapting to dynamic and complex working conditions. At the same time, by constructing a time-series confidence knowledge graph and continuously updating the conditional confidence function, it accurately characterizes the spatiotemporal coupling influence between key working parameters and their evolution over time. This breaks the limitations of static knowledge structures and achieves seamless synchronization between the logical reasoning process and the actual working conditions of the production line.

[0024] Furthermore, in the reasoning and multi-source evidence fusion stage, this invention utilizes the mutual information value, which is strongly correlated with the finished product quality index after aggregation of time-series features, as a dynamic confidence weight to accurately attenuate and redistribute the high-conflict confidence, effectively eliminating fusion distortion and decision deadlock caused by logical deviations in sensor data, and obtaining a highly objective and realistic global confidence distribution. Moreover, based on the actual conflict coefficient, the basic asymmetric utility function matrix is ​​dynamically adjusted and the Bayesian expected utility is calculated, enabling the control system to keenly perceive and adaptively avoid asymmetric risks, significantly improving the security of production control command issuance under abnormal operating conditions, and ensuring the stable operation of the feed production process and the quality of the final product in complex environments. Attached Figure Description

[0025] Figure 1 This is a flowchart of an intelligent control method for a feed production line according to the present invention; Figure 2 This is a schematic diagram illustrating the influence of conduction delay on damping characteristics in this invention; Figure 3 This is a schematic diagram comparing the production and operation effects under continuous multi-period hidden danger conditions in this invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] This invention discloses an intelligent control method for a feed production line, referring to... Figure 1 This includes steps S1-S4: S1. Generate the initial basic confidence assignment function.

[0028] In an optional embodiment, multi-source heterogeneous sensor data during the feed production process are acquired, and the threshold values ​​of each key operating condition parameter are set as production state propositions. An initial basic confidence assignment function is generated for each production state proposition by combining a Gaussian mixture model constructed based on historical sensor data and a sliding window volatility based on real-time data.

[0029] Specifically, the programmable logic controller (PLC) of the feed production line is connected using the OPCUA protocol library. Real-time data from motor current sensors, equipment speed sensors, valve opening sensors, temperature sensors, humidity sensors, and vibration sensors is collected and written to the InfluxDB time-series database to store multi-source heterogeneous data. A function to read CSV files is called to extract historical normal and abnormal operating condition data as a training set. With three Gaussian components, a Gaussian mixture model is constructed and parameters are fitted. The posterior probability of each Gaussian component corresponding to the real-time sensor data is calculated, and the calculation results are used as the basic support for each production state proposition. The standard deviation of the real-time data within a preset sliding time window is calculated and used as the sliding window volatility. Boundary conditions set by expert experience, such as excessively high crusher speed or excessively low pellet mill temperature, are mapped to state propositions. The sliding window volatility is dimensionlessly processed, and normal and abnormal adjustment coefficients are constructed according to the normal or abnormal attributes of the production state propositions. The normal adjustment coefficient decreases with increasing dimensionless volatility, while the abnormal adjustment coefficient increases with increasing dimensionless volatility. The posterior probability of the Gaussian component is adjusted using normal or abnormal adjustment coefficients and then normalized to generate an initial basic confidence assignment function that supports each production state proposition. This initial basic confidence assignment function is then used as the initial quality function for each source of evidence.

[0030] As one possible implementation, an initial basic confidence assignment function is generated for each production state proposition by combining a Gaussian mixture model built based on historical sensor data with a sliding window volatility based on real-time data.

[0031] Specifically, the expectation-maximization algorithm is used to iteratively train historical sensor data to obtain the mean and covariance matrix of each Gaussian component of the Gaussian mixture model; the standard deviation of real-time data within the current preset time window is calculated as the sliding window volatility; the current real-time data is input into the Gaussian mixture model to calculate the posterior probability of each Gaussian component; the sliding window volatility is dimensionlessly processed to obtain the dimensionless volatility; based on the above dimensionless volatility, the normal adjustment coefficient and the abnormal adjustment coefficient are calculated; according to the normal or abnormal attributes of the corresponding production state proposition, the posterior probability of the Gaussian component of the corresponding normal state proposition is multiplied by the normal adjustment coefficient, and the posterior probability of the Gaussian component of the corresponding abnormal state proposition is multiplied by the abnormal adjustment coefficient, and after normalization, the initial basic reliability assignment function of the corresponding production state proposition is obtained.

[0032] The training phase of the Gaussian mixture model requires the extraction of historical data, such as historical data on the main motor current of the pellet mill and the opening of the steam valve. The sample set is set to 10,000 consecutive records, and the optimal range for the number of Gaussian components K is set to 3 to 5. For example, selecting K=3 maps the normal low-load state proposition, the normal high-load state proposition, and the equipment abnormal state proposition, respectively. The maximum number of iterations of the expectation-maximization algorithm is configured to 200, with a convergence tolerance of 1e-5, and the output is the mean vector of the three Gaussian components. With covariance matrix And their respective prior weights.

[0033] In the online inference phase, the length N of the sliding window is set to 10 to 60 sampling periods. Assuming the sensor sampling rate is 1Hz, N is set to 30, i.e., 30 seconds of data. The standard deviation of the main motor current time series within the window is calculated, and assuming a standard deviation of 12.5A, it is used as the sliding window volatility. .

[0034] Based on the maximum minus minimum normalization calculation approach, this invention introduces a minimum positive number to prevent the denominator from being zero, constructing a dimensionless volatility, the calculation method of which is as follows:

[0035] in, It is a dimensionless volatility; For sliding window volatility; This represents the lowest volatility in history. This represents the highest volatility in history. It is a very small positive number.

[0036] The calculated dimensionless volatility Limited to between 0 and 1. Assuming historical minimum volatility. Equal to 2A, maximum volatility Equal to 32A, the current yields a dimensionless volatility. The value is equal to 0.35. At this point, the latest real-time current data is input into the trained Gaussian mixture model, and the posterior probabilities of the three components are calculated as follows: Posterior probability of the first component... Equals 0.6, the posterior probability of the second component. Equals 0.3, posterior probability of the third component Equal to 0.1. The normal adjustment coefficient is set to... The abnormal adjustment coefficient is The posterior probability of the first component with normal properties. With the posterior probability of the second component Multiply by the normal adjustment coefficient respectively The adjusted posterior probability of the first component is obtained. The adjusted posterior probability of the second component is equal to 0.39. Equals 0.195; posterior probability of the third component with anomalous attributes. Then multiply by the abnormal adjustment coefficient The adjusted posterior probability of the third component is obtained. The result equals 0.135. Summing the three values ​​yields 0.72, which is then normalized by dividing by the sum. The initial basic reliability assignments for each proposition are as follows: 0.542 for the proposition under normal low load conditions, 0.271 for the proposition under normal high load conditions, and 0.187 for the proposition under abnormal conditions. This calculation process relatively strengthens the proposition under abnormal conditions when real-time volatility increases, while avoiding misjudgments caused by relying solely on static thresholds.

[0037] S2. Update the conditional confidence function of the graph.

[0038] In an optional embodiment, a time-series confidence knowledge graph is constructed with production condition parameters as nodes, and the pre-defined relationships between parameters are used as directed edges. The conditional confidence function in the time-series confidence knowledge graph is continuously updated based on real-time data. The conditional confidence function is used to represent the influence transmission strength between nodes.

[0039] Specifically, an empty graph structure is instantiated, and then the following nodes are added sequentially: feeding auger speed, ring die roller differential speed, steam condensate volume, conditioning temperature, and pelleting chamber blockage risk. Directed edges are added between nodes according to the control logic of the feed processing flow, constructing the main framework of the time-series confidence knowledge graph. For each directed edge, the Granger causality test P-value of the upstream node's real-time data sequence on the downstream node's real-time data sequence is calculated. A bounded mapping function is used to convert the Granger causality test P-value into an influence transmission coefficient between nodes. The influence transmission coefficient ranges from 0 to 1; a smaller P-value indicates a more significant association, and a larger influence transmission coefficient. The influence transmission coefficient is applied to a preset basic conditional confidence function, and the decayed remaining confidence quality is assigned to the global unknown framework to obtain an updated conditional confidence function. A timed task module triggers the above calculation process every minute, updating the weight attribute values ​​of the conditional confidence function on the directed edges in real time.

[0040] As one possible implementation, the conditional confidence function in the time-series confidence knowledge graph is continuously updated based on real-time data. Specifically, the time series of sensor data from the source node and the target node within adjacent acquisition periods are obtained; the time warping distance between the time series of sensor data from the source node and the target node is calculated; the influence attenuation coefficient between nodes is calculated using this time warping distance; the preset basic conditional confidence function is multiplied by the influence attenuation coefficient; and the reduced confidence quality is allocated to the global unknown frame to obtain the updated conditional confidence function representing the strength of influence transmission between nodes.

[0041] The time series of source nodes, such as the feed auger speed, and target nodes, such as the conditioner outlet temperature, are obtained from a real-time time series database. The preferred range for the sliding acquisition period window width L is 60 to 120 sampling points; here, L is set to 60. Normalized sensor data within adjacent 60-second intervals are extracted to form two one-dimensional sequence vectors. A 60×60 cumulative distance matrix is ​​constructed, with a window width constraint of 10%, allowing only a time shift of 6 units diagonally to the left and right. Dynamic programming is used to find the shortest alignment path from the top left to the bottom right corner to address the transmission delay between feed and temperature, calculating the minimum cumulative normalization distance D. Assuming the minimum cumulative normalization distance D calculated for the current period is 15.8, this value characterizes the temporal consistency difference between source node fluctuations and target node responses; a smaller value indicates a higher degree of consistency, while a larger value indicates a weaker transmission relationship.

[0042] Following the general calculation approach of the exponential decay model, an exponential decay function is constructed using the obtained minimum cumulative regularization distance D to calculate the influence decay coefficient. The calculation method is as follows:

[0043] in, To affect the attenuation coefficient; This is the adjustment coefficient; This is the minimum cumulative normalized distance.

[0044] In this invention, the adjustment coefficient The range is from 0.01 to 0.1, and the adjustment coefficient is set. The calculated attenuation coefficient is 0.05. The value is 0.454. Assume the expert database has a pre-defined basic conditional confidence function. It is set that when the source node exhibits excessive rotational speed, the target node's temperature decreases. In this case, the initial propagation confidence for the lower temperature is based on the excessive rotational speed. The initial confidence level is 0.85, corresponding to the unknown frame. The base reliability is equal to 0.15. Multiplying this base reliability by the decay factor yields the updated specific reliability; the updated specific reliability for low temperature based on excessively high rotational speed is... Equals 0.386. All the confidence quality lost due to distance decay (0.464) is transferred to the globally unknown frame. This update mechanism reduces the confidence level of the unknown frame to 0.614 when the temporal response relationship between the source and target nodes weakens, and transfers the corresponding confidence quality to the unknown frame, thereby reflecting the uncertainty caused by condition drift or propagation delay.

[0045] Reference Figure 2 As the cumulative minimum obstruction misalignment distance after regularization and alignment gradually increases, the confidence depreciation ratio of the transmission direction causality strength exhibits a monotonically decreasing trend. The misalignment distance reflects the transmission delay of multi-source heterogeneous sensor data over time. The monotonically decreasing trend demonstrates that this invention, by discounting the conditional confidence using an exponential decay function, can adaptively reduce the confidence weight of abnormal fluctuations, thereby effectively mitigating the adverse effects of transmission delay on confidence propagation inference.

[0046] S3. Derive derived evidence and integrate its reliability.

[0047] In an optional embodiment, the initial basic confidence assignment function is used as source node evidence. Confidence propagation inference is performed in the time-series confidence knowledge graph based on the conditional confidence function to generate derived evidence for non-source nodes. Confidence weights are calculated using the mutual information values ​​between historical data of the evidence sources (after time window alignment and feature aggregation) and the quality indicators of finished material batches. A weighted evidence combination rule is then used to fuse the source node evidence and derived evidence. High-conflict confidence levels during the fusion process are attenuated using the confidence weights, and then redistributed using the confidence weights to obtain the global confidence distribution and conflict coefficient.

[0048] Specifically, a confidence propagation calculation module is imported to construct a confidence propagation engine, using the generated initial basic confidence assignment function as the input for source node evidence in the graph. When using probabilistic graph calculation tools, the confidence values ​​of the corresponding single-element state propositions in the basic confidence assignment function are converted into prior distributions of node states before input. Node evidence is propagated sequentially along the directed edges of the graph in topological order, or in breadth-first order or a preset number of iterations. By calculating the product of the predecessor node evidence and the conditional confidence function on the directed edge, the state space of the predecessor node is marginalized and summed to derive derived evidence for the target node; the predecessor node evidence includes the initial source node evidence and the derived evidence generated in the previous round of confidence propagation inference. Given that sensor data is high-frequency time series while laboratory test indicators are low-frequency discrete data, the following approach is first adopted: Based on the timestamps of finished product batch tests, high-frequency sensor time series data within the corresponding production time window are extracted. The mean and variance are extracted to construct a frequency-reduced aggregated sensor historical feature vector, thus aligning the data's time scale. Subsequently, the mutual information values ​​between each sensor's historical feature vector and the particle hardness index sequence returned by laboratory tests are calculated, as well as the mutual information values ​​between each sensor's historical feature vector and the pulverization rate quality index sequence of the finished product batch. The mutual information values ​​of all sensors are normalized to obtain the confidence weights of each evidence source. Orthogonal summation is performed on source node evidence and derived evidence, and the empty set basic confidence allocation value in the evidence combination formula is extracted as the conflict coefficient. A high-conflict condition is defined when the conflict coefficient exceeds a preset conflict threshold. The sigmoid activation function is used to smoothly scale the confidence weights, transferring the attenuated high-conflict confidence quality to unknown propositions within the global framework. The remaining part is processed according to the Dempster combination rule to obtain a converged global confidence distribution and conflict coefficient.

[0049] When loops exist in the time-order confidence knowledge graph, a maximum number of iterations or a convergence threshold is set, and propagation stops when the change in evidence between adjacent nodes is less than a preset threshold.

[0050] As one possible implementation, confidence propagation reasoning is performed in the temporal confidence knowledge graph based on the conditional confidence function to generate derived evidence for non-source nodes. Specifically, the temporal confidence knowledge graph is traversed, and all non-source nodes without assigned initial evidence are identified as target nodes. For each target node, all predecessor node evidence pointing to that node and their corresponding conditional confidence functions are extracted. Predecessor node evidence includes the initial source node evidence of the predecessor node or the derived evidence generated in the previous round of confidence propagation reasoning. The product of each predecessor node evidence and its corresponding conditional confidence function is calculated, and the state space of the predecessor node is marginalized and summed to obtain single-path transmission evidence mapped to the state space of the target node. All single-path transmission evidence pointing to the same target node is aggregated using combination rules to generate derived evidence for non-source nodes.

[0051] A graph traversal algorithm is used to scan the graph database structure built on Neo4j, filtering out virtual non-source nodes lacking observational evidence due to the absence of sensors as target nodes, such as the granulation chamber blockage risk node. Two predecessor nodes pointing to the target node are located, assuming they are node A (ring die roller differential speed) and node B (steam condensate flow rate). The current node evidence assigned to node A is extracted from the upstream processing flow; for example, the confidence level of node A in an elevated state is 0.7, and the confidence level of node A in a normal state is 0.3. Edge attributes include a conditional confidence function matrix; for example, the conditional confidence level of blockage based on an elevated state is 0.8, the conditional confidence level of smooth operation based on an elevated state is 0.1, and the confidence level of an unknown state is 0.1; similarly, the conditional confidence level of blockage based on a normal state is 0.1, the conditional confidence level of smooth operation based on a normal state is 0.8, and the confidence level of an unknown state is 0.1. Based on the law of total probability, product calculations and marginal summation are performed to derive the transmission evidence from node A to the target along a single path. Specifically, the transmission evidence from node A to the target with a bottleneck is 0.59, the transmission evidence from node A to the target with a smooth flow is 0.31, and the remaining 0.1 is assigned to an unknown state. Similarly, the transmission evidence from node B is extracted and derived using the same derivation. It is assumed that the transmission evidence from node B to the target with a bottleneck is 0.4, the transmission evidence from node B to the target with a smooth flow is 0.5, and the confidence level for the unknown state is 0.1. If node A or node B is not the source node, the derived evidence generated in the previous round of reasoning will continue to participate in this round of propagation as the corresponding node evidence.

[0052] The Dempster combination rule was applied to fuse the two sets of transitive evidence mapped to the target node. The orthogonal conflict factor, calculated as 0.419, was obtained by multiplying and summing contradictory propositions. Using the conflict factor, the normalization constant C was determined to be 1.721. The aggregated value of the target proposition was then calculated, yielding a target derived confidence level of 0.577 for the congestion state. Similarly, the target derived confidence level for the smooth state was calculated to be 0.406, and the target derived confidence level for the unknown state was 0.017. This generates a complete distribution of derived evidence for the virtual non-source node representing the congestion risk in the granulation chamber.

[0053] As one possible implementation, confidence weights are used to attenuate and redistribute high-conflict reliability during the fusion process, resulting in a global reliability distribution and conflict coefficient. Specifically, the conflict factors between each pair of source node evidence and derived evidence are calculated according to the weighted evidence combination rules, and the sum of all conflict factors is taken as the conflict coefficient. It is then determined whether the conflict coefficient exceeds a preset conflict threshold; if so, high-conflict reliability is identified. Evidence combinations identified as having high-conflict reliability are multiplied by an attenuation factor, which is a negative exponential function of the inverse of the corresponding confidence weight. The reliability quality released after conflict reliability attenuation is redistributed proportionally according to the average reliability proportion of the remaining unconflicted propositions in the evidence before fusion. The reliability quality is redistributed to the unconflicted propositions, and the attenuated residual conflict reliability quality is assigned to the unknown frame. After normalization, the global reliability distribution for the current production condition is obtained.

[0054] Assuming that the source node evidence weights have been assigned after mutual information value evaluation... Equals 0.8, derived evidence weight The value equals 0.4. At this point, the distributions of the two pieces of evidence to be fused across the normal and fault state spaces are as follows: First normal state confidence... Equals 0.9, first fault state reliability Equal to 0.1; Second normal state reliability Equal to 0.2, second fault state reliability It equals 0.8. The conflict factor is calculated by extracting the cross-products supporting different propositions according to Dempster's combination rules, specifically the first conflict factor. Equal to 0.72, the second conflict factor It equals 0.02. Adding the two together gives the overall conflict coefficient. The value is 0.74. The preset conflict determination threshold ranges from 0.5 to 0.7, and here the preset conflict determination threshold is set to 0.6. Because the calculated overall conflict coefficient is greater than the threshold, it is determined that a high conflict confidence scenario has occurred due to the contradiction of multimodal sensors. Attenuation processing is required to avoid the distortion of the fusion result caused by the direct normalization of high conflict evidence.

[0055] Combining the exponential discount approach used to handle highly conflicting evidence, a decay factor is constructed using the reciprocal of the confidence weight. The decay factor is calculated as follows:

[0056] in, It is the attenuation factor; This is the attenuation rate control coefficient; , where is the confidence weight.

[0057] Once intervention is triggered, the attenuation factor is calculated using the formula above. An attenuation rate control coefficient is then set. The confidence weight is 1, and the product term that causes the largest conflict is 0.72, with a lower confidence weight. The attenuation factor is calculated to be 0.082, suppressing the conflict quality to 0.059; similarly, the conflict term with a confidence weight of 0.02 is given a higher confidence weight. The baseline was reduced to 0.0057. After this reduction, the original high level of disordered conflicting quality (0.74) was compressed to a total of 0.0647. The average reliability of the propositions themselves was calculated; the average reliability for the normal state was 0.55, and the average reliability for the faulty state was 0.45. The difference before and after compression, i.e., the released redundant reliability of 0.6753, was then fed back and accumulated to the non-conflicting propositions in a ratio of 0.55 to 0.45. The non-conflicting normal propositions are: The fault items that did not conflict are Therefore, the reliability under normal conditions is obtained as follows: The reliability of the fault state is Assigning the attenuated residual conflict quality to the unknown frame yields a confidence level of 0.065 for the unknown state. This results in a global confidence distribution: 0.551 for the normal state, 0.384 for the faulty state, and 0.065 for the unknown state. Since the source node evidence weight is higher than the derived evidence weight, the fusion result after conflict attenuation retains more normal propositions from high-confidence source evidence while transferring the remaining conflict quality to the unknown frame, thus reducing the misleading influence of low-confidence conflict evidence on the global judgment.

[0058] S4. Calculate the expected utility and issue instructions.

[0059] In an optional embodiment, a basic asymmetric utility function matrix is ​​preset, and adjusted based on the conflict coefficient using a risk preference function. The global confidence distribution is converted into a Bayesian decision probability distribution. Combining the aforementioned Bayesian decision probability distribution with the adjusted utility function matrix, the Bayesian expected utility is calculated for candidate control strategies. The candidate control strategy with the largest Bayesian expected utility value is selected, and production control instructions are generated and issued.

[0060] Specifically, a two-dimensional array is initialized as the basic asymmetric utility function matrix. The utility penalty for downtime caused by false alarms is set to -50, and the utility penalty for equipment damage caused by missed alarms is set to -1000. Using an exponential risk preference function, a monotonically decreasing risk preference output value with the conflict coefficient as the independent variable is calculated to obtain the risk coefficient. The positive utility terms in the basic asymmetric utility function matrix are multiplied by the risk preference output value, and the negative utility terms are divided by the risk preference output value to complete the matrix adjustment. When the global reliability distribution contains unknown frames or composite proposition reliability quality, the reliability quality of unknown frames or composite propositions is first evenly distributed according to the number of production state propositions, or allocated to the corresponding single production state propositions according to a preset prior proportion, to obtain a Bayesian decision probability distribution consistent with the column dimensions of the utility function matrix. For three candidate control strategies—deceleration, shutdown, and maintaining the status quo—the NumPy library's computation module is used to calculate the sum of the product of the Bayesian decision probability distribution vector and the row vector corresponding to each candidate control strategy in the adjusted utility function matrix. The summation result is the Bayesian expected utility. The index value with the largest Bayesian expected utility is returned, mapping to the optimal control strategy. Based on the aforementioned strategy, low-level communication is established with the programmable logic controller (PLC), and the production control command for adjusting the frequency of the pellet mill feed inverter is written to the corresponding register address to complete the control operation.

[0061] As one possible implementation, the basic asymmetric utility function matrix is ​​adjusted based on the conflict coefficient using a risk preference function. Specifically, a basic asymmetric utility function matrix is ​​established with candidate control strategies as rows and production state propositions as columns. Matrix elements represent the baseline utility value of executing the corresponding strategy under a specific production state. Control strategies matching the current production state correspond to positive baseline utility values, while strategies for erroneous shutdowns, missed fault reports, or excessive interventions correspond to negative baseline utility values. Furthermore, the absolute value of the negative baseline utility corresponding to a missed fault report or equipment damage strategy is greater than the absolute value of the negative baseline utility corresponding to a normal malfunction. A risk preference function is constructed, which is a monotonically decreasing exponential function with the conflict coefficient as the independent variable. The positive baseline utility values ​​in the basic asymmetric utility function matrix are multiplied by the output value of the risk preference function, and the negative baseline utility values ​​are divided by the output value of the risk preference function to obtain the adjusted utility function matrix.

[0062] Initialize a basic utility function matrix U of size 3×2 in the controller memory space, and set the three rows to correspond to the strategies respectively. Maintain rated feed rate, Reduce speed by 20% Stop and reverse to clean; two columns correspond to production status Smooth production and Overload stall. The baseline utility value for filling needs to ensure that the penalty for failure far outweighs the incentive for gain, for example... , ; , ; , .

[0063] Based on the theoretical derivation of the risk penalty model, this invention constructs a risk preference function by combining the conflict coefficient, and its calculation method is as follows:

[0064] in, Output the risk preference value; This is the preference sensitivity coefficient; This represents the conflict coefficient.

[0065] In this invention, the preference sensitivity coefficient The range is 1 to 2, and a preference sensitivity coefficient is set. The input conflict coefficient is 1.5. The value is 0.65. The calculated risk preference output value for the current state is... .

[0066] Reshape the fundamental matrix using this output value according to the risk-averse control principle. For cases with a positive baseline utility value, multiply all by... To achieve a proportional contraction and reduce blindly optimistic expectations stemming from high conflict, such as the adjusted... , Conversely, for cases with negative baseline utility values, the value is divided by a factor less than 1. This amplifies potential uncontrollable losses, such as the penalty for failures in adjusted aggressive strategies. At the same time, mild penalties for prevention strategies Losses incurred due to accidental shutdowns caused by reverse shutdowns during smooth production will also be treated as negative baseline utility. By performing a two-way adjustment operation of positive utility shrinkage and negative utility amplification, a new set of conservative and mathematically computable utility function matrices was generated in a data environment with a conflict level as high as 0.65.

[0067] As one possible implementation, after converting the global confidence distribution into a Bayesian decision probability distribution, the Bayesian expected utility of candidate control strategies is calculated by combining the aforementioned Bayesian decision probability distribution with the adjusted utility function matrix. The candidate control strategy with the largest Bayesian expected utility value is selected, and production control instructions are generated and issued. Specifically, the global confidence distribution undergoes probability transformation processing, allocating the confidence quality assigned to unknown frames or composite propositions to the corresponding single production state propositions, resulting in a Bayesian decision probability distribution. For each candidate control strategy, the state utility values ​​corresponding to the strategy in the adjusted utility function matrix are multiplied item by item with the corresponding state probabilities in the Bayesian decision probability distribution, and the results are summed to obtain the Bayesian expected utility of the candidate control strategy. The Bayesian expected utilities of all candidate control strategies are sorted in descending order, and the top-ranked candidate control strategy is selected. Production control instructions containing equipment action parameters and execution time are generated based on the selected candidate control strategy and issued to the actuators of the corresponding production equipment via industrial Ethernet.

[0068] Extract the global confidence distribution and convert it into a Bayesian decision probability distribution with the same column dimensions as the utility function matrix. If this global confidence distribution contains unknown frames... The reliability quality can be converted as follows:

[0069] in, For the first Bayesian decision probabilities for a production state proposition; The overall reliability quality of propositions regarding production status; For the reliability quality of the unknown framework; The number of propositions related to the production status.

[0070] In cases involving other compound propositions, the reliability quality of the compound propositions can be evenly distributed among the individual production state propositions they contain. For example, the current time is obtained after the above probability transformation: , This set of probabilities is used as the Bayesian decision probability distribution and substituted into the utility function matrix calculated upstream. In the middle. The expected value of each strategy combination is calculated by multiplying and accumulating: for each strategy... Maintaining the rated feed rate, Bayesian expected utility For strategy Running at a 20% speed reduction, expected utility For a safe and conservative strategy Shutdown reverse cleaning, expected effect This indicates the overall risk level faced by each intervention pathway under the current threat of a highly disruptive failure with a 15% uncertainty.

[0071] After obtaining all values, the control module places the BEU set {-365.83, -23.76, -116.68} into a queue and performs a quick descending sort. -23.76 has the largest value in the negative utility game, therefore it is... The strategy was determined to be the best response. The central control module matched within the mapping library. The set of action rules associated with the strategy is transformed and assembled into machine-readable low-level PLC action parameters and delay markers. For example, the instruction explicitly sets the target frequency written to the inverter register to 40Hz, the ramp-down response time to 2.0s, and simultaneously reduces the steam control valve opening by 15% to prevent material adhesion. This combined instruction package is encapsulated in a Profinet industrial Ethernet communication frame and sent directly to the low-level actuator of the pellet mill control cabinet within a few hundred milliseconds via a star network to complete the control loop.

[0072] The experimental dataset consisted of 500 hours of continuous real-world operation records from a self-made pellet mill, including 50 real overload and stall anomaly events. To simulate discrepancies in multi-sensor observations, Gaussian white noise was artificially injected into the temperature sensor data stream, forcing a conflict between temperature-derived evidence and motor current source node evidence. The control group used a basic Gaussian mixture model combined with traditional Dempster combinatorial rules and a static Bayesian utility matrix for decision-making, while the experimental group used the full application, which included conflict confidence decay redistribution and utility function matrix adjustment. The evaluation metrics were the anomaly interception success rate and the total unplanned downtime.

[0073] Data shows that when the injected noise caused the conflict factor to reach 0.74, the control group, due to the lack of attenuation intervention during the reliability fusion process, had its normalized global reliability for overload blockage diluted to 0.25. This resulted in the Bayesian expectation assessment still judging maintaining the rated feed rate as the optimal strategy, leading to 45 actual blockages and a total unplanned downtime of 38.4 hours. The experimental group used a negative exponential function of the inverse of the confidence weight to compress the original high-conflict quality to 0.06 and perform redundant reliability feedback, significantly improving the reliability or Bayesian decision probability related to overload blockage and maximizing the expected utility of a 20% speed reduction. By amplifying the fault penalty utility through a preference sensitivity coefficient, the expected utility of a 20% speed reduction was calculated to be the highest, successfully intercepting 48 abnormal events and reducing the total unplanned downtime to 2.5 hours.

[0074] Reference Figure 3The diagram compares the metrics of the two methods. The left-hand bar represents the baseline levels of interception success rate and unplanned downtime, while the right-hand bar shows a significantly higher interception success rate and a significantly lower unplanned downtime. These differences reflect how the redistribution mechanism of the attenuation factor and the average reliability ratio of non-conflicting propositions can break the deadlock caused by conflicting multimodal sensors. By combining a monotonically decreasing exponential risk preference function with the conflict coefficient as the independent variable, a bidirectional penalty amplification operation is used to proportionally shrink expected returns and amplify potential losses. This enables the controller to deduce safety-guaranteed control commands even in uncertain data environments, effectively reducing the risk of severe equipment damage caused by misjudgments in traditional fusion methods.

[0075] This invention also discloses an intelligent control system for a feed production line, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent control method for a feed production line according to the present invention is implemented.

[0076] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0077] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for intelligent control of a feed production line, characterized in that, include: S1. Acquire multi-source heterogeneous sensor data during feed production, set key operating condition parameter thresholds as production status propositions, and generate initial basic confidence assignment functions for each production status proposition by combining a Gaussian mixture model constructed based on historical sensor data and a sliding window volatility based on real-time data; S2. Construct a time-series confidence knowledge graph with production operating condition parameters as nodes and pre-defined correlations between parameters as directed edges, and continuously update the conditional confidence function representing the strength of influence transmission between nodes in the time-series confidence knowledge graph based on real-time data; S3. Use the initial basic confidence assignment function as source node evidence, perform confidence propagation reasoning in the time-series confidence knowledge graph based on the conditional confidence function to generate derived evidence, calculate confidence weights using the mutual information values ​​between historical evidence source data after time window alignment and feature aggregation and finished feed batch quality indicators, and fuse source node evidence and derived evidence using weighted evidence combination rules to obtain global confidence distribution and conflict coefficient; S4. Preset the basic asymmetric utility function matrix. Adjust the basic asymmetric utility function matrix based on the conflict coefficient and the risk preference function to convert the global confidence distribution into a Bayesian decision probability distribution. Combine the aforementioned Bayesian decision probability distribution with the adjusted utility function matrix to calculate the Bayesian expected utility for candidate control strategies. Select the candidate control strategy with the largest Bayesian expected utility value and generate and issue production control instructions.

2. The intelligent control method for a feed production line according to claim 1, characterized in that, The generation of the initial basic reliability assignment function includes: iteratively training historical sensor data using the expectation-maximization algorithm to obtain the mean and covariance matrix of each Gaussian component of the Gaussian mixture model; calculating the standard deviation of real-time data within the current preset time window as the sliding window volatility; inputting the current real-time data into the Gaussian mixture model to calculate the posterior probability of each Gaussian component; performing dimensionless processing on the sliding window volatility to obtain the dimensionless volatility; calculating the normal adjustment coefficient and the abnormal adjustment coefficient based on the dimensionless volatility; and, according to the normal or abnormal attributes of the corresponding production state proposition, multiplying the posterior probability of the Gaussian component of the corresponding normal state proposition by the normal adjustment coefficient, and multiplying the posterior probability of the Gaussian component of the corresponding abnormal state proposition by the abnormal adjustment coefficient, and then performing normalization processing to obtain the initial basic reliability assignment function for the corresponding production state proposition.

3. The intelligent control method for a feed production line according to claim 1, characterized in that, The step of continuously updating the conditional confidence function representing the strength of influence transmission between nodes in the time-series confidence knowledge graph based on real-time data includes: acquiring the time series of sensor data from the source node and the target node within adjacent acquisition periods; calculating the time warping distance between the time series of sensor data from the source node and the time series of sensor data from the target node; using the time warping distance to calculate the influence attenuation coefficient between nodes; multiplying the preset basic conditional confidence function by the influence attenuation coefficient; and allocating the reduced confidence quality to the global unknown frame to obtain the updated conditional confidence function representing the strength of influence transmission between nodes.

4. The intelligent control method for a feed production line according to claim 1, characterized in that, The step of performing confidence propagation inference based on conditional confidence functions in a temporal confidence knowledge graph to generate derived evidence for non-source nodes includes: traversing the temporal confidence knowledge graph and identifying all non-source nodes without assigned initial evidence as target nodes; for each target node, extracting all predecessor node evidence pointing to that node and their corresponding conditional confidence functions, where predecessor node evidence includes the initial source node evidence of the predecessor node or derived evidence generated in the previous round of confidence propagation inference; calculating the product of each predecessor node evidence and its corresponding conditional confidence function, marginalizing and summing the state space of the predecessor node to obtain single-path transmission evidence mapped to the state space of the target node; and aggregating all single-path transmission evidence pointing to the same target node using combination rules to generate derived evidence for non-source nodes.

5. The intelligent control method for a feed production line according to claim 1, characterized in that, The process of obtaining the global reliability distribution and conflict coefficient includes: calculating the conflict factors between each pair of source node evidence and derived evidence according to the weighted evidence combination rules, and using the sum of all conflict factors as the conflict coefficient; determining whether the conflict coefficient is greater than a preset conflict threshold, and if it is, determining that there is high conflict reliability; multiplying the evidence combination item determined to have high conflict reliability by a decay factor, the decay factor being a negative exponential function of the inverse of the corresponding confidence weight; redistributing the reliability quality released after the conflict reliability decays to the unconflicted propositions according to the average reliability proportion of each remaining unconflicted proposition in the evidence before fusion, allocating the decayed residual conflict reliability quality to the unknown frame, and obtaining the global reliability distribution of the current production condition after normalization.

6. The intelligent control method for a feed production line according to claim 1, characterized in that, The adjustment of the basic asymmetric utility function matrix includes: establishing a basic asymmetric utility function matrix with candidate control strategies as rows and production state propositions as columns. The matrix elements represent the benchmark utility value of executing the corresponding strategy under a specific production state. Among them, the control strategy matching the current production state corresponds to a positive benchmark utility value, and the control strategy corresponding to the erroneous shutdown strategy, the missed fault reporting strategy, or the over-intervention strategy corresponds to a negative benchmark utility value. The absolute value of the negative benchmark utility corresponding to the missed fault reporting strategy or the equipment damage strategy is greater than the absolute value of the negative benchmark utility corresponding to ordinary malfunction. Construct a risk preference function, which is a monotonically decreasing exponential function with the conflict coefficient as the independent variable. Multiply the positive benchmark utility value in the basic asymmetric utility function matrix by the output value of the risk preference function, and divide the negative benchmark utility value by the output value of the risk preference function to obtain the adjusted utility function matrix.

7. The intelligent control method for a feed production line according to claim 1, characterized in that, The generation and issuance of production control instructions includes: performing probability transformation processing on the global confidence distribution, assigning the confidence quality allocated to the unknown frame or composite proposition to the corresponding single production state proposition, and obtaining a Bayesian decision probability distribution; for each candidate control strategy, multiplying and summing the state utility values ​​corresponding to each state in the adjusted utility function matrix of the strategy with the corresponding state probabilities in the Bayesian decision probability distribution, and obtaining the Bayesian expected utility of the candidate control strategy; sorting all candidate control strategies in descending order of their Bayesian expected utility, and selecting the candidate control strategy with the highest ranking; generating production control instructions containing equipment action parameters and execution time based on the selected candidate control strategy, and issuing them to the actuators of the corresponding production equipment via industrial Ethernet.

8. The intelligent control method for a feed production line according to claim 1, characterized in that, The acquisition of multi-source heterogeneous sensor data during the feed production process includes: using the OPCUA protocol library to connect to the programmable logic controller of the feed production line, collecting data from motor current sensors, equipment speed sensors, valve opening sensors, temperature sensors, humidity sensors, and vibration sensors in real time, and writing the collected data into a time-series database.

9. The intelligent control method for a feed production line according to claim 1, characterized in that, The construction of a time-series confidence knowledge graph, with production condition parameters as nodes and pre-defined relationships between parameters as directed edges, includes: sequentially adding the feeding auger speed as a graph node, adding the ring die roller differential speed as a graph node, adding the steam condensate volume as a graph node, adding the conditioning temperature as a graph node, and adding the risk of clogging in the pelleting chamber as a virtual graph node, and adding directed edges between nodes according to the control logic of the feed process flow, thus constructing the main framework of the time-series confidence knowledge graph.

10. An intelligent control system for a feed production line, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement an intelligent control method for a feed production line according to any one of claims 1-9.

Images