An artificial intelligence-based industrial parameter adaptive control system

Abstract

This invention relates to the field of industrial process control technology and discloses an artificial intelligence-based adaptive control system for industrial parameters. The system includes a deviation acquisition module, a feature compression module, a segmentation processing module, a morphology recognition module, a transition analysis module, and a parameter update module. The deviation acquisition module collects deviation signals in real time; the feature compression module performs adaptive segmentation compression on the deviation signals, outputting a sequence of non-equally spaced deviation feature points; the segmentation processing module performs quality-sensitive weighted temporal segmentation based on the process operation stage; the morphology recognition module extracts multi-scale morphological features and maps segments to deviation morphological primitive types; the transition analysis module performs path matching on the observed sequence in a probability-weighted transition graph and calculates the correction urgency; the parameter update module determines the correction vector and step size based on the matching path, and updates the process model parameters after momentum smoothing. This invention achieves early perception and progressive correction of mismatch in slow time-varying models.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and more specifically, to an industrial parameter adaptive control system based on artificial intelligence. Background Technology

[0002] In the field of industrial process control, advanced control strategies such as model predictive control rely on the accuracy of the mathematical model of the controlled process. Slow time-varying factors are common in actual industrial processes, such as catalyst activity decay, heat exchanger fouling, and batch-to-batch variations in material properties. These factors cause continuous drift in the dynamic characteristics of the process, gradually leading to a mismatch between the initially established mathematical model and the actual process.

[0003] To address the aforementioned issues, adaptive control technology maintains control quality by adjusting model or controller parameters online. Common solutions include self-tuning control based on online model identification and model reference adaptive control based on a reference model. Existing adaptive control technologies typically employ a serial mechanism of "detecting performance degradation and triggering correction": first, the magnitude of the control deviation or performance index is monitored to see if it exceeds a preset threshold; only after confirming a mismatch does the parameter identification or controller adjustment process begin.

[0004] The above mechanism has two shortcomings. First, in the early stages of model mismatch development, the amplitude of the control deviation has not yet increased significantly, making it difficult for detection methods based on amplitude thresholds or performance index thresholds to effectively identify the mismatch, leading to its continued accumulation undetected. Second, the detection and correction processes are independent and executed sequentially, resulting in an inherent response lag between the occurrence of mismatch and the effectiveness of correction. For processes such as polymerization reactions, which require high precision in temperature and reaction rate control, this lag may lead to quality problems such as shifts in product molecular weight distribution. Therefore, an artificial intelligence-based adaptive control system for industrial parameters is proposed to address the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an industrial parameter adaptive control system based on artificial intelligence, which aims to solve the problems of insufficient early perception capability of existing adaptive control technology for slow time-varying model mismatch in industrial processes and response lag between detection and correction.

[0006] To achieve the above objectives, the present invention provides an industrial parameter adaptive control system based on artificial intelligence, including a deviation acquisition module, a feature compression module, a segmentation processing module, a morphology recognition module, a change analysis module, and a parameter update module.

[0007] The deviation acquisition module collects the difference between the process set value and the actual measured value in real time as a deviation signal to obtain the tracking error of the controlled process to the set value.

[0008] The feature compression module performs adaptive segmented compression on the deviation signal and outputs a non-equal interval deviation feature point sequence composed of fixed period statistical points and morphological event trigger points. While retaining the statistical features of the deviation, it captures the moment when the deviation morphology changes, providing a data foundation for subsequent time series analysis.

[0009] Furthermore, the feature compression module operates at a fixed time period. By statistically analyzing the deviation signal and calculating the mean, standard deviation, and range of the deviation in each period as a feature point, the central tendency, dispersion, and fluctuation range of the deviation in that period can be obtained. Simultaneously, the current period is immediately truncated and a feature point is generated when any of the following conditions are met: the deviation value crosses the upper limit of the preset attention band. or lower limit The signs of the rate of change of the average deviation between two adjacent cycles are opposite, and a change signal is received during the process operation phase.

[0010] The above conditions correspond to the abnormal out-of-bounds deviation amplitude, the reversal of the deviation change trend, and the switching of process stages, respectively. By responding to the above events in real time, the moment of sudden change in the deviation pattern is accurately captured, thereby forming a non-equal interval deviation feature point sequence composed of fixed periodic statistical points and morphological event trigger points.

[0011] The segmentation module performs quality-sensitive weighted temporal segmentation on the deviation feature point sequence according to the process operation stage, dividing the deviation feature point sequence into several segments with consistent internal morphology, so that the different sensitivities of deviation morphology changes in different operation stages can be reflected.

[0012] Furthermore, the segmented processing module acquires key process indicators, and when the value of the key process indicator is less than... The time division process is divided into initial stage and later stage. to The interval is divided into the middle period, in to The interval is divided into final stages, and the deviation data is then subjected to the first level of macroscopic segmentation based on the inherent dynamic stages of the process. Within each stage, dynamic programming is used to find the set of segmentation points that minimize the cost function value within the segment. Calculate using the following formula: ; in The variance of the mean sequence within a segment represents the degree of consistency of the mean deviation within a segment; The standard deviation is the series variance, which characterizes the degree of consistency in the intensity of deviation fluctuations within a segment; The trend consistency index is the proportion of the total duration of the segment in which the linear fit slope of the mean sequence within the segment remains consistent with the sign of the slope. It is used to measure whether the direction of deviation change within the segment is stable.

[0013] , , The weighting coefficient is set to a value in the initial stage. , , The value in the intermediate period is , , The value at the end of the period is , , ,and , , The relationship between the weighting coefficients mentioned above reflects the quality-sensitive weighting characteristic that the process is most sensitive to changes in the shape of deviations in the middle stage and relatively less sensitive to changes in the end stage.

[0014] For segments with a duration less than Furthermore, the cosine similarity with the morphological encoding vector of adjacent segments is greater than [value missing]. The segments are merged to avoid creating fragmented segments that are too short and have no substantial differences due to noise or transient disturbances, thus ensuring the rationality of the segmentation results in terms of morphology.

[0015] The morphology recognition module extracts multi-scale morphological features for each segment to form a morphological encoding vector, and maps the morphological encoding vector to a deviation morphological primitive type, thereby transforming the evolution process of the continuous deviation signal into a discrete morphological primitive sequence with clear physical meaning.

[0016] Furthermore, the coarse-grained feature extracted by the morphology recognition module is the slope of the linear fitting of the mean sequence of deviations within the segment. This reflects the overall trend and direction of the mean deviation within the segment; Slope of linear fit of standard deviation series This reflects the changing trend of the deviation fluctuation intensity within the segment; Correlation coefficient between mean and standard deviation This reflects the correlation between the deviation amplitude and the fluctuation intensity; The ratio of the number of zero-crossing deviations to the segment length This reflects the oscillation frequency characteristics of the deviation within the segment.

[0017] Fine-grained features are the last fragment The mean over the period of time and the initial Difference of mean over time This reflects the change in the level of deviation between the beginning and end of a segment; finally Standard deviation over time and initial Ratio of standard deviations over time This reflects the relative change in the intensity of fluctuations at the beginning and end of a segment; and the mean of the absolute values ​​of the second differences of the mean sequence within the segment. This reflects the stability of the rate of change of the deviation.

[0018] The morphological encoding vector is By combining the above coarse-grained and fine-grained features, the deviation morphology of the segment is comprehensively characterized from two scales: the overall trend and the local boundary.

[0019] Furthermore, the deviation morphology primitive type includes at least the first to sixth types, each corresponding to a different deviation evolution morphology.

[0020] Type I satisfies and and This characterizes a well-developed steady-state pattern characterized by no obvious trend in deviation, stable fluctuations, and no significant correlation between the mean and fluctuations. The second type satisfies and and This characterizes an early mismatch pattern where the mean deviation shows a continuous drift but the fluctuation has not yet intensified. The third type satisfies and and This characterizes a pattern where the mean deviation is stable but the intensity of fluctuations continues to increase; The fourth type satisfies and and The characteristic deviation means drift and increased volatility occur simultaneously and exhibit a positively correlated coupling mismatch. Type 5 satisfies and and This characterizes an unstable state characterized by increased volatility and the presence of periodic oscillations. The sixth type satisfies and This characterizes a saturated nonlinear morphology where the deviation increases but the fluctuation is suppressed.

[0021] The morphology recognition module calculates the weighted Euclidean distance between the fragment morphology encoding vector and the center vectors of each type. , Calculate using the following formula: ; in Let be the center vector of a primitive type. to Preset weight coefficients for each dimension. Measure the similarity between the current segment and each standard form using weighted Euclidean distance. The type corresponding to the minimum value is used as the main type of this segment. The type corresponding to the second smallest value is used as the subtype to complete the morphological classification of the deviation segment.

[0022] The transition analysis module is maintained by the most recent An observation sequence window, composed of primitive types corresponding to each segment, is used for path matching within a pre-constructed probability-weighted transition graph. The matching degree is then calculated. With the rate of change index Obtain the urgency of correction ,when Greater than the preset threshold The system outputs trigger signals and matching path information to the parameter update module in a timely manner, thereby perceiving the development trend of model mismatch and generating correction decisions when the model mismatch is still in an early stage.

[0023] Furthermore, the probability-weighted transition graph is constructed through an offline process, which includes: Injecting slow time-varying parameter drift sequences into the process mechanism model, wherein the drift type includes model gain With rate Changes, model time constant With rate Changes, model time delays With rate The changes were made to simulate the slow time-varying mismatch types commonly seen in actual industrial processes. Collect deviation signals throughout the process and generate primitive type sequences according to the aforementioned morphological recognition method; statistically analyze arbitrary primitive types. Transition to primitive type probability Average length of stay The type of mismatch to which the change belongs is marked, thereby establishing the correspondence between morphological changes and physical mismatches; Using primitive types as nodes, For the edge The weights are used to correct the associated model parameters in the vector. The attributes of the edges constitute the probability-weighted transition graph. This graph, in the form of a directed graph, depicts the typical paths and statistical characteristics of the morphological evolution of deviations under different mismatch types, providing a knowledge basis for path matching and correction decisions in the online stage.

[0024] Furthermore, the transition analysis module calculates the matching degree in the following manner. With the rate of change index : ; in The length of the observation sequence window. For nodes in the probability-weighted transition graph To node The transition probabilities are represented by the product, which represents the joint probability of each step of the observation sequence. The jump penalty factor is applied when there are adjacent nodes in the probability-weighted transition graph that are not directly connected by edges in the observation sequence. ,otherwise By penalizing skip nodes, the matching degree of atypical transition paths is reduced.

[0025] ; in The actual dwell time of the transition between two adjacent nodes within the observation sequence window. This represents the average dwell time of the transition in the probability-weighted transition graph. A higher value indicates that the actual rate of change is faster than the historical average, and the mismatch is more urgent. Match Degree The rate of change index reflects the degree of agreement between the current observed sequence and known mismatch paths. Both reflect the relative speed of mismatch development and together provide a basis for calculating the urgency of correction.

[0026] Furthermore, the urgency of the correction Calculate using the following formula: ; in This is a preset speed amplification factor, used to adjust the contribution of the transition speed to the urgency level; Weights for the process operation stages, during the middle stage. In the early stages of the process Late stage of the process ,and This reflects the differences in control quality requirements at different operational stages. (Calibration urgency) Taking into account three factors—path matching degree, the speed of mismatch development, and the importance of process stages—when Exceeding the preset threshold Parameter updates are triggered in a timely manner to ensure that correction decisions are only implemented when the mismatch is confirmed with sufficient confidence and has progressed to the point where intervention is required.

[0027] The parameter update module responds to the trigger signal and updates the parameters according to the matching path and the path progress coefficient. Determine the basic correction vector and correction step size After momentum smoothing calculation, the updated process model parameters are output to the model predictive controller, realizing the gradual adaptive adjustment of model parameters as the process dynamically changes.

[0028] Furthermore, the parameter update module extracts the parameter correction vector corresponding to the last change within the observation sequence window from the probability-weighted transition graph as the basic correction vector. This vector is determined in advance through parameter sensitivity analysis, indicating the direction and relative magnitude of parameter adjustments required to counteract the current mismatch trend.

[0029] Calculate the path progress factor ,in This represents the number of edges that have been traversed on the mismatched path in the current observation sequence. This represents the total number of edges contained in the mismatched path. This reflects the current stage of mismatch within a typical development path.

[0030] Correction step size in accordance with Determine using the following formula: ,when ; ,when ; ,when ; in The preset baseline step size is used. The above segmentation rule allows for the use of a smaller step size in the early stages of mismatch development to maintain stability, and a larger step size in the later stages of mismatch development to accelerate the response speed.

[0031] Furthermore, the momentum smoothing calculation is performed using the following formula: ; in This is the parameter vector for the current process model. For the preset momentum coefficient, This approach continues the historical update direction, helping to suppress potential oscillations introduced by a single update and smoothing out the parameter change trajectory. The parameter update module also calculates... First, calculate the current correction vector. With recent The average vector of the actual correction vectors performed. The angle between them, when the angle is greater than If the current correction direction is opposite to the recent correction trend, it is determined to be an abnormal fluctuation that may be caused by misidentification or noise. The parameter update is then stopped and the event is recorded to avoid the adverse effects of incorrect parameter adjustments on control stability.

[0032] The technical effects and advantages of this invention are as follows: (i) The deviation acquisition module collects the difference between the process setpoint and the actual measured value in real time as the deviation signal. The feature compression module statistically analyzes the deviation signal at fixed time periods, calculating the mean, standard deviation, and range of the deviation within each period as feature points. Simultaneously, when a deviation is detected crossing the boundary of the zone of interest, the trend of the deviation rate of change reverses, or a change signal is received during the process operation phase, the current period is immediately truncated and feature points are generated, thus forming a non-equally spaced deviation feature point sequence. This method, while preserving the statistical characteristics of the deviation, can capture key moments when the deviation morphology changes, providing a more accurate data foundation for subsequent time series analysis and helping to improve the detection sensitivity of early mismatch signals.

[0033] (II) The segmentation module divides the feature point sequence into three macroscopic stages based on the process operation phases: the initial stage, the middle stage, and the final stage. Within each stage, a dynamic programming algorithm is used to perform fine segmentation in a quality-sensitive weighted manner, dividing the feature point sequence into several segments with consistent internal morphology. The morphology recognition module extracts multi-scale morphological features from each segment to form a morphological encoding vector, and maps the segment to a preset deviation morphological primitive type by calculating a weighted Euclidean distance. The above processing transforms the continuous deviation signal into a sequence of morphological primitives with clear physical meaning, enabling the quantitative characterization of the deviation evolution law caused by model mismatch.

[0034] (III) The transition analysis module maintains an observation sequence window composed of the most recent fragment primitive types. It performs path matching on the observation sequences within an offline-constructed probability-weighted transition graph, calculates the matching degree and transition rate exponent, and generates a correction urgency. When the urgency exceeds a preset threshold, parameter updates are triggered. The probability-weighted transition graph is constructed by injecting slow time-varying parameter drifts of different rates and types into the mechanistic model and statistically analyzing the transition probabilities and average dwell times between primitive types. By analyzing the transition path of the deviation pattern rather than isolated amplitude characteristics, the system can perceive the development trend of model mismatch when it is still in the sub-threshold stage, helping to shorten the time interval from mismatch occurrence to correction initiation.

[0035] (iv) The parameter update module determines the basic correction vector based on the matched mismatched path and the current path progress. After progress adjustment, confidence weighting, and momentum smoothing, the parameter update amount is obtained and output to the model predictive controller. The correction step size is adaptively adjusted according to the path progress. A smaller step size is used in the early stage of the path to maintain stability, and a larger step size is used in the later stage of the path to accelerate the response speed. Before the parameter update, the consistency of the correction direction is monitored. When the angle between the current correction direction and the average of the recent correction directions exceeds a preset angle, the update is stopped. This method helps to improve the smoothness and adaptability of the online update of model parameters, enabling the model predictive controller to maintain a match with the actual dynamic characteristics of the controlled process under slow time-varying conditions. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the system module composition of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1 As attached Figure 1 The illustrated industrial parameter adaptive control system based on artificial intelligence includes a deviation acquisition module, a feature compression module, a segmented processing module, a morphology recognition module, a change analysis module, and a parameter update module. Each module runs in the host computer of the industrial process control system and interacts with the real-time database of the distributed control system. The model predictive controller solves for the optimal sequence of manipulated variables based on process model parameters, controlled variable setpoints, and controlled variable measurements. The process model is in the form of a transfer function matrix, and the model parameter vector includes the gain and time constant of each input / output channel.

[0039] 1. Deviation Acquisition Module The deviation acquisition module establishes a connection with the OPC server of the distributed control system via an industrial Ethernet interface, and reads the real-time values ​​of the process setpoint tag and the actual measured value tag from the OPC server at a preset sampling period. The sampling period is determined based on the dominant time constant of the controlled process, and the sampling period value is one-tenth of the dominant time constant.

[0040] After each read operation is completed, the deviation acquisition module immediately calculates the difference between the set value and the measured value. The calculation formula is as follows: ,in Indicates the set value. Indicates the measured value. This indicates a deviation signal. If the number of consecutive failed reads exceeds a preset threshold, the deviation acquisition module outputs a preset default deviation value and issues an alarm signal.

[0041] When the controlled process is a multivariable system, the deviation acquisition module simultaneously reads the setpoints and measured values ​​of multiple control loops, calculates the deviation component for each loop, and outputs it as a deviation vector. Each element in the deviation vector corresponds to the deviation signal of a controlled variable. The deviation acquisition module transmits the calculated deviation signal along with the corresponding timestamp to the feature compression module.

[0042] 2. Feature Compression Module The feature compression module internally sets up a fixed time window timer, the duration of which is denoted as . If the deviation acquisition module outputs a multivariate deviation vector, the feature compression module performs the same adaptive segmented compression operation on each deviation component in the deviation vector, and each deviation component independently generates its own non-equal interval deviation feature point sequence.

[0043] In every complete Within a period, the feature compression module stores all received deviation signal sample values ​​in a buffer array. When At the end of the cycle, the feature compression module performs statistical calculations on the deviation sample values ​​in the buffer array: Calculate the arithmetic mean of all sampled values ​​as the mean deviation for that period, denoted as . ; Calculate the unbiased sample standard deviation of all sampled values ​​as the deviation standard deviation for that period, denoted as . ; The difference between the maximum and minimum values ​​among all sampled values ​​is calculated as the deviation range for that period, denoted as . .

[0044] Will , , Combined with the end timestamp of the current period, it forms a fixed period feature point, which is then output to the segmentation processing module.

[0045] exist During the period before the cycle ends, the feature compression module continuously monitors the triggering conditions of the following three types of events in real time.

[0046] The first trigger condition is that the deviation crosses the boundary of the zone of interest, and the area inside the zone of interest is... The focus is on the upper boundary, which is preset to be the boundary. and lower boundary ,in The value is taken as 1.5 times the standard deviation of the deviation when the system is in optimal control state. Values The feature compression module monitors the instantaneous value of the deviation signal. When the instantaneous value crosses from the inner region of the zone of interest to a value greater than [a certain value], [the module will detect the deviation]. The area or smaller When the region is specified, the condition is deemed to be met.

[0047] The second trigger condition is a reversal of the trend in the rate of change of deviation. The feature compression module records the average deviation of the previous complete period, denoted as... During the current cycle, the feature compression module calculates the arithmetic mean of the accumulated sampled values ​​in real time, denoted as... ,calculate Compared to Change Simultaneously, record the change from the previous period, denoted as . .when and When the product of and is less than zero, the condition is considered satisfied.

[0048] The third trigger condition is the receipt of a change signal during the process execution phase. The feature compression module listens for this signal, and when it receives the signal, it determines that the condition is met.

[0049] When any of the above conditions is satisfied, the feature compression module immediately uses the current time as the cutoff point to truncate and statistically analyze the accumulated deviation samples from the start of the cycle to the current time within this cycle, and calculates the mean deviation within the truncation interval. Standard deviation and range A morphological event trigger feature point is generated and output. At the same time, the feature compression module resets the time window timer and starts a new statistical cycle.

[0050] The feature compression module arranges fixed-period feature points and morphological event-triggered feature points in the order of their generated timestamps to form a non-equal-interval deviation feature point sequence, which is then used as the output of this module.

[0051] 3. Segmentation Processing Module The segmentation processing module receives the non-uniformly spaced deviation feature point sequence output by the feature compression module. When the input is a feature point sequence with multiple deviation components, the segmentation processing module performs the following temporal segmentation operation on the feature point sequence of each deviation component, and each component is independently divided into segments.

[0052] The segmented processing module first reads the current values ​​of key process indicators (TPIs) representing the operational stage of the industrial process from the real-time database of the distributed control system. These TPIs are pre-selected based on the specific industrial process type; for example, for polymerization processes, the TPI is monomer conversion rate; for distillation processes, the TPI is the purity of the top product. The numerical range of the TPIs is normalized to [value missing]. arrive between.

[0053] The segmented processing module compares the current values ​​of the key process indicators it reads with the preset first-stage threshold. Second stage threshold Comparison, among which , ,satisfy .

[0054] When the indicator value is less than When the indicator value is greater than or equal to the initial stage of the process, it is determined that the process is in its early stages; when the indicator value is greater than or equal to the initial stage of the process, it is determined that the process is in its early stages. and less than At that time, it was determined to be in the middle of the process; When the indicator value is greater than or equal to At this point, it is determined that the process is in its final stage. The above three macroscopic stages constitute the first-level segmentation boundary.

[0055] Within each macro-level stage, the segmentation processing module uses a dynamic programming algorithm to perform fine segmentation on the biased feature point sequence output by the feature compression module. Assume the feature point sequence contains... The feature points are numbered sequentially by time. to The goal of dynamic programming is to... A set of segmentation positions is determined among the feature points, such that the internal cost function of each segment is such that... The sum reaches its minimum.

[0056] The dynamic programming algorithm takes the segment cost between points in the feature point sequence as input and outputs the optimal set of segmentation point positions.

[0057] Internal cost function of fragment The calculation formula is: ; in The mean deviation of each feature point within the segment The variance of the constructed sequence, The standard deviation of the deviation of each feature point within the segment The variance of the constructed sequence, The trend consistency index is calculated by performing a least-squares linear fit on the mean sequence within the segment to obtain the slope of the fitted line, counting the number of continuous feature points whose slope remains unchanged within the segment, and dividing this number by the total number of feature points in the segment. The resulting ratio is the trend consistency index. Least squares linear fitting takes a two-dimensional data point set as input and outputs the slope and intercept of the fitted line.

[0058] Weighting coefficient , , Select from the preset weight table based on the current macroeconomic stage: In the early stages of the process, , , ; During the process , , ; Towards the end of the process , , .

[0059] in , , ; , , ; , , The above values ​​satisfy... , , .

[0060] The specific process of solving dynamic programming is as follows: Define a two-dimensional array, where element D[i][j] represents the element whose i is the first element. The first to the second Cost function for dividing a feature point into a segment The value; Define a one-dimensional array, where element F[h] represents the previous... The minimum total cost of dividing a feature point into several segments, where F[0] takes the value of .

[0061] The state transition equation is Through bottom-up iterative calculation, F[Q] is obtained, which is the minimum total cost. Simultaneously, the cost of each transition is recorded. The value is determined by the position of the backtracking split point.

[0062] After the dynamic programming solution is completed, the segmentation module obtains a set of segmentation point positions and divides the feature point sequence into several initial segments.

[0063] Subsequently, the segmentation module performs a minimum length check on each initial segment: for segments whose duration is less than the preset minimum duration... For a given segment, calculate the cosine similarity between the morphological encoding vector of that segment and the morphological encoding vectors of adjacent segments. Values Minutes. If the cosine similarity is greater than the preset similarity threshold. , Then, calculate the cosine similarity between the short segment and its immediate and adjacent segments, and merge the short segment with its adjacent segments that have a high similarity into one segment. After merging, recalculate the morphological encoding vector of the new segment and continue to compare its similarity with adjacent segments until the duration of all segments is greater than or equal to the specified length. Until further mergers are no longer possible.

[0064] The segmentation processing module ultimately outputs a sequence of segments after verification and merging. Each segment contains its starting feature point index, ending feature point index, process stage, all feature point data contained within the segment, and the sequence of original deviation signal sample values ​​within the time range covered by the segment.

[0065] 4. Morphology Recognition Module The morphology recognition module extracts multi-scale morphological features for each segment output by the segmentation processing module. When the input is a segment with multiple deviation components, the morphology recognition module performs the following feature extraction and primitive type mapping operations on each segment of the deviation components, and each component independently outputs a primary type and a secondary type. The feature extraction process is divided into two parts: coarse-grained feature extraction and fine-grained feature extraction.

[0066] For coarse-grained feature extraction, the mean deviation of each feature point within the segment is extracted. Arrange the values ​​in chronological order to form a mean sequence. Perform a least-squares linear fit on this mean sequence, and use the slope of the fitted line as a feature. Extract the standard deviation of the deviation of each feature point within the segment. Arrange the standard deviations in chronological order to form a series. Perform a least-squares linear fit on this series of standard deviations, and use the slope of the fitted line as a feature. .

[0067] Calculate the Pearson correlation coefficient between the mean sequence and the standard deviation sequence within this segment. The Pearson correlation coefficient calculation takes two equal-length numerical sequences as input and outputs a value between [values ​​missing]. and The correlation coefficient between them is calculated and used as a feature. Based on the original deviation signal sample value sequence within the time range covered by this segment, the total number of times the instantaneous value of the deviation signal crosses zero is counted. This number is then divided by the total number of sampling points within the segment, and the resulting ratio is used as a feature. .

[0068] For fine-grained feature extraction, a preset first and last reference duration is used. , The value is one-third of the total duration of the segment. The last value of the segment is taken. Calculate the mean deviation of all feature points within the time period. The average value is denoted as ; Take the original fragment Calculate the mean deviation of all feature points within the time period. The average value is denoted as ,feature The calculation formula is Take the last part of this segment. Calculate the standard deviation of the deviation for all feature points within the time period. The average value is denoted as ; Take the original fragment Calculate the standard deviation of the deviation for all feature points within the time period. The average value is denoted as ,feature The calculation formula is .

[0069] The mean deviation of each feature point within the segment Arranged chronologically to form a mean sequence, the difference between adjacent first-order differences is calculated point-by-point in this mean sequence as the second-order difference. The time interval between feature points is not considered in the calculation. The absolute value of each second-order difference is taken, and the arithmetic mean of all absolute values ​​of second-order differences is calculated as the feature value. .

[0070] The morphological recognition module arranges the above seven features in sequence to form a morphological encoding vector, denoted as . The morphology recognition module has pre-stored center vectors for six types of deviation morphological primitives. Each center vector is obtained through K-means clustering, which takes a large number of morphological encoding vectors extracted from historical or simulation data as input and outputs six cluster center vectors.

[0071] The six types of center vectors are denoted as follows: First type center vector , Second type of center vector , Third type of center vector , Fourth type center vector , Fifth type center vector , Sixth type center vector .

[0072] The six primitive types correspond to the following judgment conditions, among which , , , , : Type I satisfies and and ; The second type satisfies and and ; The third type satisfies and and ; The fourth type satisfies and and ; Type 5 satisfies and and ; The sixth type satisfies and .

[0073] For the morphological encoding vector of the current segment, the morphological recognition module calculates its weighted Euclidean distance to each of the six center vectors mentioned above. The calculation formula is: ; in , , , , , , These represent the values ​​of each dimension of the center vector of the primitive type whose distance is being calculated. When calculating the distance to the first type, the value is taken as... When calculating the distance to the second type, the value is taken as... And so on, with the same logic applied to the other dimensions.

[0074] to The preset weighting coefficients for each dimension, , , , , , , ,satisfy .

[0075] After calculation, the morphology recognition module compares the six distance values, selects the smallest value, and determines the primitive type corresponding to the smallest value as the primary type of the segment; it then selects the second smallest value and determines the primitive type corresponding to the second smallest value as the secondary type of the segment. The morphology recognition module outputs the primary and secondary types of each segment, along with the segment identification information, to the transition analysis module. The secondary type is used as an alternative matching path when the system runs for the first time or when the transition probability of the corresponding primary type in the probability-weighted transition graph is lower than a preset backup threshold, which is set to 0.1.

[0076] The offline construction of the probability-weighted transition graph is as follows: The probability-weighted transition graph is pre-built offline and stored in the system's configuration database for use by the transition analysis module during online operation. The transition analysis module reads the adjacency list data of the probability-weighted transition graph from the configuration database when the system starts and loads it into memory. The specific construction process of the probability-weighted transition graph is as follows.

[0077] First, a mechanistic simulation model of the controlled industrial process is established. This model is based on first-principles calculations and includes externally adjustable process parameters, including at least the process gain. Process time constant and process delay .

[0078] Secondly, slow time-varying parameter drift sequences are injected into the mechanistic model to target the process gain. Applying a rate of change of The monotonous drift, Values nominal value Every 24 hours; Regarding the time constant Applying a rate of change of The monotonous drift, Values nominal value Every 24 hours; For time delay Applying a rate of change of The monotonous drift, Values nominal value Every 24 hours.

[0079] In each simulation run, you can choose to drift a single parameter or a combination of drifting multiple parameters simultaneously.

[0080] Then, the simulation is run and the deviation signal is collected. Under each parameter drift configuration, the deviation signal is recorded at the same sampling period as in the online operation phase throughout the entire time interval from the start of the simulation to the moment when the process performance deteriorates significantly. The collected deviation signal data is processed using the same feature compression, segmentation and morphology recognition process as in the online operation phase to generate a complete primitive type sequence.

[0081] Next, the transition probabilities and average dwell times were statistically analyzed for all primitive type sequences obtained from all simulation experiments. For any two primitive types... and Statistics on all sequences from type Direct transition to type The total number of occurrences is recorded as Statistics on the types of all sequences The total number of occurrences is recorded as transition probability The calculation formula is For each time from type To type The transition instance records the system's state at which it remained during that transition. The duration of the state is calculated by summing the dwell times of all such instances and then taking their arithmetic mean as the average dwell time. .

[0082] Next, the mismatch type is labeled and the correction vector is determined. For each type... To type For each transition edge, the mismatch type is labeled according to the parameter drift type injected when generating the transition sequence. Simultaneously, for each transition edge, the corresponding basic correction vector of the model parameters is determined using parameter sensitivity analysis. .

[0083] The specific process of parameter sensitivity analysis is as follows: under the parameter drift scenario corresponding to this transition, along the gain... Time constant Time delay Small positive and negative perturbations are applied across three dimensions. The rate of change of the deviation response is calculated, and the direction that results in the fastest reduction of the deviation is taken as the correction direction. The unit vector in that direction is multiplied by a preset base amplitude as the correction value. .

[0084] Finally, a directed graph data structure is constructed, using six primitive types as the six nodes of the directed graph. For each pair satisfying... Greater than zero and Create a branch node in the graph. Pointing to node A directed edge, the weight of which is assigned a value. The set of attributes of an edge includes the average dwell time. Mismatch type identifier and basic correction vector The resulting directed graph is a probability-weighted transition graph, which is stored in adjacency list format.

[0085] 5. Change Analysis Module The change analysis module internally maintains a first-in-first-out fixed-length queue as the observation sequence window, with a fixed queue length. , When the morphology recognition module outputs principal type sequences corresponding to multiple deviation components, the transition analysis module maintains an independent observation sequence window for each principal type sequence and performs the following path matching and urgency calculation respectively.

[0086] Whenever the morphology recognition module outputs the main type of a new segment, the transition analysis module adds the main type of the new segment as a new element to the tail of the corresponding component's queue. If the number of elements in the queue exceeds [a certain threshold], [the transition analysis module will then add the new segment's main type as a new element to the tail of the queue.] Then, the earliest added element is removed from the head of the queue, ensuring that the queue always contains the most recently added element. The main type sequence of fragments.

[0087] The transition analysis module calculates the matching degree in the following manner. : Let the observation sequence window currently contain The main types are listed in chronological order as follows: Query sequentially from the probability-weighted transition graph common Each transition probability value, the above The product is obtained by multiplying the transition probability values ​​together.

[0088] Then determine the jump penalty factor. The values ​​are checked one by one. For adjacent types, if every pair of adjacent types has a corresponding direct directed edge in the probability-weighted transition graph, then Values If there exists at least one pair of adjacent types that have no direct directed edge in the probability-weighted transition graph, then The value is 0.5.

[0089] Match The calculation formula is: ; For each known mismatch path marked on the probability-weighted transition graph, the transition analysis module calculates the matching degree between the observation sequence window and the path in the manner described above. Take one of them Maximum value and The path is used as the matching path. If multiple paths exist... If the values ​​are the same and all are the maximum, then the transition rate index is selected. The larger path is used as the matching path.

[0090] The change analysis module calculates the change rate index in the following manner. Take the latest change that occurred in the observation sequence window, i.e., from Change to This change has led to the acquisition system... The actual duration of the state is denoted as . ;Query from the probability-weighted transition graph to Average residence time of change , recorded as The rate of change index The calculation formula is: ; in This is a function that takes two values ​​as input and outputs the larger of the two.

[0091] The change analysis module calculates the correction urgency in the following manner. Based on the current macroscopic stage of the process, select a stage weight factor from the preset weight values. When in the middle of the process When in the early stages of the process When it is in the late stage of the process ,in , , ,satisfy ; Obtain the preset speed amplification factor , .

[0092] Urgency of correction The calculation formula is: ; The transition analysis module calculates the urgency of correction for each deviation component. Take all deviation components The value with the largest value is taken as the overall correction urgency at the current moment. This overall correction urgency is then compared with a preset urgency threshold. Comparison, .

[0093] When the urgency of comprehensive correction is greater than At that time, the transition analysis module generates a correction trigger signal and obtains the maximum value. The deviation component of the value corresponds to the matching mismatch path identifier and the current matching degree. The directed edge attribute information corresponding to the most recent change is packaged into a trigger message and sent to the parameter update module; when the urgency of comprehensive correction is less than or equal to At this time, no trigger signal is output, and the transition analysis module continues to wait for the recognition result of the next segment.

[0094] 6. Parameter Update Module After receiving the trigger message from the transition analysis module, the parameter update module first parses the trigger message and extracts the basic correction vector contained in the directed edge attributes corresponding to the most recent transition, denoted as . , It is a vector whose dimensions are the same as those of the process model parameters.

[0095] Secondly, calculate the path progress factor. Based on the mismatch path identifier carried in the trigger message, the total number of transition edges contained in the mismatch path is obtained from the configuration data of the probability-weighted transition graph and denoted as . The number of transition edges that the type sequence within the current observation sequence window has already experienced on this mismatch path is denoted as . Path progress coefficient The calculation formula is .

[0096] Then, determine the correction step size. Obtain the preset reference step size , The values ​​are the nominal values ​​of each model parameter. According to the path progress coefficient The numerical range is determined according to the following segmentation rules. : when hour, ; when hour, ; when hour, .

[0097] Next, the parameter update module performs a consistency check of the correction direction: calculating the correction vector to be executed this time. The calculation formula is: Retrieve the most recent data from the system's stored history. The actual correction vector executed this time is denoted as , , Calculate this The arithmetic mean of the correction vectors is denoted as... The calculation formula is: ; Calculate vectors with vector The angle between them.

[0098] When the included angle is greater than If the current correction direction is determined to significantly conflict with the recent correction trend, the parameter update module will abort the current parameter update process and record the time of the abort event, the current observation sequence window content, and the calculated parameters. and Write to the system log file; when the included angle is less than or equal to At that time, the parameter update module continues to execute subsequent steps.

[0099] Next, the parameter update module performs momentum smoothing parameter updates: it reads the current parameter vector from the process model currently used by the model predictive controller, denoted as... Retrieve the process model parameter vector recorded since the last parameter update from the system's stored history, denoted as... ; Obtain the preset momentum coefficient , .

[0100] Updated parameter vector The calculation formula is: ; Finally, the parameter update module will calculate the... Write the process to the model parameter storage area, overwriting the original parameter values. This storage area contains specific parameter tags in the distributed control system's real-time database. The parameter update module updates the tag values ​​through the data write interface.

[0101] At the beginning of each control cycle, the model predictive controller reads the latest model parameters from this storage area for solving the control variables. The model predictive controller takes the process setpoint, process measurements, and process model parameters as inputs and outputs the manipulated variable values. The parameter update module also updates the actual correction vector executed in this cycle. Stored in the historical record queue for subsequent monitoring of consistency in the correction direction.

[0102] The specific process of system collaboration is as follows: (1) The deviation acquisition module collects the deviation signal in real time and transmits it to the feature compression module; (2) The feature compression module adaptively compresses the deviation signal to generate a non-equal interval deviation feature point sequence and outputs it to the segmentation processing module; (3) The segmentation module performs quality-sensitive weighted temporal segmentation on the feature point sequence according to the process operation stage, and outputs the segmented sequence to the morphology recognition module. (4) The morphology recognition module extracts multi-scale morphological features for each segment, maps the segment to a deviation morphological primitive type, and outputs the type sequence to the transition analysis module; (5) The change analysis module maintains the observation sequence window, performs path matching in the probability-weighted change graph, and sends a trigger signal to the parameter update module when the correction urgency exceeds the threshold. (6) The parameter update module responds to the trigger signal, determines the basic correction vector and correction step size, and updates the process model parameters after direction consistency monitoring and momentum smoothing, so that the model predictive controller can use them in subsequent control cycles.

[0103] Through the closed-loop process described above, the system continuously senses early trends of model mismatch and gradually adjusts the model parameters during the continuous operation of the industrial process, so that the model predictive controller always matches the actual dynamic characteristics of the controlled process.

[0104] Alternative Implementation Method 1: Sliding Window Segmentation As one implementation method, the segmentation module can use a sliding window local feature mutation detection method for temporal segmentation.

[0105] The segmentation processing module internally uses a fixed-length analysis window, whose length is set to include a fixed number of feature points, for example, 10 feature points. The module aligns the starting position of this analysis window with the first feature point of the deviation feature point sequence and calculates the internal cost function of the segment formed by all feature points within the window. The calculation method is the same as in the above implementation method. The calculation formula is the same.

[0106] Then the segmented processing module slides the analysis window backward along the time axis by a step size of one feature point, and recalculates the value within the new window. Values, and calculate the values ​​corresponding to two consecutive slides. The absolute value of the difference between the values ​​is denoted as . .when When the value exceeds the preset mutation detection threshold, the mutation detection threshold is set to the historical average. 0.5 times the value, historical average Value taken from the initial operation of the system All sliding windows within 24 hours The arithmetic mean of the values ​​is used to mark a dividing point at the boundary between the current window and the previous window in the segmentation module.

[0107] The segmented processing module repeatedly executes the above sliding window and calculation. Value, comparison The process of thresholding and marking segmentation points continues until the end of the analysis window reaches the last feature point in the deviation feature point sequence. All marked segmentation point positions divide the feature point sequence into several segments. After segmentation, the segmentation processing module performs minimum length verification and merging operations on the resulting segments, following the same rules as described above.

[0108] Alternative Implementation Method 2: Convolutional Neural Network Feature Extraction As another implementation method, the morphology recognition module can use a one-dimensional convolutional neural network as a feature extractor.

[0109] The one-dimensional convolutional neural network was trained offline before system deployment. The training data came from a large number of deviation fragment samples generated by the mechanism simulation model of the controlled industrial process under various slow time-varying parameter drift conditions. Each sample was labeled with its corresponding deviation morphology primitive type. The training employed a cross-entropy loss function and the Adam optimizer.

[0110] The network structure consists of three one-dimensional convolutional layers, two pooling layers, and two fully connected layers. The kernel size of the first convolutional layer is... Step size is The number of filters is 32; the kernel size of the second convolutional layer is... Step size is The number of filters is 64; the kernel size of the third convolutional layer is... Step size is The number of filters is 128. Each convolutional layer is followed by a pooling layer, with a pooling window size of [size missing]. Step size is The last fully connected layer has a seven-dimensional output, and its seven output nodes correspond sequentially to the features. , , , , , , .

[0111] During the online operation phase, the morphology recognition module takes the mean deviation sequence of each segment output by the segmentation processing module as input. Before forward propagation, the mean deviation sequence of the input network is first normalized to zero mean, i.e., the sequence mean is subtracted and then divided by the sequence standard deviation. Then, through one forward propagation calculation, the seven-dimensional morphological encoding vector is directly obtained from the last fully connected layer. The subsequent primitive type mapping process is exactly the same as the implementation described above.

[0112] Alternative Implementation Method 3: Comprehensive Change Rate Index As another implementation method, the transition speed index in the transition analysis module The calculation can comprehensively consider all windows. The rate of change is abnormal.

[0113] When the number of segments in the observation sequence window is less than hour, Values When the number of segments in the observation sequence window is greater than or equal to At that time, for each transition within the observation sequence window, i.e. from to ,from to ,......,from to If there exists a transition from the probability-weighted transition graph... to If there is a direct directed edge, then the system remains at a certain state during that transition. The duration of the state is taken as the actual dwell time. Obtain the corresponding average dwell time from the probability-weighted transition diagram. Calculate the abnormal components of the speed If the directly directed edge does not exist in the graph, then let The subscript Values to .

[0114] Then all indivual Calculate the arithmetic mean as the overall rate of change index. ,Right now urgency of correction The calculation formula used This is the comprehensive change rate index.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An industrial parameter adaptive control system based on artificial intelligence, characterized in that, It includes a deviation acquisition module, a feature compression module, a segmentation processing module, a morphology recognition module, a transition analysis module, and a parameter update module; The deviation acquisition module collects the difference between the set value and the actual measured value in real time as a deviation signal. The feature compression module performs adaptive segmented compression on the deviation signal and outputs a sequence of non-equally spaced deviation feature points composed of a mixture of fixed-period statistical points and morphological event trigger points. The segmentation module performs quality-sensitive weighted temporal segmentation on the deviation feature point sequence according to the process operation stage, dividing the deviation feature point sequence into several segments with consistent internal morphology. The morphology recognition module extracts multi-scale morphological features for each segment to form a morphological encoding vector, and maps the morphological encoding vector to a deviation morphological primitive type. The transition analysis module is maintained by the most recent An observation sequence window, composed of primitive types corresponding to each segment, is used for path matching within a pre-constructed probability-weighted transition graph. When the matching degree... With the rate of change index Calculated correction urgency Greater than the preset threshold At that time, a trigger signal and matching path information are output to the parameter update module; The parameter update module responds to the trigger signal and updates the parameters according to the matching path and the path progress coefficient. Determine the basic correction vector and correction step size After momentum smoothing calculation, the updated process model parameters are output to the model predictive controller.

2. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The feature compression module operates at a fixed time period. Statistical analysis is performed on the deviation signal, and the mean, standard deviation, and range of the deviation within each period are calculated as a feature point. Simultaneously, the current period is immediately truncated and a feature point is generated when any of the following conditions are met: the deviation value crosses the upper limit of the preset attention band. or lower limit The signs of the rate of change of the average deviation between two adjacent cycles are opposite, and a change signal is received during the process operation phase.

3. The industrial parameter adaptive control system based on artificial intelligence according to claim 1, characterized in that, The segmented processing module acquires key process indicators, and when the value of the key process indicator is less than... The time division process is divided into initial stage and later stage. to The interval is divided into the middle period, in to The interval is divided into final stages; within each stage, dynamic programming is used to find the set of segmentation points that minimize the cost function value within the segment. Calculate using the following formula: ； in The mean of the segment is the variance of the sequence. The standard deviation is the series variance. As a trend consistency indicator, The value is the proportion of the total duration of the segment in which the slope of the linear fit of the mean sequence within the segment remains consistent with the sign of the segment's slope. , , The weighting coefficient is set to a value in the initial stage. , , The value in the intermediate period is , , The value at the end of the period is , , ,and , , ; For segments with a duration less than Furthermore, the cosine similarity with the morphological encoding vector of adjacent segments is greater than [value missing]. The fragments are merged.

4. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The coarse-grained feature extracted by the morphology recognition module is the slope of the linear fit of the mean sequence of deviations within the segment. Slope of linear fit of standard deviation sequence Correlation coefficient between mean and standard deviation The ratio of the number of zero-crossing deviations to the segment length ; Fine-grained features are the last fragment The mean over the period of time and the initial Difference of mean over time ,at last Standard deviation over time and initial Ratio of standard deviations over time and the mean of the absolute values ​​of the second differences of the mean sequence within the segment. The morphological encoding vector is .

5. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The deviation morphological primitive type includes at least the first type to the sixth type; Type I satisfies and and ; The second type satisfies and and ; The third type satisfies and and ; The fourth type satisfies and and ; Type 5 satisfies and and ; The sixth type satisfies and ; The morphology recognition module calculates the weighted Euclidean distance between the fragment morphology encoding vector and the center vectors of each type. , Calculate using the following formula: ； in It is the center vector of a primitive type. to Preset weight coefficients for each dimension; Will The type corresponding to the minimum value is used as the main type of this segment. The type corresponding to the second smallest value is taken as the secondary type.

6. The industrial parameter adaptive control system based on artificial intelligence according to claim 1, characterized in that, The process of constructing the probability-weighted transition graph includes: Injecting slow time-varying parameter drift sequences into the process mechanism model, wherein the drift type includes model gain With rate Changes, model time constant With rate Changes, model time delays With rate Change; Collect the deviation signal throughout the process and generate a primitive type sequence according to the method described in claim 4 or claim 5; statistically analyze any primitive type. Transition to primitive type probability Average length of stay And indicate the type of mismatch to which the transition belongs; Using primitive types as nodes, For the edge The weights are used to correct the associated model parameter vector. The attributes of the edges constitute the probability-weighted transition graph.

7. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The transition analysis module calculates the matching degree in the following manner. With the rate of change index : ； in The length of the observation sequence window. For nodes in the probability-weighted transition graph To node The probability of transition, The jump penalty factor is applied when there are adjacent nodes in the probability-weighted transition graph that are not directly connected by edges in the observation sequence. ,otherwise ; ； in The actual dwell time of the transition between two adjacent nodes within the observation sequence window. This represents the average dwell time of the transition in the probability-weighted transition graph.

8. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The urgency of the correction Calculate using the following formula: ； in This is the preset speed amplification factor; Weights for the process operation stages, during the middle stage. In the early stages of the process Late stage of the process ,and .

9. The artificial intelligence-based industrial parameter adaptive control system according to claim 1, characterized in that, The parameter update module determines the correction step size in the following manner. With respect to the basic correction vector: The parameter correction vector corresponding to the last transition within the observation sequence window is extracted from the probability-weighted transition map and used as the basic correction vector. ; Calculate the path progress factor ,in This represents the number of edges that have been traversed on the mismatched path in the current observation sequence. This represents the total number of edges contained in the mismatched path; Correction step size in accordance with Determine by the following formula: ,when ; ,when ; ,when ; in This is the preset baseline step size.

10. The industrial parameter adaptive control system based on artificial intelligence according to claim 1, characterized in that, The momentum smoothing calculation is performed using the following formula: ； in This is the parameter vector for the current process model. The preset momentum coefficient is used; the parameter update module is also calculating... First, calculate the current correction vector. The average vector of the three most recent actual correction vectors The angle between them, when the angle is greater than If the parameter update fails, the current update will be stopped and the event will be recorded.

Images