PID (Proportion Integration Differentiation) parameter setting method based on knowledge graph

Abstract

The invention discloses a PID (Proportion Integration Differentiation) parameter setting method based on a knowledge graph, belongs to the technical field of industrial automation control, and aims to solve the problem of difficulty in parameter setting caused by nonlinear characteristics of large inertia, large lag and the like of a controlled object in a complex industrial process. According to the method, dynamic response data and static attributes of equipment are synchronously collected by applying a step test signal, and after preprocessing and feature extraction, dynamic features and the static attributes are combined to generate structured query conditions; converting query conditions into vectors by utilizing a graph embedding technology, carrying out multi-dimensional similarity matching in the knowledge graph, and introducing semantic similarity to process an imprecise matching situation; a PID parameter is generated by adopting a single rule or fusing multiple rules according to a matching result, and finally a closed-loop self-learning mechanism is formed through parameter verification and effect feedback, so that the transformation from artificial experience dependence to knowledge-driven decision is realized, and the setting precision and the adaptive capacity of the system under complex working conditions are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation control technology, and in particular relates to a PID parameter tuning method based on knowledge graphs. Background Technology

[0002] In the field of industrial process control, the proportional-integral-derivative (PID) controller is the most widely used control algorithm due to its simple structure and strong robustness. The performance of a PID controller hinges on the proper setting of its parameters (proportional band, integral time, and derivative time), a process known as parameter tuning. Traditional tuning methods heavily rely on the personal experience of control engineers, who must manually adjust the parameters based on the dynamic response curve of the controlled object (such as a step response). This approach leads to low tuning efficiency and makes it difficult to accumulate, standardize, and pass on experience. To address this issue, some automated or semi-automated intelligent tuning software has emerged, such as methods that identify loop types based on the morphological characteristics of control loop curves and recommend tuning strategies accordingly. Furthermore, advanced process control and other high-end solutions aim to handle more complex control requirements. However, in industrial settings where high levels of automation are required, these solutions still have significant limitations.

[0003] Existing technical solutions have revealed several shortcomings in practical applications. First, traditional manual tuning methods are inefficient and inconsistent, failing to meet the needs of large-scale control system maintenance. Second, most intelligent tuning software can only handle control loops with simple characteristics. For complex objects commonly encountered in industrial production, such as those with large inertia, large hysteresis, nonlinearity, and multivariate coupling (e.g., furnace temperature, distillation column pressure), their tuning performance is poor, lacking in-depth reasoning capabilities. Furthermore, existing technologies generally lack self-learning mechanisms, failing to continuously learn and evolve from historical successes and failures, resulting in stagnant system tuning capabilities. Finally, while external expert services or APC systems can partially solve the problems, they suffer from high costs, long implementation cycles, and untimely responses. For example, Chinese patent application CN120368457A discloses a PID control system tuning method and device. However, this patent mainly relies on the analysis and matching of the dynamic curve shape of the loop. It fails to deeply integrate the static attributes of the device itself (such as equipment model and actuator type) with dynamic characteristics, and does not build a structured knowledge base containing rich domain knowledge for reasoning. As a result, it has poor adaptability when dealing with complex working conditions and does not have the ability to optimize tuning knowledge through continuous operation feedback. Summary of the Invention

[0004] The purpose of this invention is to provide a PID parameter tuning method based on knowledge graphs, which solves the technical problems of insufficient tuning accuracy and poor adaptability in the face of complex industrial process control loops with large inertia, large hysteresis and nonlinearity, due to reliance on fixed tuning strategies or limited human experience.

[0005] The purpose of this invention is to construct a structured knowledge base that integrates the static attributes of the device, the dynamic characteristics of the loop, and the expert experience rules, so as to achieve unified representation and associative reasoning of multi-source heterogeneous tuning knowledge, thereby generating PID parameter suggestions that are highly matched with the characteristics of the current controlled object under complex and variable operating conditions.

[0006] The purpose of this invention is to establish a parameter tuning closed-loop mechanism with self-evolution capabilities. By integrating the effect feedback of each tuning practice into the knowledge graph, the rule confidence and tuning strategy are continuously optimized, thereby achieving an improvement from single tuning decisions to long-term performance self-optimization.

[0007] This invention proposes a knowledge graph-based PID parameter tuning method. The method includes: acquiring data for the tuning task, including dynamic features collected by applying a step test signal and static attributes read from the device; generating query conditions based on the dynamic features and static attributes; extracting word vectors from the query conditions to obtain a query vector; performing similarity matching between the query vector and a word vector library, which is constructed based on a knowledge graph library through word vector extraction and contains entities related to the tuning project, the device to be tuned, and the relationships between them; and determining the PID tuning parameters based on the similarity matching results. By integrating dynamic response features and device static attributes, a comprehensive data foundation describing the characteristics of the controlled object is constructed, overcoming the limitations of relying solely on curve shape and providing data support for accurate tuning under complex operating conditions.

[0008] Preferably, the data acquisition for the tuning task includes preprocessing the collected raw data. This preprocessing employs a digital filtering algorithm to suppress high-frequency noise and removes outliers based on statistical principles. This effectively suppresses the interference of industrial noise and outliers on feature analysis, improving the accuracy and reliability of the subsequently extracted dynamic feature parameters, thereby ensuring the robustness of the tuning decision.

[0009] Preferably, this dynamic feature reconstructs the continuous system response model through curve fitting and calculates time-domain feature parameters based on the fitted model. By reconstructing the continuous system model through curve fitting, sampling noise is smoothed, enabling more accurate calculation of key time-domain features of the system (such as overshoot and settling time), and enhancing the credibility of similarity comparison with historical cases in the knowledge graph.

[0010] Preferably, this static attribute includes reading the device model from the equipment configuration file or electronic identifier and determining the loop type based on the controlled physical quantity and process requirements. By automatically identifying the device model and loop type, the tuning decision is associated with specific hardware characteristics and process requirements, making parameter recommendations more targeted and avoiding the incompatibility of general strategies on different devices and loops.

[0011] As a preferred approach, this method logically integrates dynamic feature sets and static attribute sets to form structured composite query conditions. Integrating multi-source heterogeneous information into structured query conditions provides rich context for intelligent retrieval of knowledge graphs, realizing the transformation from "data" to "query intent" and improving retrieval efficiency and accuracy.

[0012] Preferably, word vector extraction for query conditions involves using graph embedding technology to map elements in the query conditions to a low-dimensional vector space, and then generating a unified query vector through a combination function. By utilizing graph embedding technology to transform semantic query conditions into numerical query vectors, efficient computation and matching of unstructured knowledge in the vector space is achieved, laying the foundation for rapid similarity retrieval of large-scale knowledge bases.

[0013] Preferably, this similarity matching involves calculating the matching degree between the query vector and the rule vector across multiple dimensions, including dynamic features, device type, and loop type, and obtaining a comprehensive similarity score based on weighted summation. Through multi-dimensional weighted similarity matching, the correlation between dynamic performance, hardware configuration, and control objectives is comprehensively measured, enabling the matching results to more fully reflect the overall fit between current tuning requirements and historical rules.

[0014] As a preferred approach, this similarity matching also incorporates a semantic similarity metric to handle imprecise matches between query conditions and rule conditions. The introduction of a semantic similarity metric enables the system to identify and handle imprecise matching cases such as "moderate overshoot" and "significant overshoot," enhancing the fuzzy reasoning capabilities of the knowledge graph and its flexibility in practical applications.

[0015] Preferably, this method determines PID tuning parameters based on similarity matching results, including determining whether there are significantly high-matching rules. If so, a single rule parameter is adopted; otherwise, a rule fusion strategy is initiated to generate parameters. By determining whether there are significantly high-matching rules, this method achieves rapid adoption of a single rule when a clear best practice exists, while initiating a fusion strategy when multiple rules compete, thus balancing tuning efficiency and decision robustness.

[0016] Preferably, this rule fusion strategy calculates a weighted average of the parameters of multiple candidate rules, with the weights based on the similarity score or confidence level of each rule. This weighted fusion strategy, based on similarity or confidence level, can comprehensively utilize the experience of multiple related rules to generate a more balanced set of PID parameters that are more likely to adapt to subtle differences in the current system, thus improving the adaptability of the parameter scheme.

[0017] The beneficial effects of this invention are mainly reflected in the following aspects: This invention deeply integrates dynamic response characteristics and equipment static attributes through a knowledge graph, enabling parameter tuning decisions to move beyond isolated curve morphology analysis and instead place them within a specific context, such as equipment model and loop type, for comprehensive reasoning. This multi-dimensional and structured analysis method can more comprehensively characterize the controlled object's features. Therefore, when facing complex operating conditions common in industrial processes, such as large inertia, large lag, and nonlinearity, it can recommend more accurate and adaptable PID parameters compared to traditional methods or single-curve analysis techniques, fundamentally improving the stability, response speed, and steady-state accuracy of the control loop.

[0018] The multi-condition similarity matching mechanism based on graph embedding introduced in this invention can not only perform precise matching, but also achieve intelligent association and reasoning under imprecise conditions through semantic similarity measurement. This enables the system to handle fuzzy concepts such as "moderate overshoot" and "significant overshoot," possessing a certain degree of analogical reasoning ability and more closely resembling the thinking patterns of human experts. Simultaneously, the rule fusion strategy ensures that, in the absence of absolute matching rules, it can synthesize the experience of multiple high-similarity rules to generate robust parameter schemes. This intelligent decision-making mechanism reduces reliance on single, fixed tuning rules, giving the system a strong adaptability to handle diverse and uncertain operating conditions.

[0019] This invention constructs a complete closed loop from parameter decision-making to effect verification. The effect of each tuning is fed back to the knowledge graph to update the confidence of relevant rules. Successful experiences are reinforced, while unsuccessful experiences are marked for analysis. This closed-loop self-learning mechanism enables the system to continuously optimize its knowledge base and decision-making logic from historical practice, realizing the transformation from a "one-time tuning tool" to a "continuously learning expert system." This not only effectively preserves, standardizes, and reuses the individual experience of senior engineers, but also allows the system's tuning capabilities to continuously evolve with runtime, effectively solving the core pain point of traditional methods' difficulty in knowledge transfer and optimization. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0021] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0022] according to Figure 1 As shown in Example 1, the entire process of automatic PID parameter tuning based on knowledge graphs and dynamic-static feature fusion in an industrial automation system is described in detail. The process begins with the start of the tuning task, sequentially going through preliminary stages such as test signal application and data synchronization acquisition, and data preprocessing. Subsequently, the processing tasks of the dynamic feature extraction channel and the static attribute acquisition channel are executed in parallel. The dynamic channel completes dynamic feature extraction, curve fitting, and time-domain feature calculation, while the static channel reads the device model and determines the loop type. The system then combines the output results of the two channels into query conditions, generates a query vector Q using graph embedding technology, and determines whether a significantly high-matching rule exists based on the multi-condition similarity matching results—if it exists, a single rule parameter is adopted; otherwise, a rule fusion strategy is initiated. Finally, the output parameters and reasoning basis are presented. After parameter distribution and verification, the rule confidence is increased or decreased based on whether the tuning effect meets the standards, thus forming a closed-loop tuning mechanism from parameter decision-making to effect feedback.

[0023] In industrial automation production scenarios, when the PID parameters of a control loop need to be retuned, the system first triggers the "Start Tuning Task" process. This step marks the beginning of the tuning cycle, and the core task is to initialize the PID parameter tuning module. The initialization process includes loading a pre-trained and complete knowledge graph library. This knowledge graph stores the complex relationships between equipment models, control loop types, historical tuning cases, and their corresponding PID parameter rules in a graph structure. Simultaneously, the system activates its built-in digital signal processing unit, dynamic feature extraction engine, and rule inference engine, allocating computing resources and establishing internal communication links. This initialization phase ensures that all software modules and hardware interfaces are ready to respond to subsequent real-time data acquisition, signal excitation, and knowledge-based parameter decision requests, building a stable execution environment for the entire tuning process.

[0024] After initialization, the system immediately enters the "apply step test signal and synchronously acquire data" phase. The core operation of this phase is to apply a controllable step setpoint change to the actuator of the control loop. This step signal serves as the system excitation source, aiming to disrupt the original equilibrium state of the control loop and stimulate its dynamic response characteristics. The rise time and amplitude of the step signal need to be pre-configured according to the characteristics of the controlled object to ensure that the response process contains rich dynamic information without causing system instability. Simultaneously with the application of the step signal, the system's high-precision data acquisition unit starts synchronously at a millisecond-level sampling rate. This unit reads the instantaneous values ​​of the controlled variable in real time through the analog input module or industrial bus, forming time-series data. Simultaneously, the system reads static attribute data from the equipment configuration file, electronic nameplate, or equipment description file, including but not limited to the actuator's manufacturer and model code, the sensor's measurement range and accuracy class, and the type of controlled variable in the loop. This synchronous acquisition mechanism of dynamic signals and static attributes ensures strict timestamp alignment, providing a time-consistent data foundation for subsequent feature extraction and attribute correlation analysis.

[0025] The raw data collected typically contains inherent high-frequency electromagnetic interference, sensor quantization noise, and transmission link noise from industrial environments, thus requiring data preprocessing. The preprocessing process mainly includes two steps: digital filtering and outlier removal. Digital filtering often employs finite-length unit impulse response (FIR) filters or infinite-length FIR filters based on sliding windows, such as first-order low-pass filters. Their cutoff frequency is set according to the expected bandwidth of the controlled object to attenuate high-frequency noise components without significantly distorting the main dynamic characteristics of the system response. For non-stationary noise, wavelet transform thresholding can be used to separate noise from useful signals by thresholding wavelet coefficients at a specific scale. Outlier removal is usually based on statistical principles, such as using the Laida criterion or interquartile range (IQR) to identify and remove outliers that significantly deviate from the data trend. After preprocessing, the resulting smooth data sequence more accurately reflects the dynamic characteristics of the system, laying the foundation for subsequent precise feature extraction.

[0026] After data preprocessing, the process flows in parallel into two processing channels: a dynamic feature extraction channel and a static attribute acquisition channel. The first step in the dynamic feature extraction channel is "dynamic feature extraction." This step performs time-domain analysis on the preprocessed step response curve to extract key characteristic parameters that characterize the system's dynamic performance. These features typically include: the rise time of the response curve, i.e., the time required for the output to rise from 10% to 90% of the steady-state value; the peak time, i.e., the time required for the response to reach its first peak; the overshoot, i.e., the percentage by which the maximum response value exceeds the steady-state value; and the settling time, i.e., the shortest time required for the response to enter and remain within an allowable error band near the steady-state value. For non-oscillating systems, the equivalent time constant and pure time delay may also be of interest. Quantifying these features provides objective indicators for evaluating the transient performance of the system under the current PID parameter control.

[0027] Next, curve fitting is performed. To reconstruct a continuous system response model from discrete sampling points and facilitate theoretical analysis, the system uses a mathematical model to fit the extracted feature data points. Commonly used fitting models include first-order plus pure time delay models or second-order plus pure time delay models. The fitting process typically uses optimization algorithms such as the least squares method to find a set of model parameters that minimizes the sum of squared errors between the model output and the actual sampled data. The fitted transfer function model not only smooths out sampling noise but, more importantly, provides a simplified mathematical model for the system, facilitating integration with standard tuning methods.

[0028] Based on the fitted curve, the system performs "calculation of time-domain features." This step involves a second, more precise calculation of the initially extracted features. For example, the fitted model can more accurately calculate the overshoot and settling time without noise influence. These model-validated time-domain features constitute the core dataset describing the system's dynamic performance and are crucial for determining the similarity between the system's dynamic behavior and historical cases in the knowledge graph during subsequent rule matching.

[0029] While the dynamic channel operates, the static attribute acquisition channel performs "static attribute acquisition" in parallel. This process first "reads the device model." By parsing the device description file or querying the device's electronic identification, the system accurately obtains the manufacturer and model information of the actuator, sensor, and controller. Different device models differ in hardware characteristics such as dynamic response speed, linearity, dead zone, and resolution. These differences directly affect the optimal range of PID parameters and tuning strategies.

[0030] Next, the system determines the loop type. Loop types are categorized based on the controlled physical quantity and process requirements. For example, temperature control loops typically exhibit large inertia and significant hysteresis; during tuning, the integral action needs to be carefully considered to eliminate steady-state error, and derivative action may be introduced to mitigate hysteresis effects. In contrast, flow control loops usually have rapid response and high noise levels; during tuning, the proportional action is strong, the integral time is short, and derivative action is generally avoided to prevent amplifying noise. Determining the loop type allows the system to focus its search on the subset of tuning rules in the knowledge graph most relevant to that type of loop, improving matching efficiency and accuracy.

[0031] After the dynamic feature extraction channel and the static attribute acquisition channel output their results, the system enters the "combined query condition" stage. This stage logically integrates the temporal feature set produced by the dynamic channel with the attribute set produced by the static channel, forming a structured, knowledge graph-oriented composite query condition. This query condition comprehensively reflects the identity characteristics and current performance status of the controlled object, providing a multi-dimensional search context for subsequent intelligent retrieval.

[0032] After the query conditions are combined, the process enters the graph embedding vectorization stage to generate the query vector. Entities and relationships in the knowledge graph are first mapped to a low-dimensional, dense vector space using graph embedding technology. Each entity is represented by a high-dimensional vector, and the relationships between entities are represented by specific operations in the vector space. The system merges the elements in the composite query conditions into a unified, high-dimensional query vector Q using a specific combination function. This query vector Q numerically encodes the entire query intent of this tuning task.

[0033] The query vector Q is fed into a vector database or similarity calculation engine and compared with all stored tuning rule vectors in the knowledge graph. Similarity calculation typically uses metrics such as cosine similarity or Euclidean distance. The matching process is multi-dimensional: it calculates not only the matching degree between the query vector Q and the rule vectors on static attributes such as device model and loop type, but also the similarity on dynamic features such as overshoot and settling time. The system may assign different weights to features of different dimensions and calculate a comprehensive similarity score through weighted summation or more complex machine learning models. Finally, the system selects the rules with the highest comprehensive similarity to form a candidate rule set.

[0034] Based on the similarity matching results, the system makes a decision. A significantly high match is typically defined as a rule whose similarity score is much higher than other rules. If this is determined, the process proceeds to the branch that adopts a single rule parameter. The system directly uses the PID parameter associated with the high-matching rule as the recommended value for this tuning. Simultaneously, the system generates reasoning, explaining the rule's origin, its successful application records in historical similar cases, and the similarity details of this match.

[0035] If the judgment is negative (i.e., there are no rules significantly higher than others, but multiple candidate rules with high similarity exist), the process enters the rule fusion strategy branch. Under this branch, the system selects the top N rules based on similarity ranking. Parameter fusion strategies may include: weighted averaging of the Kp, Ti, and Td parameters of the N rules, with the weights representing the similarity scores between each rule and the query vector Q; or, employing more complex fusion algorithms, such as case-based reasoning, to adjust the parameters to adapt to subtle differences in the current system. Rule fusion aims to comprehensively utilize the experience of multiple related rules to generate a more robust set of PID parameters that may be better suited to the current situation.

[0036] Regardless of whether the PID parameters are determined by adopting a single rule parameter or a rule fusion strategy, the process proceeds to the output parameter and reasoning stage. The system distributes the final PID parameter set to the target controller via configuration tools or communication protocols, completing the online parameter update. Simultaneously, the system generates a structured tuning report, detailing the adopted parameter values, their sources, matching similarity, and the reasoning logic based on dynamic feature analysis and static attribute matching. This report provides engineers with a transparent explanation of the parameter settings and is used for auditing and knowledge traceability.

[0037] After the parameters are issued, the system initiates the "parameter issuance and verification" process. Once the new parameters take effect in the controller, the system typically applies a test signal again and collects the response data of the control loop. By analyzing the response curves of this verification test, the control effect of the new parameters is evaluated. The evaluation metrics mainly include stability, response speed, and steady-state accuracy.

[0038] Based on the verification results, the judgment criteria are based on preset performance indicator thresholds. If the effect meets the standard, the system will adjust the confidence weight of the adopted rules or the main rules participating in the fusion upwards in the knowledge graph. Confidence is metadata in the knowledge graph that represents the reliability of a rule. An increase in confidence means that the rule is more likely to be preferentially recommended in similar tuning tasks in the future.

[0039] If the verification result fails to meet the standard, the system will reduce the confidence weight of the applied rule. In addition, cases that fail to be calibrated may be marked and stored in a special area of ​​the knowledge graph for subsequent analysis, or used to trigger revisions and optimizations of the knowledge graph under certain conditions, such as modifying relevant rules or marking usage restrictions after accumulating enough counterexamples.

[0040] The system completes all steps of this tuning task, and the PID parameter tuning module returns to standby mode, awaiting the triggering of the next tuning task. The entire process forms a complete closed loop from data acquisition, feature analysis, knowledge matching, parameter decision-making to effect verification and knowledge feedback, demonstrating an intelligent PID parameter tuning method based on data and knowledge.

[0041] Example 2 according to Figure 2 As shown, this embodiment details a specific implementation method for parameter tuning of a proportional-integral-derivative (PID) controller applied in an industrial process control system. The entire process begins with data acquisition from the control system to be tuned. The system applies a standard step test signal to the setpoint channel of the control system through a preset communication interface. Simultaneously, the system begins high-speed synchronous acquisition of instantaneous data of process variables, setpoints, and manipulated variables. This data is transmitted in real time to the system's data buffer via a fieldbus or data acquisition card, forming a set of raw data sequences strictly corresponding to time.

[0042] The acquired raw data sequence first enters the data preprocessing stage. The preprocessing module loads various digital filtering algorithms, such as moving average filtering or low-pass filtering, to process the raw data sequence. This processing aims to suppress unavoidable high-frequency electromagnetic interference and measurement noise in the field environment, while preserving the true trajectory of the control system's dynamic response to the greatest extent possible. After filtering, the data curves, which might have contained spikes, become smooth and continuous, laying the foundation for subsequent accurate analysis.

[0043] Next, the system enters the feature extraction stage. The feature extraction module calls a predefined curve fitting function library, which contains various typical function models used to describe dynamic systems. The module uses an optimization algorithm to find a fitting function that can approximate the preprocessed data sequence with minimum error. Once the optimal fitting function is determined, the system no longer directly processes the original data points, but instead operates on the continuous and parameterized ideal response curve represented by the fitting function.

[0044] Based on the fitted curves, the system automatically calculates a set of time-domain characteristic parameters that quantify the system's dynamic performance. The calculation process includes: determining the time interval required for the response to reach a specific proportion of the steady-state value from the initial state, i.e., the rise time; identifying the ratio of the maximum positive deviation exceeding the steady-state value to the steady-state value itself during the response process, i.e., the overshoot; analyzing the time required for the response to enter and remain within the allowable error band of the steady-state value, i.e., the settling time; evaluating the attenuation ratio between two adjacent oscillation peaks in the same direction, i.e., the attenuation ratio; and measuring the residual deviation between the response and the target setpoint after stabilization, i.e., the steady-state error. These calculated features correspond to the curve feature library definitions shown in Table 1, constituting a numerical description of the dynamic characteristics of the control system.

[0045] Table 1 Curve Feature Library After dynamic feature extraction, the system reads the static attribute information related to the control loop from the project configuration file. This information mainly includes device identification data such as the equipment model and manufacturer information of the key actuators used in the loop, and its typical classification is shown in Table 2: Table 2 Equipment Parts Database The control function type of this loop in the process flow diagram is also specified, with common types and characteristics shown in Table 3, such as temperature control loop, pressure control loop, or flow control loop. The dynamic characteristic parameter set and the static attribute information set are combined to form a complete comprehensive query condition describing the current tuning requirements.

[0046] Table 3 Loop Type Library At this point, the process enters the core matching stage based on the knowledge graph. The system accesses a pre-built and stored tuning knowledge graph, which organizes a large number of tuning rules in a graphical structure. Each rule has a condition part and a conclusion part. The condition part defines the applicable dynamic characteristic range, device type, and loop type, while the conclusion part contains verified controller parameter values ​​such as proportional band and integral time. To illustrate the rule structure and matching logic in the knowledge graph, a typical heating furnace temperature control loop is used as an example. Its basic information is shown in Table 4, and the corresponding expert experience rules are shown in Table 5.

[0047] Table 4 Example of basic circuit information To achieve efficient retrieval within a large-scale rule base, the system employs graph embedding technology. This technology maps nodes and relationships in the knowledge graph to a low-dimensional continuous vector space. The system first transforms the comprehensive query conditions for the current tuning requirements into a fixed-dimensional query vector using the same graph embedding model. Simultaneously, the conditional portion of each rule in the knowledge graph has been pre-processed and stored as a corresponding rule vector.

[0048] Subsequently, the system initiates multi-condition similarity matching calculation. This calculation is not a single comparison, but is decomposed into three independent dimensions performed asynchronously. The system calculates the projection distance between the query vector and a certain rule vector in the subspace defined by dynamic features, and the reciprocal of this distance is converted into the feature matching degree. Similarly, the system calculates the device matching degree in the subspace defined by device type information, and the loop matching degree in the subspace defined by loop type information.

[0049] It is worth emphasizing that the matching algorithm of this system incorporates semantic similarity measurement. For example, when the rule condition requires "moderate overshoot" while the query condition is "relatively large overshoot," the algorithm can calculate a non-zero similarity between "moderate" and "relatively large" based on their relative positions in the word vector space, rather than performing a strict Boolean judgment. This mechanism enables the system to retrieve rules that are not perfectly matched but have a certain degree of relevance.

[0050] After obtaining the matching scores for the three sub-items, the system performs a weighted sum of these three scores according to preset weighting coefficients, resulting in a comprehensive similarity score between zero and one. This score represents the overall degree of match between the current query condition and the rule condition. The system traverses all rules in the knowledge graph, completes the similarity scoring for each rule, and generates a preliminary candidate rule list sorted in descending order of this score.

[0051] Next, the system introduces two attributes for in-depth decision-making: rule priority and rule confidence. Rule priority is a static value preset by domain experts when rules are added to the database, used to identify the general applicability level of the rule. Rule confidence is a dynamically maintained statistical value that records the frequency with which the rule has historically achieved good results in actual operation. For example, in Table 5, rule E1 is assigned a "high" priority, indicating that the oscillation problem it addresses is considered a critical performance defect in the control system that needs to be addressed first.

[0052] Table 5 Examples of Expert Experience For each rule in the initial candidate list, the system takes its combined similarity score, priority value, and confidence value as input and feeds them into a weighted scoring function. This function outputs a score representing the final priority of the rule. The system then reorders the candidate rule list based on this final score, placing rules with high similarity, high priority, and a good historical record of validity at the top of the list.

[0053] The system then processes the final list of candidate rules after sorting. The rules in the list may offer the same or different parameter suggestions. The system checks if the final score of the top-ranked rule is significantly higher than that of the subsequent rules. If a significant advantage is found, the system decides to adopt the suggested parameters of that single rule.

[0054] If the top-ranked rules have very similar final scores, the system initiates a rule fusion strategy. The fusion process may involve calculating a weighted average of the parameter values ​​suggested by these rules, with the weighting coefficients proportional to the final score or confidence level of each rule. Through fusion, the system generates a new set of parameter suggestions that integrates the experience from multiple rules.

[0055] After determining the final recommended parameter values, the system prepares to output the results. The output includes not only the specific values ​​for parameters such as the proportional band and integration time, but also a structured document outlining the reasoning. This document is automatically generated and covers the main rule identifiers adopted, the key reasons for rule matching, the historical confidence levels of the rules, and explanations of any conflicting recommendations between the rules.

[0056] The system uses a standardized industrial communication protocol to download the generated parameter suggestions to the target controller, such as the corresponding function block of a programmable logic controller or distributed control system. After the parameters are downloaded, the controller uses the new parameter values ​​to enter the operating state.

[0057] To evaluate the effectiveness of parameter tuning, the system again collected operating data of the control system under typical operating conditions, especially the changes in process variables. Based on the new operating data, the system repeated the previous feature extraction process to calculate a new set of dynamic performance indicators.

[0058] The system compares the new metrics with the original metrics before tuning, generating an objective tuning effectiveness evaluation report. The core of the evaluation report is to determine whether the performance metrics are improving in the expected direction.

[0059] Finally, the system performs a closed-loop self-learning update. The system stores all the data from this tuning task—including initial query conditions, adopted rules, issued parameters, verification data, and performance evaluation conclusions—as a data sample in the historical database. Based on the performance evaluation conclusions, the system updates the confidence attributes of relevant rules in the knowledge graph. Rules that are adopted and have good verification results have their confidence levels appropriately increased; rules that are adopted but whose verification results do not meet expectations have their confidence levels correspondingly decreased. The overall rule quality of the knowledge graph continuously evolves as the system continues to run. At this point, a complete parameter tuning closed-loop process is completed.

Claims

1. A PID parameter tuning method based on knowledge graphs, characterized in that, The method includes: Acquire data for the tuning task, including dynamic characteristics acquired by applying a step test signal and static attributes read from the device; Query conditions are generated by combining dynamic features and static attributes. The query conditions are used to extract word vectors to obtain query vectors; The query vector is matched with a word vector library for similarity. The word vector library is constructed based on a knowledge graph library through word vector extraction. The knowledge graph library contains entities of the project to be calibrated, entities of the device to be calibrated, and the relationships between them. Based on the similarity matching results, the PID tuning parameters are determined.

2. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The acquisition of data for the tuning task includes preprocessing the collected raw data. The preprocessing uses a digital filtering algorithm to suppress high-frequency noise and removes outliers based on statistical principles.

3. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The dynamic features are used to reconstruct the continuous system response model through curve fitting, and time-domain feature parameters are calculated based on the fitted model.

4. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The static attributes include reading the device model from the device configuration file or electronic identifier, and determining the loop type based on the controlled physical quantity and process requirements.

5. A PID parameter tuning method based on a knowledge graph according to claim 1, 3, or 4, characterized in that, The method logically integrates dynamic feature sets and static attribute sets to form structured composite query conditions.

6. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The word vector extraction of the query conditions is performed by using graph embedding technology to map the elements in the query conditions to a low-dimensional vector space, and then generating a unified query vector through a combination function.

7. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The similarity matching includes calculating the matching degree between the query vector and the rule vector in multiple dimensions such as dynamic features, device type and circuit type, and obtaining a comprehensive similarity score based on weighted summation.

8. A PID parameter tuning method based on a knowledge graph according to claim 1 or 7, characterized in that, The similarity matching also introduces a semantic similarity metric to handle imprecise matching between query conditions and rule conditions.

9. The PID parameter tuning method based on knowledge graphs according to claim 1, characterized in that, The method determines PID tuning parameters based on similarity matching results, including determining whether there are significantly high matching rules. If so, a single rule parameter is adopted; otherwise, a rule fusion strategy is initiated to generate parameters.

10. The PID parameter tuning method based on knowledge graphs according to claim 9, characterized in that, The rule fusion strategy calculates a weighted average of the parameters of multiple candidate rules, with the weights based on the similarity score or confidence level of each rule.

Images