Casting production line energy consumption optimization control method

Abstract

The invention discloses a casting production line energy consumption optimization control method, which comprises the following steps of: deploying intelligent sensing nodes in each process unit, acquiring and marking various process data in real time, and establishing a multi-target feature modeling and optimization framework based on layered multi-agent by adopting data preprocessing means such as time sequence synchronization, exception elimination and normalization; balanced optimization of key indexes such as energy consumption, productivity, quality and safety is achieved, the system has the capabilities of parameter self-adaptive switching, real-time response, rolling optimization and continuous self-lifting, and the operation efficiency of a production line, energy consumption control and robustness under complex working conditions are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation and intelligent manufacturing, and particularly relates to a casting production line energy consumption optimization control method. BACKGROUND

[0002] With the continuous improvement of industrial automation and intelligent manufacturing level, the energy consumption optimization and multi-objective collaborative control of the casting production line have become an important means for energy saving and efficiency improvement. At present, in the field of automatic control of casting process, the mainstream technology mainly relies on centralized data acquisition and optimization model, and adopts single-objective optimization algorithm or segmented energy consumption scheduling method based on process experience. These systems usually collect energy consumption, production status, product quality and other parameters of each process, combine with experience rules or static optimization model in the process field, and realize the control of energy consumption and the satisfaction of part of production constraints. In recent years, with the popularity of industrial Internet of Things and Multi-Agent System (MAS) technology, some intelligent control systems introduce multi-objective optimization, distributed modeling and local / global scheduling collaboration methods, in order to realize the comprehensive optimization of energy consumption and production capacity under the premise of ensuring product quality and safety.

[0003] In the related technical solutions disclosed, multi-objective energy consumption optimization often adopts an integrated model, which cooperatively considers multiple objectives such as energy consumption, production capacity, quality and equipment safety through large-scale mixed integer programming, genetic algorithm and other global optimization tools, and outputs the optimal parameter setting. However, this kind of method is limited by single model level, large parameter space, complex process coupling and real-time data delay, etc., and it is difficult to balance the global convergence of optimization solution and the dynamic response speed of the field. Some systems try to use rule-based segmented optimization or real-time rolling optimization window to dynamically adjust the parameters, but when facing sudden working conditions and multiple conflicting target trade-offs, the response ability and adjustment timeliness are weak. In addition, the existing distributed optimization algorithm lacks adaptive adjustment mechanism of hierarchical weight when facing multi-objective sensitive balance, process segment and equipment level collaborative decision, which makes it difficult to dynamically balance between global target and local constraint, affecting the simultaneous optimization of energy efficiency improvement and comprehensive production capacity.

[0004] Typical technical application scenarios mainly include: in a large-scale casting production line, the central control system or distributed workshop host collects data of each production unit, analyzes energy consumption and key indicators of production process, and sets parameters for optimization. In this process, the traditional scheme mainly adopts fixed-period segmented optimization, static energy consumption index upper and lower limit control or manual experience tuning, and has limited ability to deal with coupling relationship between multiple objectives, production variation and abnormal state.

[0005] The prior art mainly has the following prominent problems: (1) it is difficult to achieve global optimal real-time response of production line level parameters under multi-objective trade-off, that is, the convergence speed and response time of the optimization algorithm cannot meet the field application requirements when dealing with field real-time data mutations and target conflicts, resulting in limited energy consumption and production capacity improvement; (2) the hierarchical multi-agent system lacks a sound target weight self-adaptive adjustment and process section coordination mechanism, and it is difficult to achieve flexible index switching and self-balancing between different process units and production sections; (3) the current optimization control mode has weak adaptability to low latency and strong dynamic changes, and the parameter setting is rigid and prone to lag, which cannot continuously evolve to adapt to the changing market demand and equipment state; (4) the optimization index interaction relationship modeling under multi-objective scenarios such as energy consumption, production capacity, quality and safety is not comprehensive, resulting in low matching degree of the parameter solving model and the actual working condition, and the model does not have perfect self-adaptive fine-tuning ability. SUMMARY

[0006] The present application provides a casting production line energy consumption optimization control method to solve the above technical problems.

[0007] The technical scheme of the present application is as follows: a casting production line energy consumption optimization control method, comprising:

[0008] S1, real-time collection of original collection data of multiple process units in the casting production line, including energy consumption data, production state data, quality index data and equipment safety parameters, and identification of process categories and production section labels for each sampling point;

[0009] S2, preprocessing of the collected multi-process unit original collection data to obtain a preprocessed multi-process unit data set;

[0010] S3, based on different production section labels of each process unit, corresponding multi-objective energy consumption-quality-production capacity-safety feature modeling is established, parameter distribution modeling is performed using multi-objective optimization theory, and a feature weight distribution matrix of each unit is obtained;

[0011] S4, relying on a hierarchical multi-agent architecture, inputting the multi-objective feature weight distribution matrix into the corresponding process unit agent, independently solving the energy consumption optimal parameters of each unit, and generating a candidate optimization parameter vector;

[0012] S5, synchronously submitting the candidate optimization parameter vector to a global scheduling agent, dynamically adjusting the decision weight coefficients of each unit through an adaptive reward mechanism combined with a reinforcement learning method;

[0013] S6, using a short window rolling optimization algorithm, based on the weight coefficients and the candidate optimization parameter vector output by the global scheduling agent, real-time updating the energy consumption optimal parameter setting of each process unit;

[0014] S7, judge whether the energy consumption optimal parameter setting of the process unit in the current time window meets the real-time response threshold condition, if convergence delay is caused by target conflict, switch to the approximate parameter fast solving mode;

[0015] S8, based on the energy consumption optimal parameter setting result in implementation, continuously collect the production line actual energy consumption, capacity and quality feedback data input into the adaptive retraining mechanism, dynamically fine-tune the multi-agent parameters and model.

[0016] Advantages

[0017] Firstly, the present application is aimed at the problem of real-time deficiency of optimization parameters under multi-objective trade-off, and innovatively introduces a hierarchical multi-agent system architecture, which subdivides the casting production line into process unit layer, production section layer and whole line layer, and each unit is independently responsible for energy consumption optimization and real-time collection of key data of the unit by local intelligent agent (Agent). Through the combination of local autonomous optimization and global collaborative mechanism, the technical bottleneck of slow response and difficulty in real-time matching of production dynamics of the existing single centralized optimization is broken through, the synchronization of convergence speed, algorithm response and actual production change of each level optimization algorithm is realized, and the real-time performance of parameter solving and application is effectively improved;

[0018] Secondly, the data collection, preprocessing and multi-objective weight distribution modeling link proposed by the present application effectively improves the consistency and standardization of input data. Through multi-source data real-time collection, abnormal processing, time synchronization and normalization technology, the influence of data noise and abnormal fluctuation is greatly reduced, and the reliability and precision of subsequent optimization modeling are improved. The distribution modeling of multi-objective feature weight (such as PCA, mutual information analysis, AHP hierarchical analysis, etc.) makes the relationship between energy consumption, capacity, quality and equipment safety and other indicators be scientifically quantified, which greatly improves the precision and adaptability of parameter tuning compared with traditional experience weight method or single objective optimization;

[0019] Thirdly, aiming at the convergence delay problem caused by dynamic changes and target conflict of the production line, the present application innovatively adopts a "dual-mode switching" optimization mechanism: under normal working conditions, the high-precision multi-objective optimization algorithm is preferentially executed to ensure the global optimization of energy consumption and multi-objective parameters; when encountering sudden abnormality or dynamic sharp change, the system can automatically switch to the fast approximate solving mode to ensure that the optimization decision can be completed within a strict response time window. This design significantly shortens the parameter adjustment delay, improves the sudden adaptability of the system, and solves the problem of "response-precision" imbalance of traditional multi-objective optimization;

[0020] The present invention's short-window rolling optimization and adaptive reward mechanism are key to achieving global dynamic weight adjustment for hierarchical agents. Through intelligent algorithms such as reinforcement learning, the system automatically and dynamically adjusts the weights of various objectives based on real-time production line feedback, ensuring that global performance maintains a high level of balance between energy consumption, production capacity, and quality over the long term. This improves overall system optimization efficiency by over 20% compared to static weighting or manual adjustment methods, significantly enhancing the reliability of energy consumption adjustment and line capacity utilization.

[0021] In summary, the present invention realizes efficient, stable, real-time and intelligent decision-making and deployment of multi-objective energy consumption optimal parameters of the casting production line through innovative technologies such as hierarchical multi-agent collaborative optimization, data standardization modeling, dual-mode real-time parameter solution and adaptive evolution mechanism, solves the difficult problem that the existing technology cannot take into account both global dynamic balance and real-time rapid response, provides strong technical support for energy consumption control and production efficiency improvement in the field of intelligent manufacturing, and has outstanding creativity and wide industrial applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Attachment Figure 1 It is a main flow chart of the energy consumption optimization control method of a casting production line;

[0023] Attachment Figure 2 It is a sub-flowchart of the energy consumption optimization control method of the casting production line;

[0024] Attachment Figure 3 It is a sub-flowchart of a method for optimizing energy consumption control of a casting production line. DETAILED DESCRIPTION

[0025] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0026] The preferred implementation methods of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0027] As used herein, the singular forms "a," "an," and "the" may also include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "include," "comprising," "having," and the like specify the presence of stated features, integers, steps, operations, components, parts, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or combinations thereof. Furthermore, the term "and / or" as used in this specification includes any and all combinations of the relevant listed items.

[0028] Please as attached Figure 1 The application provides a casting production line energy consumption optimization control method, specifically comprising:

[0029] S1: Real-time collection of energy consumption, production state, quality indicators and equipment safety parameters of multiple process units in the casting production line, and identification of process categories and production section labels for each sampling point.

[0030] S2: Abnormal value elimination, time synchronization and normalization preprocessing of the collected multi-process unit energy consumption data, production state data, quality indicator data and equipment safety parameters to ensure consistency and timeliness of data input.

[0031] S3: Based on different production section labels of each process unit, corresponding multi-objective energy consumption-quality-capacity-safety feature modeling is established, and multi-objective optimization theory is used for parameter distribution modeling to obtain a feature weight distribution matrix for each unit.

[0032] S4: Based on the hierarchical multi-agent architecture, the modeled multi-objective feature weight distribution matrix is input into the corresponding process unit agent, and the energy consumption optimal parameters of each unit are independently solved to generate a candidate optimization parameter vector.

[0033] S5: The candidate optimization parameter vector output by each process unit agent is synchronized and submitted to the global scheduling agent, and through an adaptive reward mechanism combined with reinforcement learning method, the decision weight coefficients of each unit are dynamically adjusted to realize multi-objective balanced optimization of the whole line.

[0034] S6: Using a short window rolling optimization algorithm, based on the weight coefficients and candidate optimization parameter vectors output by the global scheduling agent, the energy consumption optimal parameter settings of each process unit are updated in real time to realize rapid response of multi-objective collaborative parameters of the production line.

[0035] S7: Determine whether the energy consumption optimal parameter settings of the process unit in the current time window meet the real-time response threshold conditions. If the convergence is delayed due to conflicts between energy consumption, capacity and safety targets, switch to the approximate parameter fast solving mode to ensure the timeliness of index adjustment.

[0036] S8: Based on the energy consumption optimal parameter setting results in implementation, continuously collect actual energy consumption, capacity and quality feedback data of the production line, and input the data into the adaptive retraining mechanism to dynamically fine-tune the multi-agent parameters and models to adapt to environmental and task changes.

[0037] S9: When the production line detects abnormal state or significant changes in working conditions, trigger the global scheduling agent to update the parameter group, realize adaptive switching of the multi-agent energy consumption optimization model, and ensure continuous optimization of energy consumption and production indicators of the casting production line under different working conditions and market demands.

[0038] The step S1: collecting the energy consumption, production state, quality index and equipment safety parameter of multiple process units in the casting production line in real time, and identifying the process category and production section label for each sampling point. Specifically, it includes:

[0039] S1.1: Deploying intelligent sensor nodes in each process unit in the casting production line, collecting energy consumption parameters, production state parameters, quality index parameters and equipment safety parameter data packets based on industrial Ethernet or wireless industrial protocol, to obtain the original collection data stream of multiple source physical quantities.

[0040] Intelligent sensor nodes are deployed in each process unit in the casting production line to synchronously collect energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters as the basis for collecting multi-source physical quantity data of process units.

[0041] Industrial Ethernet communication interfaces (such as Profinet, Ethernet / IP, etc.) or wireless industrial protocols (such as WirelessHART, LoRaWAN, etc.) are used to physically access and configure sensor nodes in each process unit, realizing a high-bandwidth data collection network for cross-unit multi-type real-time signals.

[0042] Further, through a networking addressing algorithm (parameters including physical network segment allocation, MAC address binding, node identity authentication), the unique identification and reliable data flow routing of each sensor node are realized, preventing data packet loss and cross-unit mixing phenomenon.

[0043] A multi-channel data collection module is used to periodically poll and collect various types of physical quantity parameter signals, including real-time active / reactive power, current, voltage, temperature, vibration, error alarm and production line state, etc., realizing synchronous sampling between data channels. The multi-channel data collection module includes multiple intelligent sensor nodes, with a collection accuracy of ±0.5%-10% and a sampling period of 1-20 seconds.

[0044] Based on the sensor signal sampling frequency and process variation rate index, the polling collection period, maximum data packet cache time and abnormal timeout retransmission mechanism are parameterized set to ensure the integrity of key energy consumption indicators, product quality parameters and equipment safety parameters and the real-time traceability of sudden events.

[0045] The collected data packets are encrypted through data encryption and link integrity verification algorithm, and the integrity of each data packet is verified by using check code (such as CRC-32), eliminating damaged or abnormal sampling caused by link noise.

[0046] The collected multi-source original physical quantity data is aggregated and managed by taking the timestamp and sensor node ID as the data packet index key, and an original data stream of the process unit is outputted with source identification and complete link information.

[0047] Through the above deployment and collection chain, real-time, full-coverage, high-reliability, traceable data basis of physical quantity in each dimension among multiple process units of the casting production line is realized, which provides solid data support for time sequence synchronization and spatial calibration of downstream algorithm links.

[0048] For example, three types of intelligent sensor nodes are deployed in the melting unit, pouring unit and cooling unit in the casting production line, respectively. The melting unit is provided with a power measurement sensor (with an accuracy of ±0.5%), a thermocouple temperature sensor (with a response time of <1s), a state signal acquisition module and a flue gas detection module; the pouring unit is provided with a production counter, a position sensor and a process quality sensor; and the cooling unit is provided with a circulating water temperature monitoring sensor, a pressure monitoring sensor and a device vibration sensor. The nodes are connected through Profinet industrial Ethernet combined with LoRa wireless backhaul to realize cascading networking, with subnet segment 192.168.10.x for the melting unit and 192.168.20.x for the pouring unit, node device number mapped through MAC address, field networking detection period set to 500ms, and abnormal data packet retransmitted for 3 times. The data collection adopts periodic polling, and a complete data packet is generated every 5 seconds and uploaded after AES-128 encryption. The synchronized data packet is outputted in the following structure:

[0049] Timestamp: 2024-06-15T10:00:05.000Z

[0050] Process unit: melting unit

[0051] Device ID: 192.168.10.5

[0052] Collected parameters: current = 130A, temperature = 1570℃, status code = 0, flue gas concentration = 22mg / m 3

[0053] Data integrity check code: 0x2E4A18C2

[0054] The deployment achieves a packet loss rate of less than 0.01% for key physical quantities of the whole unit and a data integrity check pass rate of 99.99% within one production cycle. The above chain collection and networking guarantee large-scale, multi-dimensional monitoring requirements of the production line, and lay a high-quality data foundation for subsequent time sequence synchronization, process space mapping and abnormality detection based on data stream.

[0055] S1.2: Apply a time synchronization algorithm to the collected raw data stream to achieve the synchronized integration of energy consumption parameters, production state parameters, quality index parameters, and equipment safety parameters under a unified time reference to generate a structured multi-source process unit synchronization data set.

[0056] The multi-source raw data packets collected by multiple process units of the casting production line, including energy consumption parameters, production state parameters, quality index parameters, and equipment safety parameters, are used as input conditions.

[0057] The time synchronization algorithm (reference clock signal synchronization, industrial network protocol time adjustment, data packet timestamp alignment) is used to achieve the synchronized integration of energy consumption parameters, production state parameters, quality index parameters, and equipment safety parameters under a unified time reference.

[0058] Further, through the distributed network clock drift compensation algorithm (parameters: master clock source, synchronization frequency of sub-nodes, maximum allowed drift threshold), the clock offset of each sub-unit data acquisition node is corrected in real time, the data sampling time is strictly aligned across units and protocols, and a global synchronization timestamp field is added to each data packet.

[0059] Further, the multi-channel buffering and queue matching algorithm (parameters: sampling period, data packet migration time window) is applied to merge multiple types of index data within the same sampling period, achieve one-to-one matching between multi-dimensional indexes, and filter redundant or missing records caused by sampling delay and data packet retransmission.

[0060] Further, the abnormal synchronization verification algorithm (such as sampling missing detection and sudden jump time identification) is applied to the data set merging result to interpolate or mark and remove abnormal data records, ensuring the integrity and time sequence consistency of the synchronization data set.

[0061] Through the above processing, a structured multi-source process unit synchronization data set is output, each data set containing standard fields such as process unit label, device ID, synchronization timestamp, energy consumption parameter, production state parameter, quality index, and equipment safety parameter under a unified time base.

[0062] Through the above time synchronization processing algorithm, the original asynchronous, multi-source, and cross-unit acquisition results are converted into a unified time base, structured, and multi-dimensional collaborative synchronization data set, achieving high timeliness and high consistency input of key parameters such as industrial energy consumption and production state, laying a data foundation for subsequent process space calibration and multi-target modeling.

[0063] Exemplary, the multiple types of physical quantity signals are collected synchronously in the smelting unit, the pouring unit, and the cooling unit. The energy consumption and temperature collection period of the smelting unit is 2 seconds, the production and process parameter collection period of the pouring unit is 5 seconds, and the pressure and vibration parameter collection period of the cooling unit is 1 second. The main clock source of the whole line is based on the IEEE 1588 Precision Time Protocol (PTP), and the synchronization frequency is set to be calibrated once per minute. The real-time clock error adjustment threshold of each collection sub-node is ±10 ms. The maximum time window of all unit data packets in the data hub cache area is 5 seconds, the queue matching window is set to ±2 seconds, and the multi-thread data merging algorithm is used to merge the index data in real time. For the group of data lost by the pressure sensor of the cooling unit due to network fluctuations, the system automatically corrects it by using the mean value interpolation of the previous and subsequent time. The final generated standard synchronous data set is as follows: time stamp 2024-06-15T10:05:00.000Z, corresponding to the current of the smelting unit (ID 192.168.10.5) 133 A, the temperature 1573℃, the production of the pouring unit (ID 192.168.20.3) 32, the flow 21L / min, the water pressure of the cooling unit (ID 192.168.30.2) 0.29MPa, and the vibration 15mm / s. This step ensures that all key parameters in the three process links are synchronized and matched in a unified time window, with an abnormality rate of less than 0.5%, providing high consistency time sequence multi-dimensional data for process calibration and real-time multi-objective optimization.

[0064] S1.3: Based on the process unit layout information and sensor deployment planning, a unit space correlation algorithm is performed on the structured multi-source process unit synchronous data set, and all data samples are identified according to the physical process unit to form a process unit data subset with a spatial label.

[0065] S1.4: A production section labeling algorithm is applied to the process unit data subset, and real-time production scheduling information and process flow logs are used to add production section labels to each data sample, realize dual parameter identification of process category and production stage, and generate environment-specific process sampling full parameter set.

[0066] S1.5: The process sampling full parameter set with process category and production section labels is input into the real-time data buffer module, and the latest multi-objective optimization data is accessed by the subsequent preprocessing and feature modeling module through a streaming data caching mechanism to form a data support closed loop.

[0067] The step S2: The collected multi-process unit energy consumption data, production state data, quality index data, and equipment safety parameters are subjected to outlier rejection, time synchronization, and normalization preprocessing to ensure the consistency and timeliness of data input. Specifically, it includes:

[0068] S2.1: Based on statistical discrimination method and physical constraint rule, multi-source outlier detection is performed on the collected energy consumption data, production state data, quality index data and equipment safety parameters, and outlier data and sampling error data exceeding the normal discrimination threshold are removed to obtain a high-confidence process unit original data set.

[0069] The collected energy consumption data, production state data, quality index data and equipment safety parameters are taken as input objects to enter the data preprocessing stage.

[0070] A multi-source outlier detection method (parameters: statistical discrimination threshold, physical constraint upper and lower bounds, abnormal probability coefficient) is used to realize abnormal and outlier sample identification of multi-type original sampling data across process units.

[0071] Further, statistical discrimination methods (such as mean-standard deviation interval detection, box plot quartile distance anomaly detection) are used to perform data distribution analysis on each parameter type, and an abnormality discrimination threshold T is set. For any original sampling value x i , the following conditions are met:

[0072] |x i -μ|>k·σ

[0073] where μ is the expected mean of the current parameter category, σ is its standard deviation, and k is an empirically determined abnormality discrimination coefficient (generally 2-3). This formula realizes outlier detection based on the normal assumption and is suitable for continuous parameters such as energy consumption, temperature and current.

[0074] Further, for parameters with theoretical upper and lower limits or industry safety thresholds, physical constraint rules are used for verification. The specific constraint relationship is:

[0075] L j ≤x i,j ≤U j

[0076] where x i,j is the sampling value of the ith record under parameter j, L j and U j are the physical lower and upper limits of parameter j, respectively. For example, the safety upper limit of a temperature sensor, the allowable current value of a device, the compliance extreme value of flue gas concentration, etc.

[0077] Further, rule-based detection of special sampling errors such as missing codes, no signal, sampling dead values (constant for multiple periods) and failed verification codes is used. By performing dynamic consistency checks on continuous data streams, obvious logical errors or physically impossible data are removed (such as temperature instantaneous cliff-like drop, negative growth of production count, etc.).

[0078] Through the above outlier detection and elimination, a high-confidence process unit raw data set is output, which eliminates noise interference and data artifacts for subsequent time synchronization and normalization processing, and realizes high-quality data input.

[0079] For example, for the data set collected in the five-minute cycle of the smelting unit of the casting production line, the average value of the current collection is 125 A, the standard deviation is 10 A, and the abnormality discrimination coefficient is k=3. Therefore, the current discrimination calculation interval is [95 A, 155 A]. The detection found that the single recording current = 180 A, which satisfies |180-125| = 55 > 30, and is determined as abnormal.

[0080] The temperature parameter safe value interval is set to [1400℃, 1700℃], and the collected value 1570℃ is passed, and 1800℃ is physically constrained and filtered.

[0081] Further inspection of the yield counter found that the count value did not change for 3 consecutive sampling periods, and combined with the equipment working state, it was determined to be a dead value, and the data of this link was eliminated.

[0082] Through multiple rounds of elimination, the total amount of raw data is 1000, and after abnormal value detection and physical constraint filtering, 990 usable high-confidence data are retained, the abnormality elimination rate is 1%, and the residual abnormality rate is less than 0.1% after manual review. The output process unit raw data set provides an ideal data basis for the subsequent steps of time synchronization and normalization algorithm, significantly improving the stability and precision of the multi-objective model.

[0083] S2.2: Based on the process unit raw data set after eliminating abnormal values, a unified clock reference and an industrial synchronization protocol are used to align the sampling time stamps. The time trajectory resampling and interpolation are performed on the multi-source data of different process units and different collection frequencies to form a time-synchronized data stream, ensuring that all data samples can be compared under the same time dimension.

[0084] S2.3: For the energy consumption data, production state data, quality index data and equipment safety parameters after time synchronization processing, interval scaling, Z-score standardization or maximum and minimum normalization algorithms are applied, and the dimensions of each feature are unified. The numerical range is mapped to the standard interval, and the normalized feature matrix after normalization is output to eliminate the modeling bias caused by the differences in physical quantity units and value scales.

[0085] S2.4: For the normalized feature matrix, data consistency verification is performed, and interval overlap ratio verification, feature correlation analysis and other methods are used to correct abnormal coupling relationship and residual deviation, and ensure that the data used for subsequent multi-objective feature modeling meets the industrial control level consistency standard in statistical characteristics and interactive coupling relationship.

[0086] S2.5: Based on the feature matrix passed by the consistency check, generate the final standard input data set with single collection time index, provide high timeliness, high reliability, high consistency input data for subsequent multi-objective energy consumption-quality-production capacity-safety feature modeling module, close the information port of data preprocessing and feature modeling.

[0087] The step S3: based on the different production section labels of each process unit, respectively establish the corresponding multi-objective energy consumption-quality-production capacity-safety feature modeling, use multi-objective optimization theory to carry out parameter distribution modeling, so as to obtain the feature weight distribution matrix of each unit.

[0088] As shown in Figure 2 , specifically includes:

[0089] S3.1: Group and archive the process unit energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters after the preprocessing of the previous data, and construct an association mapping table based on the production section labels of each sampling point to generate a process unit-production section-multi-parameter triple as the input of feature modeling.

[0090] After the preprocessing of the previous data (including outlier rejection, time synchronization, normalization), the process unit energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters are taken as input objects, and enter the input archiving stage of feature modeling.

[0091] A multi-level parameter archiving method (parameters: process unit code, production section label, time stamp index) is used to realize the grouping and arrangement of various parameter data according to the corresponding process unit number and real-time production section identifier.

[0092] Further, through the association mapping table construction algorithm (parameters: the primary key is the joint label of process unit-production section, and the additional index is the collection time stamp), the one-to-one correspondence of energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters in multi-dimensional space is realized, all parameter samples are mapped and normalized to the standard data structure framework, and the association mapping table Schema is generated, for example:

[0093] Process unit ID|Production section label|Collection time|Energy consumption parameter X1|Production state parameter X2|Quality index X3|Equipment safety parameter X4

[0094] Further, through the data iteration archiving algorithm, according to the process unit space identifier and production section flow log in the collection data set, the sampling points in the historical window are filtered one by one, each sampling sample is classified into a unique process unit-production section group according to the process unit number, production section label and time index, and the parallel archiving and grouping management of multi-parameter fields are realized.

[0095] Further, through the field aggregation and verification mechanism (parameters: field integrity threshold, sampling missing re-filling strategy), the data integrity of the above grouped archived objects is tested, and abnormal samples with serious sampling missing, label mismatch or inconsistent flow information are automatically removed, ensuring that the archived triple parameter set has consistency and high confidence.

[0096] Through the above multi-level grouping archiving and mapping table construction processing method, the process unit-production section-multiple parameter triple data format is output as the standardized input of multi-objective feature coupling modeling, realizing the data fusion of multi-objective parameters in the process unit and production section task scenario.

[0097] For example, in the smelting process of a certain foundry workshop, the collected data covers 3 process units (numbers A, B, C) and their respective 3 typical production sections (such as preheating, smelting, and casting), and each data sample is collected with a time window of five minutes, with energy consumption as current (A), production status as speed (rpm), quality as temperature (℃), and equipment safety as pressure (bar). Using the above grouping archiving method, the archiving is based on (process unit ID, production section label) as the primary key and the collection time as the index, and a correlation mapping table with structured fields is constructed (for example: A-smelting-10:00-127 A-1500 rpm-1600℃-4bar). During the archiving process, a small amount of data is found to be missing, and the missing safety parameters are filled in by the mean interpolation strategy, and the total number of archived samples reaches 5400, with an average of 1800 per production section, and the field integrity rate is more than 99%. The final output of the standardized process unit-production section-multiple parameter triple realizes a high-quality data base for subsequent PCA, mutual information analysis and optimization modeling, effectively reduces the model bias and reasoning errors caused by inconsistent data structures, and improves the modeling efficiency and prediction accuracy.

[0098] S3.2: Based on the triple feature modeling input, a multi-objective feature extraction algorithm (such as principal component analysis PCA, mutual information analysis, etc.) is used to analyze the correlation between energy consumption parameters, production status parameters, quality index parameters and equipment safety parameters, and to extract multi-objective coupling feature vectors between parameters to reflect the technical constraints between energy consumption and production index targets.

[0099] After data preprocessing (including outlier removal, time synchronization, normalization processing) and grouping archiving (forming process unit-production section-multiple parameter triple), the feature modeling input is taken as the object, and enters the multi-objective feature mutual correlation analysis stage.

[0100] The data matrix of the energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters is centrally processed by using a principal component analysis (PCA) method (parameters: covariance matrix construction dimension = 4, principal component contribution rate threshold ≥ 90%), and a characteristic covariance matrix is calculated to obtain the linear correlation structure between the parameters. Further, the eigenvalue and eigenvector decomposition of the covariance matrix is performed by using an eigenvalue decomposition algorithm, the first k principal components with a cumulative variance contribution rate greater than 90% are screened, a parameter dimension reduction mapping matrix is generated, and the principal component feature extraction of the multi-parameter space is realized.

[0101] Further, the nonlinear mutual correlation between the parameters in the above principal components and original feature combinations is measured by using a mutual information analysis method (parameters: discretization step, mutual information evaluation threshold). Specifically, the mutual information of two variables (such as process parameters X i and X j ) is calculated by the following formula:

[0102]

[0103] where P(x i , x j ) is the joint probability distribution, P(x i ) and P(x j ) are the marginal probability distributions.

[0104] Further, the multi-objective coupling attributes of each target (energy consumption, production capacity, quality, safety) parameter are screened by using the principal component feature contribution analysis and mutual information analysis results, and the feature selection rule is set: the features with the principal component variance contribution rate and mutual information value both exceeding the preset threshold are identified as important multi-objective coupling features. The screening result is output as a multi-objective coupling feature vector in the process unit-production section scenario, and the elements thereof reflect the coupling constraints, causal correlations or constraint relationships between the parameters.

[0105] Further, on the basis of the feature vector screening, a parameter normalization transformation method is used to ensure that each coupling feature component meets the unified dimension and distribution interval, and a standardized multi-objective coupling feature matrix is formed.

[0106] Through the above algorithm chain, the process unit-production section multi-parameter triple data is converted into a multi-objective coupling feature vector that can reflect the technical constraint relationship between energy consumption and production multi-objectives, and provides a structured input for subsequent multi-objective optimization and parameter weight modeling, and realizes the scientific quantification of the trade-off coupling relationship between different production objectives.

[0107] Exemplarily, 1440 groups of data (every 5 minutes / day) are collected at a smelting unit-smelting section of a casting production line. Covariance calculation is performed on the matrix of the normalized energy consumption parameter (current), production state (rotational speed), quality (alloy temperature), and equipment safety (pressure) by using the PCA method, and the covariance matrix is:

[0108]

[0109] Eigenvalues and eigenvectors are calculated for the covariance matrix, and the cumulative contribution rate of the first two principal components reaches 92%. Mutual information analysis is performed on each pair of parameters, and the energy consumption-quality mutual information is calculated to be 0.21, the energy consumption-safety mutual information is calculated to be 0.17, and the quality-safety mutual information is calculated to be 0.33 (unit: bit), all of which are higher than the threshold value 0.15. Four-dimensional sub-features with high principal component contribution rate and mutual information are screened, and the normalized multi-objective coupling feature vector [0.78, 0.72, 0.89, 0.81] is formed. The feature vector is used as the core input of subsequent optimization parameter distribution modeling, and effectively describes the quantitative constraint relationship between energy consumption and production capacity, quality, and equipment safety targets in the smelting section.

[0110] S3.3: The multi-objective coupling feature vector is subjected to parameter distribution modeling by using multi-objective optimization theory (such as Pareto frontier analysis and analytic hierarchy process AHP), a feature weight evaluation system of each process unit under a specified production section is constructed, and a four-dimensional parameter weight matrix of energy consumption-quality-production capacity-safety of the process unit-production section is output.

[0111] The multi-objective coupling feature vector of the process unit-production section obtained by the multi-objective feature extraction algorithm is used as the input object, and enters the parameter distribution modeling and weight system construction stage.

[0112] The parameter distribution identification of non-dominated solutions among the targets is realized by using the multi-target Pareto frontier analysis method (parameters: feature vector dimension = 4, target type = energy consumption, quality, production capacity, and equipment safety). By constructing a multi-objective optimization problem, the four targets of energy consumption, quality, production capacity, and safety are set as the optimization dimensions F = [f1, f2, f3, f4], and the following Pareto optimal discrimination formula is applied to the multi-objective feature vector x of each process unit-production section:

[0113]

[0114] wherein, is the feasible parameter set, and f k represents the evaluation function of the kth target.

[0115] ​​​Further, the quantitative evaluation of the weight relationship among the multi-objective features is realized by an analytic hierarchy process (AHP) algorithm (parameters: consistency ratio CR < 0.1 of the pair-wise judgment matrix). A pair-wise comparison matrix A = [a ij ] 4×4 is constructed for energy consumption, quality, production capacity, and equipment safety, and the consistency ratio value is calculated by using expert scoring / historical data attribution:

[0116]

[0117] wherein λ max is the maximum eigenvalue of the matrix, n = 4, and RI is the random consistency index (RI = 0.90 is obtained by table lookup). The feature weight vector w is calculated by solving:

[0118] Aw = λ max w

[0119] and normalized to obtain the relative weights of each target parameter under the process unit-production section.

[0120] Further, the four-dimensional parameter weight matrix W = [w ij ] m×4 is generated by a multi-objective coupling optimization weight integration method combining the Pareto frontier solution set and the AHP weight vector, wherein m is the number of production sections, and each row represents the multi-objective weight distribution under the specified process unit-production section.

[0121] Further, a multi-objective weighted distribution consistency checking method (parameters: KL divergence threshold, weight and normalization error limit) is used to ensure the consistency of the distribution form, the sum of 1, and the rationality of the correlation structure between the targets.

[0122] Through Pareto frontier analysis and AHP hierarchical analysis method, the multi-objective feature coupling vector is converted into a parameter weight distribution matrix with priority ranking and constraint balance, realizing the quantitative reflection of the evaluation system output of energy consumption, quality, production capacity, and equipment safety multi-objectives.

[0123] For example, for a smelting unit-heating section, the multi-objective coupling feature vector obtained by PCA and mutual information analysis is [0.76, 0.82, 0.91, 0.81], and the non-dominated solution extraction is performed on 365 groups of daily data samples by Pareto frontier analysis, obtaining 30 groups of typical Pareto optimal solutions. For this section, the AHP algorithm is used to construct the target pair-wise judgment matrix:

[0124]

[0125] After eigenvalue decomposition, λmax = 4.11, CI = 0.0367, CR = 0.0407 < 0.1, matrix consistency passed. The de-normalized eigenvector is [0.186 0.372, 0.082, 0.360], reflecting that the production section weighs the quality and safety targets more heavily than energy consumption and production capacity. In combination with the sensitivity of each target to process performance in the Pareto optimal solution set, the above weights are further smoothed using weighted fusion to generate the energy consumption-quality-production capacity-safety four-dimensional weights of the section as [0.18, 0.37, 0.08, 0.37].

[0126] In the 8 process units and 24 production sections of the entire production line, the weight distribution matrix is calculated in batches using the above method section by section. The sum of the weights of all distribution matrices is 1, and the KL divergence is less than 0.01, ensuring the parameter distribution form and target balance. The weight evaluation system serves as the standard input basis for subsequent parameter normalization and multi-agent collaborative optimization, providing quantitative basis for global energy consumption optimization and multi-target constraint balance.

[0127] S3.4: Based on the obtained parameter weight evaluation system, a feature normalization algorithm (such as Min-Max normalization) is used to normalize the energy consumption-quality-production capacity-safety multi-target weight matrix under the process unit, so that the weight distribution between different targets has comparability and parameter consistency.

[0128] S3.5: The normalized multi-target weight matrix of each process unit-production section is summarized and managed to generate a process unit layered multi-target feature weight distribution database, providing one-to-one parameter weight input for subsequent layered multi-agent optimization decision-making, realizing the technical closed loop of feature recognition, target balance and parameter tuning.

[0129] The step S4: relying on the layered multi-agent architecture, the modeled multi-target feature weight distribution matrix is input into the corresponding process unit agent to independently solve the energy optimal parameters of each unit and generate a candidate optimization parameter vector. As shown in Figure 3 , specifically comprising:

[0130] S4.1: The multi-target parameter input initialization processing is performed on the multi-target feature weight distribution matrix output by step S3 to standardize the feature weight distribution matrix format and accurately map to the data interface of each process unit agent, ensuring the consistency and processability of the input data in the agent.

[0131] The normalized multi-target feature weight distribution matrix of the process unit-production section output by step S3 is used as input data to enter the multi-target parameter input initialization processing stage, which is used to map to the data interface of the process unit agent to ensure parameter format standardization and data consistency.

[0132] The data normalization format conversion method (parameters: field order, data type, dimension constraint) is used to realize the unified field sorting, numerical structure standardization and data type correction of the original feature weight distribution matrix, so that the weight distribution matrix under each process unit-production section meets the unified demand specification of the agent for input parameters.

[0133] Further, through the parameter template mapping algorithm (parameters: process unit label mapping table, production section index rule), the accurate collaborative mapping relationship between the multi-objective weight matrix and the internal parameter port of each process unit agent is realized, and each set of normalized weight vector is attached with the standardized unit ID and production section index label, and the standardized parameter mapping tuple is output.

[0134] Further, a consistency checking mechanism (parameters: feature dimension constraint = 4, weight sum constraint = 1.0, data legal interval = [0, 1]) is used to perform consistency check on each parameter tuple after mapping, including feature quantity, weight sum and effective interval limitation of each weight component, and abnormal reminder or missing value refilling processing is performed on the data samples that do not pass the test, to ensure that the structure and specification of all data inputs meet the processing needs of the downstream optimization algorithm.

[0135] Further, through the data buffering and high-concurrency distribution mechanism, the parameter tuples that have passed the verification are real-time enqueued to the data input queue of each agent according to the interface calling strategy of the process unit agent, realizing low-latency and high-reliability data pushing, and providing a basic parameter set for independent solution for subsequent agent multi-objective parameter optimization task instantiation.

[0136] Through the above multi-level normalization processing and consistency checking, the multi-objective feature weight distribution matrix standardized by the S3 step is converted into a data interface parameter that can be directly input into the agent, realizing the overall compatibility of the data structure, and providing a high-quality data basis for independent solution of the hierarchical multi-agent energy consumption optimization parameter.

[0137] For example, in the smelting process unit of the casting production line, the smelting, refining and casting production sections correspond to the multi-objective feature weight distribution matrix respectively:

[0138] Smelting section [0.21, 0.44, 0.12, 0.23]

[0139] Refining section [0.15, 0.51, 0.10, 0.24]

[0140] Casting section [0.25, 0.36, 0.15, 0.24]

[0141] The data is organized in json structure, and each element contains process unit ID (such as E001), production section code (such as S01), standard normalized weight array and verification attribute.

[0142] The weights of all elements are first sorted in the order of [energy consumption, quality, production capacity, safety] using a field sorting and type correction mechanism, and the float32 format is used for data type.

[0143] A consistency check formula is used:

[0144]

[0145] where w i is the normalized weight of each target parameter.

[0146] After the check is passed, the weight tuple of each segment is pushed to the corresponding smelting process unit agent queue using the data distribution thread, ensuring that the queue delay is within 50 ms and the packet loss rate is less than 0.01%. Finally, these standardized parameter tuples serve as key inputs for independent solution of multi-objective optimization by agents, verifying the consistency and structure matching of parameter inputs for process unit agents, and providing pre-data assurance for normal operation of subsequent algorithms.

[0147] S4.2: Based on the hierarchical multi-agent architecture, the multi-objective feature weight distribution matrix corresponding to each process unit is input into the local process unit agent Agent using a parameterized interface to establish a multi-objective optimization task instance owned by the agent, providing an input parameter set for subsequent independent solution of energy-optimal parameters.

[0148] Under the hierarchical multi-agent architecture, the process unit-production segment multi-objective feature weight distribution matrix processed by standardization and interface mapping is input, and the object of action is the local agent Agent distributed in each process unit.

[0149] A parameterized interface adaptation algorithm (parameters: field order mapping table, unit check template, interface version protocol) is used to rearrange the multi-objective feature weight distribution matrix of each process unit-production segment at the byte level and perform unit consistency check according to the data interface format preset by the agent Agent, and the fields are bound one by one to ensure that the order, data type, and resolution of the weight parameters completely adapt to the data structure definition of the Agent end.

[0150] Further, through dynamic memory allocation and buffer registration methods (parameters: memory segmentation threshold, data write priority), the multi-objective weight distribution matrix is loaded in a structured manner. For multiple process unit weight data written concurrently, an independent buffer area is automatically allocated to achieve efficient scheduling and writing of data streams, avoiding data loss and congestion when accessing the interface.

[0151] Further, a multi-objective task instance generation algorithm (parameters: weight matrix data integrity check code, Agent identity, production segment index) is implemented in the local intelligent agent Agent to automatically create an independent multi-objective optimization task instance based on the input multi-objective weight distribution matrix, and assign a unique context ID to each task instance for subsequent task state tracking and parameter dynamic updating.

[0152] Further, by the parameter input consistency verification method (parameters: field consistency comparison threshold, illegal value error detection table), the input weight parameters in the successfully established multi-objective optimization task instance are checked for field consistency, abnormal data is automatically identified and corrected according to the plan, and the algorithm call in the subsequent optimization link has high data reliability.

[0153] Through the above parameterized interface mapping, multi-objective task instance generation and consistency verification chain processing, the multi-objective feature weight distribution matrix of the process unit after normalization and standardization is robustly converted into an input parameter set that can directly drive the local multi-objective optimization of the intelligent agent Agent, realizing the rapid landing of multi-objective dynamic task instances at the intelligent agent end of each process unit, and laying a traceable and expandable data foundation for subsequent independent optimization and solution of the energy consumption optimal parameters of each process unit.

[0154] For example, in the smelting unit-preheating section of the casting production line, the input data is the standardized mapped multi-objective weight distribution matrix [energy consumption: 0.20, quality: 0.30, capacity: 0.25, safety: 0.25], the field precision is set to two decimal places, and the data type is float. The parameterized interface adaptation algorithm automatically compares the target interface requirements, rearranges the field order to [capacity, quality, energy consumption, safety], allocates 32 bytes of memory buffer, and adds a check code to each input weight. The interface landing time is less than 10 ms. After the weight data is written, the intelligent agent Agent creates a TaskInstance_20240613_0001 task instance using the multi-objective task generation algorithm, and the unique identifier is bound to the smelting-preheating section-20240613. Each field is found to meet the protocol requirements after consistency verification. This instance successfully calls the above parameters in the subsequent multi-objective optimization phase to drive the smelting unit preheating section energy consumption optimal parameter solving process. During the task instance running period, the weight dynamic attribute supports real-time changes with process conditions, the verification interface reliability is 100%, and there is no record of task instance initialization failure. The final output is the local Agent multi-objective optimization task context and input parameter set, which efficiently supports the optimal parameter adaptive solving and rolling update under the dynamic weight of the smelting preheating process.

[0155] S4.3: According to the multi-objective collaborative optimization algorithm built in each process unit agent, the multi-objective function is constructed by inputting the multi-objective characteristic weight distribution matrix and the basic data such as energy consumption and production state collected by the unit, and the multi-objective optimization solution is realized under the constraints of energy consumption, production capacity, quality and equipment safety.

[0156] S4.4: In each process unit agent, intelligent optimization algorithms such as reinforcement learning and multi-objective evolution are used to iteratively search for the optimal parameter combination, further optimize the parameter solution by comprehensively considering the priority of each target and the constraint condition, and finally obtain the candidate solution set of the optimal energy consumption parameters of the process unit.

[0157] S4.5: The candidate optimization parameter vector output by each process unit agent is subjected to effectiveness and consistency test, and the solutions that do not meet the real-time constraints or violate the safety boundary are excluded, and the effective candidate optimization parameter vector is output to the upper layer global scheduling agent in a standard protocol, providing a basic solution set for global multi-objective optimization and fast parameter response.

[0158] The step S5: The candidate optimization parameter vector output by each process unit agent is synchronized and submitted to the global scheduling agent, and through the adaptive reward mechanism combined with the reinforcement learning method, the decision weight coefficient of each unit is dynamically adjusted to realize the balanced optimization of multi-objective of the whole line. Specifically, it includes:

[0159] S5.1: Based on the legal process unit label and timestamp, the candidate optimization parameter vector generated by each process unit agent is subjected to synchronous packaging and transmission scheduling processing to obtain a multi-process unit candidate optimization parameter set in a standardized manner, ensuring the timeliness and integrity of the data in the cross-unit global scheduling process.

[0160] Based on the legal process unit label and timestamp information, the parameter attribution identification standardization method (parameters: process unit ID, production segment label, timestamp format definition, priority label) is used to realize the unique identification and traceability information completion of the multi-objective parameter vector.

[0161] A multi-dimensional data synchronous packaging algorithm (parameters: packaging window length, synchronization tolerance, packet loss compensation strategy) is used to collect the candidate optimization parameter vectors output by each process unit agent within a specified time window, and efficient synchronous packaging is realized according to a unified data structure (such as JSON, ProtocolBuffer, etc.), and a temporary parameter data set across process units is obtained.

[0162] Further, by the protocolized data scheduling distribution method (parameters: data encryption check field, network QoS level, real-time threshold setting, buffer writing management), the parameter data set after synchronous packaging is processed, and according to the access protocol and physical network topology of the global scheduling intelligent agent, the candidate optimization parameter vector of each process unit is sent to the data receiving interface of the global scheduling intelligent agent in a synchronous and concurrent manner, ensuring the integrity and real-time arrival of the data.

[0163] The transmission consistency check algorithm (parameters: data packet sequence number, hash checksum, retransmission window) is used to check the sequence consistency and content integrity of the transmitted data packet. For abnormal packet loss or order disorder data, automatic retransmission and data compensation mechanism is used for repair to ensure the correct reception of the candidate optimization parameter set of the multi-process unit at the global scheduling end.

[0164] Through the data buffer consistency comparison method (parameters: time synchronization tolerance interval, parameter floating interval), the received multi-process unit candidate optimization parameter vector set is compared for window consistency, and samples with time drift or abnormal floating are removed, and the standardized multi-process unit candidate optimization parameter set for global multi-objective collaborative optimization is output.

[0165] Through the above algorithm chain processing, the independent parameter optimization results distributed in each process unit intelligent agent are synchronized, efficiently and reliably gathered to the global scheduling intelligent agent, realizing the standard input set of the multi-process unit candidate optimization parameters supporting the subsequent global multi-objective optimization and weight dynamic adjustment.

[0166] For example, the smelting unit (ID: E001, production section: S01), the refining unit (ID: E002, S02) and the casting unit (ID: E003, S03) generate candidate optimization parameters in the same rolling window by their respective intelligent agents, such as:

[0167] E001-S01 generates parameters [0.31, 0.45, 0.09, 0.15] with label {ID: E001, S01, ts: 20240613T095900, Prio: 1}.

[0168] E002-S02 generates parameters [0.28, 0.53, 0.07, 0.12] with label {ID: E002, S02, ts: 20240613T095900, Prio: 2}.

[0169] E003-S03 generates parameters [0.33, 0.38, 0.14, 0.15] with label {ID: E003, S03, ts: 20240613T095900, Prio: 3}.

[0170] A packing window width of 100ms is used, with a tolerance of 5ms for time difference. All parameters are packaged and synchronized in float32 format. Distribution uses an encrypted channel of the TCP protocol, the network QoS level is real-time, and the data packets are checked by CRC32. After the packaging and synchronization are completed, the parameter sets are sorted in ascending order by timestamp on the global scheduling agent side, and the consistency check passes without packet loss or disorder. The final output is three standardized structure arrays with unique labels and multi-objective parameters, which successfully support global multi-objective optimization and subsequent dynamic adjustment of the reward mechanism weights. This step improves the timeliness of the parameter set to sub-second level and 100% packet integrity, laying a data foundation for global collaboration between complex process links.

[0171] S5.2: Based on the candidate optimization parameter sets of multiple process units, the inter-agent communication protocol is used to normalize and aggregate the parameter vectors to form a multi-objective parameter decision matrix to support the subsequent adaptive reward mechanism operation of the global scheduling agent.

[0172] S5.3: Relying on the global scheduling agent, with the multi-objective parameter decision matrix as the input condition, an adaptive reward mechanism (such as a reward signal generator based on reinforcement learning) is adopted to calculate the reward score based on the historical contribution and real-time performance of the candidate optimization parameter vector of each unit to obtain the adaptive reward vector of multiple process units.

[0173] S5.4: Based on the adaptive reward vector of multiple process units, the weight update algorithm in the reinforcement learning method is used to dynamically adjust the decision weight coefficient of each unit in the multi-objective parameter decision matrix, output the adaptive decision weight distribution matrix, and achieve the optimal weight configuration under the global collaboration of multiple intelligent agents.

[0174] S5.5: Using the output adaptive decision weight distribution matrix, combined with the initial multi-process unit candidate optimization parameter set, execute the multi-objective balance optimization algorithm of the global objective function to obtain the global optimal energy consumption parameter vector of the multi-objective weighted balance, providing the input basis for the next step of real-time parameter setting and response formation.

[0175] Step S6: Using a short window rolling optimization algorithm, based on the weight coefficients and candidate optimization parameter vectors output by the global scheduling agent, the optimal energy consumption parameter settings of each process unit are updated in real time to achieve rapid response of the multi-objective collaborative parameters of the production line. Specifically including:

[0176] S6.1: Based on the weight coefficients output by the global scheduling agent and the candidate optimization parameter vectors of each process unit, the short window rolling optimization algorithm is called to perform preliminary mapping calculations on the multi-objective parameter distribution set within the current time window to obtain the time-series updated input of the optimal energy consumption parameters of each process unit.

[0177] The input data includes an adaptive decision weight coefficient distribution matrix of the global scheduling agent output and a candidate optimization parameter vector set of each process unit in the current time window.

[0178] A short window rolling optimization algorithm (parameters: rolling window length T_w, step size Δt, objective function F_multi) is used to realize preliminary mapping calculation of the multi-objective parameter distribution set in the current time interval.

[0179] Further, through the window data acquisition module (parameters: current system time t_now, window start and end time [t_{start}, t_{end}]), the candidate optimization parameter vector and the corresponding weight of each process unit in each rolling window are extracted.

[0180] Further, the window parameter mapping algorithm (parameters: each unit weight coefficient w_i(t), candidate optimization parameter vector P_i(t)) is applied to combine all parameter vectors in the window according to the weight, and the weighted optimal parameter initial value of each process unit is obtained, which satisfies the following formula:

[0181]

[0182] wherein, w_i(t) is the weighted optimal parameter initial value of the i-th process unit at time t, w_i(t) is the weight coefficient allocated by the current global scheduling agent, and P_i(t) is the candidate optimization parameter vector of the corresponding process unit. i i

[0183] Further, the window mapping sequence generator (parameter: window cumulative step number N_w) is used to statistically analyze the weighted optimal parameter sequence under each step in the time window, select the optimal target set point in the interval, and form the time sequence update input of the process unit energy optimal parameter.

[0184] Through the short window rolling optimization algorithm, the multi-objective parameter distribution set normalized in the previous step is converted into the weighted optimal parameter sequence of each process unit, realizing the energy parameter time sequence input that meets the current multi-objective trade-off and dynamic scheduling state, and laying a data foundation for subsequent dynamic reorganization and feasibility judgment.

[0185] ​​Exemplarily, in a certain short window optimization cycle of the casting production line, assuming that the step length At is 5 seconds and the window length Tw is 20 seconds. The weight coefficients of each process unit output by the global scheduling agent are as follows: the smelting unit w 1 (t) = 0.32, the refining unit w 2 (t) = 0.41, and the casting unit w 3 (t) = 0.27. The candidate optimization parameter vectors output by the smelting unit in the window are [0.70, 0.76, 0.93, 0.82], [0.68, 0.75, 0.94, 0.81], [0.69, 0.74, 0.92, 0.84], and [0.71, 0.73, 0.91, 0.85] in sequence.

[0186] The weighted optimal parameter initial value of each step is calculated by using the following recursive formula:

[0187]

[0188] Taking the first step (t1) as an example, the following is obtained:

[0189] By analogy, the parameter weighting results in the four steps are counted, and the interval average is finally used as the timing update input of the smelting unit in the window: [0.223, 0.237, 0.294, 0.264]. The refining and casting units are calculated in the same way.

[0190] In this example, through the short window rolling optimization algorithm, the dynamic mapping and small step update of the multi-objective parameter are efficiently realized. The timing input of the energy optimal parameter finally output significantly improves the dynamic response speed, effectively supporting the adaptive adjustment and rolling optimization control of the subsequent sub-steps.

[0191] S6.2: Taking the multi-objective parameter distribution set output by the previous sub-step as input, applying timing normalization and dynamic feature weighting processing, and performing dynamic reorganization on the candidate optimization parameter vectors of each process unit to realize adaptive adjustment of the energy optimal parameters of the process unit in the new weight background.

[0192] S6.3: For the energy optimal parameters of each process unit after dynamic reorganization, the parameter feasibility verification algorithm (such as safety constraint detection and core process consistency judgment) is used to exclude the infeasible settings caused by multi-objective conflicts, and output the parameter adjustment suggestions after constraint verification.

[0193] S6.4: According to the parameter adjustment suggestions after verification, the parameter rolling optimization control strategy is implemented to update the actual parameter settings of each process unit in real time, and generate parameter application records with time period labels, realizing the closed-loop rolling deployment of the energy optimal parameters.

[0194] S6.5: Collect and aggregate the energy consumption parameter running state, process unit production index and external working condition changes after the implementation of rolling optimization regulation, and dynamically correct the weight coefficient and parameter window in the short window rolling optimization algorithm using real-time state feedback mechanism to enhance the immediate adaptability of parameter setting to sudden changes in working conditions and task changes.

[0195] The step S7: judging whether the process unit energy consumption optimal parameter setting in the current time window meets the real-time response threshold condition, if the convergence is delayed due to the conflict of energy consumption-capacity-safety targets, switching to the approximate parameter fast solving mode to ensure the timeliness of index adjustment. Specifically includes:

[0196] S7.1: Perform convergence determination processing on the energy consumption optimal parameter setting results output by each process unit agent in the previous time sequence window, analyze the target function convergence speed and index target error change trend of the parameter vector based on the convergence determination algorithm to obtain the convergence state identification result.

[0197] The energy consumption optimal parameter setting results output by each process unit agent in the previous time sequence window are used as input, and the parameter history cache module is used to retrieve all parameter vectors and target function values under synchronous time sequence.

[0198] The convergence determination algorithm (optional algorithms include target function error threshold determination based on first-level criterion, target function convergence speed analysis based on second-level criterion, and parameter disturbance trend detection based on three-dimensional trajectory) is used to analyze the parameter vector dimension by dimension, realizing the preliminary screening of the convergence of parameter setting results.

[0199] Further, the sliding window target function difference algorithm is used to analyze the target function value Perform multi-step difference:

[0200]

[0201] Wherein, is the comprehensive multi-objective function value of the i-th process unit at time t, is the target function change amount in a single cycle.

[0202] By comparing the absolute value of with the set convergence threshold ∈, when and the criterion is met for consecutive windows, further dynamic normalized error trend determination is performed to analyze the target function convergence speed:

[0203]

[0204] Wherein, h is the sliding window step size.

[0205] Further, the cumulative convergence speed With the target error change sequence, the trend entropy and change rate sensitivity combined algorithm is used to determine the convergence trend of the current parameter setting. If the convergence speed meets the actual threshold requirement, a convergence success state identifier is generated, otherwise it is marked as not converged or convergence delayed.

[0206] Through the above full-link convergence determination algorithm, the energy-optimal parameter setting result of the last window is converted into a standardized convergence state identifier, realizing the chain closed loop of process unit level convergence monitoring and subsequent response.

[0207] For example, in the variable speed lifting section process unit of the casting production line, the energy-optimal parameter is the heating power setting Feed rate Pressure control After 10 optimization cycles of continuous operation, the system collects The moving average is 0.05, and the sliding window step h is 5, Less than the convergence determination threshold ∈ = 0.1. Through trend entropy and rate analysis, the parameter fluctuation rate in 10 cycles is reduced to 0.5%, and the system outputs the convergence success state identifier. If in the high temperature impact section, due to process disturbance The change in the continuous window is more than 0.3, and the rate Higher than the threshold, the system outputs the non-convergence or convergence delay state. The above results are respectively used as the input basis for subsequent real-time target threshold detection and fast approximate solution mode. Through the above convergence determination step, it is ensured that the parameter optimization algorithm result has dynamic monitoring and real-time response capability, and provides algorithm basis for subsequent index timeliness adjustment.

[0208] S7.2: Based on the convergence state identifier result, calculate the target response delay degree corresponding to the current energy-optimal parameter setting and the real-time response threshold condition, use the threshold detection algorithm to judge whether the energy-optimal parameter setting meets the preset real-time response requirement, and generate a threshold determination signal.

[0209] S7.3: For the process unit parameter set that does not meet the real-time response threshold, analyze its target conflict type, use the multi-objective conflict identification algorithm to classify the deviation relationship among the energy, capacity, safety and other objective functions, and attribute it to the cause structure of the current multi-objective optimization model, to generate a target conflict type identifier.

[0210] S7.4: For the energy-optimal parameter setting result that has been determined to be convergence delayed and has a multi-objective conflict type identifier, trigger the parameter fast approximate solution algorithm, input the threshold determination signal, the target conflict type identifier and the last round parameter vector, and use the parameter fast approximate inference mechanism based on heuristic or responsive algorithm to generate an approximate parameter optimization solution.

[0211] S7.5: Fuse the generated approximate parameter optimization solution with the original energy consumption optimal parameter setting result, select the most priority parameter result based on the parameter fusion decision algorithm, and update the process unit energy consumption optimal parameter setting output to realize dynamic switching of energy consumption optimal parameters and timeliness guarantee of index adjustment.

[0212] The step S8: based on the energy consumption optimal parameter setting result in implementation, continuously collect the actual energy consumption, capacity and quality feedback data of the production line, input the data into the adaptive retraining mechanism, dynamically fine-tune the multi-agent parameters and models to adapt to the environment and task changes. Specifically, it includes:

[0213] S8.1: Continuously monitor and sample the actual process unit of the energy consumption optimal parameter setting result in implementation, obtain real-time energy consumption measurement data, capacity statistical data and product quality feedback data based on the multi-agent distributed sensing nodes to obtain the original feedback data stream representing the current production state.

[0214] S8.2: Input the original feedback data stream into the adaptive feedback processing module, use the time series preprocessing algorithm to perform outlier rejection, time synchronization and data normalization processing on the energy consumption measurement data, capacity statistical data and quality feedback data to generate a high-consistency feedback feature matrix, providing standardized feature input for subsequent model fine-tuning.

[0215] S8.3: Based on the high-consistency feedback feature matrix, apply real-time feature mapping algorithm to analyze the correlation between feedback features and multi-agent historical parameter tuning trajectories to match the functional dependence relationship between master parameters and performance indicators, and generate data label pairs suitable for adaptive retraining.

[0216] Based on the high-consistency feedback feature matrix, the real-time feature mapping algorithm (input includes normalized energy consumption data, capacity data and quality indicator feature vectors) is used to realize the mapping and matching of feedback features to multi-agent historical parameter tuning trajectories.

[0217] Further, through the time series dynamic correlation analysis method, the correlation coefficient matrix between each feedback feature and historical parameter tuning sequence is calculated, and the time-varying dependence relationship between master parameters and production performance indicators is established.

[0218]

[0219] Wherein, is the value of the i-th feedback feature at time t, is the j-th master parameter historical tuning sequence, R ij (t) is the Pearson correlation coefficient of the two.

[0220] Further, through the sliding window feature alignment algorithm (window width W, sliding step S), the feedback features and parameter trajectories in different time periods are analyzed in a window cascade manner, and the main control parameter variation points that cause significant changes in feedback targets are extracted.

[0221]

[0222] wherein S ij (k) is the variation coupling degree of the i-th feature and the j-th parameter in the k window, and is the respective sliding window mean.

[0223] Further, an adaptive label generation algorithm is used to automatically filter strong coupling pairs between feedback features and main control parameters according to the correlation threshold τ and the variation coupling degree threshold σ, and to construct data label pairs for retraining.

[0224] The label pair set obtained by the above screening is used as the data stream input for subsequent model fine-tuning.

[0225] Through the algorithm chain, the high consistency feedback matrix and the historical tuning trajectory are efficiently mapped, and the data label pairs accurately reflecting the actual working condition of the main control relationship are automatically generated, realizing high adaptability fine-tuning of subsequent multi-agent parameters and models.

[0226] For example, in the smelting-casting section of a casting production line, the feedback features include energy consumption P fb , production capacity Q fb and first pass rate Y fb (normalized to 0.65, 0.85 and 0.92, respectively); the historical main control parameters are heating power P hist , pouring rate V hist and mold pressure F hist (normalized to 0.6, 0.8 and 0.9, respectively). For the last 20 time windows, the Pearson correlation analysis is used to calculate the correlation coefficient matrix of the feedback features and the main control parameters, and it is found that all are higher than the set threshold of 0.8. Further calculation of the sliding window variation coupling degree The peak window output is 0.15, which is higher than the threshold of 0.12. Finally, the label pairs (P fb , P hist ) and (Y fb , F hist ) are selected as input data streams, and the data label set suitable for retraining is generated through the above processing, realizing high-precision capture of the dynamic dependence relationship between the energy consumption parameters and the quality indicators of the production line.

[0227] S8.4: The generated data label pair is input into the multi-agent parameter adaptive retraining mechanism, and the reinforcement learning or incremental gradient fine-tuning algorithm is used to perform structural optimization adjustment on the current multi-agent parameter set and its collaborative decision-making model to output the updated multi-agent parameter set and model weight matrix.

[0228] S8.5: Based on the updated multi-agent parameter set and model weight matrix, the key tuning results are pushed to the real-time optimization control module to realize online dynamic adjustment of the multi-agent system; at the same time, the global scheduling agent is fed back the optimized parameter correction amplitude and performance index improvement evaluation data to verify the adjustment effect of the adaptive fine-tuning mechanism in a closed loop.

[0229] The step S9: When the production line detects abnormal state or significant change of working condition, the global scheduling agent updates the parameter set to realize adaptive switching of the multi-agent energy consumption optimization model to ensure continuous optimization of energy consumption and production indicators of the casting production line under different working conditions and market demands. Specifically, it includes:

[0230] S9.1: The process unit energy consumption parameters, production state parameters, quality index parameters and equipment safety parameters collected in real time by the casting production line are applied to the abnormal detection algorithm to identify abnormal states that exceed the preset threshold interval or show trend mutation to obtain abnormal state identification labels as working condition change signals.

[0231] S9.2: Based on the abnormal state identification label, the global scheduling agent triggers the parameter set update module to call the multi-objective feature weight distribution matrix and candidate optimization parameter vector that have been deployed to perform weight dynamic reestimation and parameter linkage adjustment algorithm to obtain a new parameter set suggestion set that adapts to the current working condition change.

[0232] S9.3: The new parameter set suggestion set output by the weight dynamic reestimation and parameter linkage adjustment algorithm is automatically selected by the global scheduling agent using the working condition adaptive switching strategy to select the optimal parameter set and issue an update instruction to each process unit agent to realize adaptive switching and model version management of the multi-agent energy consumption optimization model.

[0233] S9.4: Based on the received optimal parameter set, each process unit agent reloads the updated multi-objective feature weight distribution matrix and optimization parameter settings, and relies on the parameter synchronization protocol to ensure consistent and coordinated stage performance consumption optimization deployment of the multi-agent system at the whole production line level.

[0234] S9.5: After the multi-agent system completes the adaptive switching of the parameter set, the energy consumption data, production capacity feedback information and equipment operation safety indicators of the process unit are collected and analyzed in real time, and the performance after model switching is determined by applying the effect evaluation algorithm, and the evaluation results are fed back to the global scheduling agent to form a continuous adaptive optimization closed loop.

[0235] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0236] The above description is only the preferred embodiments of the present application and is not used to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included in the protection scope of the present application.

Claims

1. A method for optimizing energy consumption of a casting production line, characterized in that: The following steps are involved: S1. Real-time collection of raw data from multiple process units within a casting production line, including energy consumption data, production status data, quality indicator data, and equipment safety parameters, and identification of process categories and production segment labels for each sampling point; S2. preprocessing the collected original data of the multiple process units to obtain a preprocessed multi-process unit data set; S3. Based on the different production segment labels of each process unit, a corresponding multi-objective energy consumption-quality-capacity-safety feature model is established. Parameter distribution modeling is performed using multi-objective optimization theory to obtain the feature weight distribution matrix of each unit. S4. Relying on a hierarchical multi-agent architecture, the multi-objective feature weight distribution matrix is ​​input into the corresponding process unit agent, and the optimal energy consumption parameters of each unit are independently solved to generate a candidate optimization parameter vector; S5. Synchronously submit the candidate optimization parameter vector to the global scheduling agent, and dynamically adjust the decision weight coefficient of each unit through an adaptive reward mechanism combined with a reinforcement learning method; S6. Use a short-window rolling optimization algorithm to update the optimal energy consumption parameter settings of each process unit in real time based on the weight coefficients and candidate optimization parameter vectors output by the global scheduling agent; S7, determining whether the optimal parameter settings for energy consumption of the process unit within the current time window meet the real-time response threshold condition. If convergence is delayed due to target conflict, switching to the approximate parameter fast solution mode; S8. Based on the optimal energy consumption parameter setting results during implementation, continuously collect the actual energy consumption, production capacity and quality feedback data of the production line and input them into the adaptive retraining mechanism to dynamically fine-tune the multi-agent parameters and models.

2. The real-time optimization method for sintering process based on edge computing according to claim 1 is characterized in that: After step S8, the following steps are also included: S9. When an abnormal state or significant change in working conditions is detected on the production line, the global scheduling agent is triggered to update the parameter group to complete the adaptive switching of the multi-agent energy consumption optimization model.

3. The real-time optimization method for sintering process based on edge computing according to claim 1 is characterized in that: The step S1 specifically includes: Intelligent sensor nodes are deployed in each process unit of the casting production line to collect energy consumption parameters, production status parameters, quality index parameters and equipment safety parameters based on industrial Ethernet or wireless industrial protocols, and obtain the original collected data stream of multi-source physical quantities; Applying a time series synchronization algorithm to the original collected data stream, synchronously integrating the energy consumption parameter, the production status parameter, the quality index parameter, and the equipment safety parameter under a unified time reference, and generating a structured multi-source process unit synchronized data set; Based on the process unit layout information and sensor deployment plan, a unit space association algorithm is executed on the structured multi-source process unit synchronized data set to label all data samples according to physical process units to form a process unit data subset with spatial labels; Applying a production segment labeling algorithm to the process unit data subset, using real-time production scheduling information and process flow logs, attaching a production segment label to each data sample to generate a full set of process sampling parameters specific to the environment; The completed process category and the complete parameter set of the process sampling are input into the real-time data buffer module.

4. The method for optimizing energy consumption of a casting production line according to claim 3, characterized in that: The intelligent sensor node includes multiple types of sensors such as active power, temperature, state, vibration, pressure and output, with a collection accuracy of ±0.5%-10% and a sampling period of 1-20 seconds.

5. The method for optimizing energy consumption of a casting production line according to claim 1, characterized in that: The step S2 specifically includes: Based on statistical discrimination methods and physical constraint rules, we perform multi-source outlier detection on the collected energy consumption data, production status data, quality indicator data, and equipment safety parameters. We remove outlier data and sampling error data that exceed the normal discrimination threshold to obtain a high-confidence original data set for the process unit. Based on the original data set of the process unit, a unified clock reference and industrial synchronization protocol are used to align the sampling timestamps, and time trajectory resampling and interpolation are performed on multi-source data from different process units and different acquisition frequencies to form a time-synchronized data stream; Applying a normalization algorithm to the time-synchronized data stream, unifying the dimensions of each feature, mapping the numerical range to a standard interval, and outputting a normalized feature matrix; Performing data consistency check on the characteristic matrix to correct abnormal coupling relationships and residual deviations; Based on the feature matrix that passes the consistency check, a standard input preprocessed multi-process unit data set with a single acquisition time index is generated.

6. The method for optimizing energy consumption of a casting production line according to claim 5, characterized in that: The outlier detection uses the mean-standard deviation interval or box plot method, the outlier discrimination coefficient k is 2-3, the upper and lower bounds of the physical constraints are set according to actual production conditions, the normalization algorithm is Z-score normalization or maximum and minimum normalization, and the feature matrix normalization interval is [0,1].

7. The method for optimizing energy consumption of a casting production line according to claim 1, characterized in that: The step S3 specifically includes: After pre-processing the data, the process unit energy consumption parameters, production status parameters, quality index parameters, and equipment safety parameters are grouped and archived. Based on the production segment labels of each sampling point, an association mapping table is constructed for these parameters to generate a process unit-production segment-multi-parameter triple as the feature modeling input. Based on the triple feature modeling input, a multi-objective feature extraction algorithm is used to perform feature cross-correlation analysis on energy consumption parameters, production status parameters, quality index parameters, and equipment safety parameters, and to extract the multi-objective coupling feature vectors between the parameters. The multi-objective coupling eigenvector is modeled with parameter distribution using multi-objective optimization theory, a characteristic weight evaluation system for each process unit in a specified production section is constructed, and a four-dimensional parameter weight matrix of energy consumption, quality, production capacity, and safety under the process unit-production section is output; Based on the obtained parameter weight evaluation system, the feature normalization algorithm is used to normalize the energy consumption-quality-capacity-safety multi-objective weight matrix under the process unit; The normalized multi-objective weight matrix of each process unit-production section is aggregated and managed to generate a hierarchical multi-objective feature weight distribution matrix for the process unit.

8. The method for optimizing energy consumption of a casting production line according to claim 1, characterized in that: The step S4 specifically includes: Performing multi-objective parameter input initialization processing on the multi-objective feature weight distribution matrix, standardizing the feature weight distribution matrix format and mapping it to the data interface of each process unit intelligent body; Based on the hierarchical multi-agent architecture, the multi-objective feature weight distribution matrix corresponding to each process unit is input into the local process unit agent using a parameterized interface to establish the agent's own multi-objective optimization task instance; Relying on the multi-objective collaborative optimization algorithm built into each process unit intelligent agent, a multi-objective function is constructed based on the multi-objective feature weight distribution matrix and the energy consumption data, production status data, quality index data and equipment safety parameters collected by this unit; An intelligent multi-objective optimization algorithm is used within each process unit agent to iteratively search for the optimal parameter combination. The priority of each objective and the constraints are further combined to optimize the parameter solution, ultimately obtaining a candidate solution set for the optimal energy consumption parameters of the process unit. The set of candidate solutions for optimal energy consumption parameters output by each process unit agent is checked for validity and consistency, solutions that do not meet real-time constraints or violate safety boundaries are excluded, and the valid candidate optimization parameter vectors are output to the upper-level global scheduling agent using a standard protocol.

9. The method for optimizing energy consumption of a casting production line according to claim 8, characterized in that: The intelligent multi-objective optimization algorithm includes reinforcement learning, a multi-objective evolutionary algorithm or an optimization algorithm based on weighted aggregation. The multi-objectives include energy consumption, production capacity, quality and equipment safety, and support dynamic adjustment of weights and parameter inputs.

10. The method for optimizing energy consumption of a casting production line according to claim 1, characterized in that: The step S5 specifically includes: The candidate optimization parameter vectors generated by each process unit agent are synchronously packaged and transmitted based on the statutory process unit tags and timestamps to obtain a set of candidate optimization parameters for multiple process units. Based on the candidate optimization parameter sets of the multiple process units, the parameter vectors are normalized and aggregated using an inter-agent communication protocol to form a multi-objective parameter decision matrix; Relying on the global scheduling agent, taking the multi-objective parameter decision matrix as input conditions, and adopting an adaptive reward mechanism, the reward score is calculated based on the historical contribution and real-time performance of the candidate optimization parameter vector of each unit, and the adaptive reward vector of multiple process units is obtained; Based on the multi-process unit adaptive reward vector, a weight update algorithm in a reinforcement learning method is used to dynamically adjust the decision weight coefficient of each unit in the multi-objective parameter decision matrix, and output an adaptive decision weight distribution matrix; The adaptive decision weight distribution matrix is ​​used in combination with the initial multi-process unit candidate optimization parameter set to execute a multi-objective balance optimization algorithm of the global objective function to obtain a global optimal energy consumption parameter vector of the multi-objective weighted balance.

Images