A lead-zinc flotation process data set construction method and system for intelligent optimization

Abstract

The application relates to a lead-zinc flotation process data set construction method and system for intelligent optimization. The method includes the following steps: uniformly standardizing lead-zinc flotation production related data, accurately aligning offline assay grade data and high-frequency process data based on material residence time and total delay range, completing sample construction, data cleaning and repair, quality evaluation and version management in combination with working condition identification and process mechanism rules, forming a high-quality feature-label sample set with working condition metadata and a traceable data set version. Compared with existing data sets, the method can significantly improve the intelligent model training data quality for process index prediction and optimization control, improve the accuracy, stability and generalization ability of the model, enhance the explainability of production process fluctuation analysis and abnormal cause tracing, reduce the data arrangement, cleaning and maintenance cost, promote the standardization and reuse of lead-zinc flotation process data, and provide a reliable data basis for enterprises to continuously carry out process optimization and intelligent upgrading.

Description

Technical Field

[0001] This application relates to the field of lead-zinc flotation technology, and in particular to a method and system for constructing a lead-zinc flotation process dataset for intelligent optimization. Background Technology

[0002] With the development of intelligent lead-zinc flotation, data-driven methods (such as expert systems, neural networks, and digital twins) have become key to optimizing production. The effectiveness of these advanced applications heavily depends on the quality of input data. However, the construction of datasets in the industry is currently in a primitive state: time-series data is mechanically collected and stored only through distributed control systems (DCS), online sensors (such as pH meters and grade analyzers), and manual input. Outliers, missing values, and noise in the data are handled crudely, usually by simply filling in the mean or deleting them directly, which destroys the physical meaning and temporal correlation of the data. Data labels are usually only average test values ​​at the shift or hour level, which cannot be accurately aligned with high-frequency process parameters on the time scale, resulting in inaccurate "feature-label" pairing and making it difficult for the model to learn the true causal relationship. The contextual information of production conditions (such as changes in raw ore type, equipment start-up and shutdown, and reagent batch changes) is not recorded during data collection, resulting in a large amount of data with different "patterns" mixed in the dataset, and the model training objective is unclear.

[0003] Due to the aforementioned extensive data management methods, the constructed datasets suffer from the dilemma of "large data volume but low quality," resulting in inaccurate predictions by AI models trained on "dirty data" and an inability to adapt to changes in production conditions. Noise and lag in the data can mislead advanced process control algorithms, leading to incorrect decisions and even production fluctuations. A large amount of data becomes "data garbage" due to a lack of effective labeling and contextual information, which cannot support in-depth process analysis and optimization. Summary of the Invention

[0004] To address or partially address the problems existing in related technologies, this application provides a method and system for constructing a dataset for intelligent optimization of lead-zinc flotation processes. This method and system can achieve standardized processing of multi-source heterogeneous data, accurate alignment of test data and process data, improvement of sample quality, and version management of datasets, providing a high-quality, interpretable, and reusable data foundation for intelligent model training and optimized control.

[0005] The first aspect of this application provides a method for constructing a dataset for intelligent optimization of lead-zinc flotation processes, comprising the following steps: Acquire production-related data during the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data; Based on time-series process data and flotation process mechanism, the material residence time and total delay range are calculated. Within the delay range, offline test grade data and corresponding high-frequency process data are time-aligned to obtain the process time window and its feature vector corresponding to each test grade sample, and feature-label samples are formed. Feature extraction is performed on process variables and image features of time-series process data and feature-label samples. Based on the extraction results, working condition identification is performed to generate working condition labels. The working condition labels are then used as metadata and associated with the corresponding time-series data and feature-label samples to form a data package. Data quality checks are performed on data packets based on a pre-built process mechanism rule base, identifying missing and / or outlier values, and estimating and repairing repairable missing and / or outlier values ​​to obtain a feature-label sample set with operating condition metadata. The sample set is quality assessed to generate a dataset quality evaluation result. The sample set, along with its working condition screening criteria, data cleaning and repair strategies, and quality evaluation result, is registered as a dataset version for subsequent training and use by intelligent models.

[0006] The production-related data includes at least real-time process parameters from distributed control systems and / or programmable logic controllers, online grade and / or particle size data from online analyzers, foam images and their image feature data acquired by machine vision devices, and offline grade data of concentrates, tailings and middlings recorded by laboratory testing systems.

[0007] The calculation of material residence time and total delay range based on time-series process data and flotation process mechanism includes: The average residence time of the material in each flotation unit is estimated based on the effective volume and slurry flow rate of each unit, and the residence time of each unit is accumulated along the material flow direction to obtain the material residence time of the process section. By combining the slurry transport time in the pipeline with the detection delay of the online analyzer and the testing process, the total delay range of the testing sampling point relative to the upstream process measuring point is determined.

[0008] This includes time alignment of offline laboratory sample data with corresponding high-frequency process data within the latency range, which includes: Within the total delay range, delay search and / or dynamic time warping are performed on the time series of laboratory grades and the time series of related process variables to determine the best matching process time for each laboratory grade sample. A process time window of a preset length is constructed with the best matching process moment as the center, and statistics are calculated on the process variables and image features within the window to generate feature vectors.

[0009] The process of extracting features from time-series process data and feature-labeled samples, and then identifying the working conditions, includes: Based on the sliding time window, the mean, variance, rate of change of each process variable, as well as the statistical characteristics of equipment start-up and shutdown status, alarm frequency and foam image, are calculated to construct the operating condition feature vector. Cluster analysis and / or classification identification are performed on the operating condition feature vectors to distinguish between stable operating state, raw ore property fluctuation dynamics, reagent adjustment transition state, equipment abnormal state and / or maintenance shutdown state, and corresponding operating condition labels are generated accordingly.

[0010] The estimation and repair of repairable missing and / or outlier values ​​includes: Based on data that meet the process mechanism rules and are in a stable operating condition, process variables and / or image features that are correlated with the target variable are selected to construct training samples; The target variable is predicted and estimated using a random forest regression model and / or a gradient boosting tree model. The prediction results are used as repair values ​​for missing values ​​and soft outliers. When multiple target variables are repaired simultaneously, the repair results are adjusted in conjunction with material balance and / or metal balance constraints.

[0011] The quality assessment of the feature-label sample set with operating condition metadata includes: Completeness index is calculated based on the proportion of missing and outlier values; consistency index is calculated based on the deviation of material balance and / or metal balance; accuracy index is calculated based on the deviation between online analysis results and offline test results; and timeliness index is calculated based on data update delay and availability. Each indicator is weighted according to a preset weight to obtain a comprehensive score for dataset quality, which serves as part of the dataset quality evaluation results.

[0012] The second aspect of this application provides a data set construction system for lead-zinc flotation processes, applicable to the data set construction method for lead-zinc flotation processes provided in the first aspect of this application, including: The data acquisition and standardization module is used to acquire production-related data in the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data. The time alignment and sample construction module is used to calculate the material residence time and total delay range based on time-series process data and flotation process mechanism. Within the delay range, offline test grade data and corresponding high-frequency process data are time-aligned to obtain the process time window and its feature vector corresponding to each test grade sample, and feature-label samples are formed. The working condition identification and data packet construction module is used to extract features from process variables and image features of time-series process data and feature-label samples, perform working condition identification based on the extraction results, generate working condition labels, and associate the working condition labels as metadata with the corresponding time-series data and feature-label samples to form a data packet. The data cleaning and repair module is used to perform data quality checks on data packets based on a pre-built process mechanism rule base, identify missing values ​​and / or outliers, estimate and repair repairable missing values ​​and / or outliers, and obtain a feature-label sample set with operating condition metadata. The quality assessment and version management module is used to assess the quality of feature-label sample sets with working condition metadata, generate dataset quality evaluation results, and register the sample set, along with its working condition screening conditions, data cleaning and repair strategies, and quality evaluation results, as dataset versions for subsequent training and use by intelligent models.

[0013] A third aspect of this application provides a computer-readable storage medium storing a control program for a lead-zinc flotation process dataset construction system. When executed by a processor, the control program implements the steps of the lead-zinc flotation process dataset construction method provided in the first aspect of this application.

[0014] The technical solution provided in this application may include the following beneficial effects: This application provides a method and system for constructing a dataset for intelligent optimization of lead-zinc flotation processes. It processes lead-zinc flotation production-related data to form a high-quality feature-labeled sample set and a traceable dataset version with operating condition metadata. Compared to existing dataset construction methods, this method significantly improves the quality of training data for intelligent models used for process index prediction and optimization control, enhances the model's accuracy, stability, and generalization ability, strengthens the interpretability of production process fluctuation analysis and anomaly tracing, reduces data processing, cleaning, and maintenance costs, and promotes the standardized accumulation and reuse of lead-zinc flotation process data, providing a reliable data foundation for enterprises to continuously carry out process optimization and intelligent upgrading. Attached Figure Description The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0015] Figure 1 This is a schematic flowchart illustrating the method for constructing a dataset for lead-zinc flotation process as shown in an embodiment of this application. Detailed Implementation

[0016] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0017] A data set construction system for lead-zinc flotation processes includes: The data acquisition and standardization module is used to acquire production-related data in the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data. The time alignment and sample construction module is used to calculate the material residence time and total delay range based on time-series process data and flotation process mechanism. Within the delay range, offline test grade data and corresponding high-frequency process data are time-aligned to obtain the process time window and its feature vector corresponding to each test grade sample, and feature-label samples are formed. The working condition identification and data packet construction module is used to extract features from process variables and image features of time-series process data and feature-label samples, perform working condition identification based on the extraction results, generate working condition labels, and associate the working condition labels as metadata with the corresponding time-series data and feature-label samples to form a data packet. The data cleaning and repair module is used to perform data quality checks on data packets based on a pre-built process mechanism rule base, identify missing values ​​and / or outliers, estimate and repair repairable missing values ​​and / or outliers, and obtain a feature-label sample set with operating condition metadata. The quality assessment and version management module is used to assess the quality of feature-label sample sets with working condition metadata, generate dataset quality evaluation results, and register the sample set, along with its working condition screening conditions, data cleaning and repair strategies, and quality evaluation results, as dataset versions for subsequent training and use by intelligent models.

[0018] like Figure 1 The method for constructing a dataset for intelligent optimization of lead-zinc flotation processes, as shown, is applicable to the aforementioned system and includes the following steps: S1. Obtain production-related data during the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data.

[0019] In this embodiment, production-related data includes, but is not limited to, one or more of the following: process control data from DCS, PLC, SCADA, MES, or industrial historical databases; equipment data reflecting equipment operating status; offline laboratory test data; flotation foam image data or video frame data; event-type data such as operation records, alarm records, fault records, and shift change records; and supplementary process data entered manually.

[0020] The process control data includes slurry flow rate, liquid level, concentration, pH value, air volume, reagent dosage, valve opening, motor current, equipment speed, etc.; the offline test data includes raw ore grade, concentrate grade, tailings grade, recovery rate, etc.; the image data is preferably flotation cell foam images.

[0021] Since various types of data often differ significantly in terms of source system, sampling frequency, field naming, timestamp format, and unit dimensions, they need to be standardized first. Specifically, a variable mapping relationship can be established to unify fields representing the same process meaning in different systems into standard variable identifiers, and the physical meaning, unit, data type, process segment to which the variable belongs, and reasonable upper and lower limits can be uniformly defined.

[0022] The timestamps from various data sources are uniformly converted, and the data is resampled and time-series aligned according to a preset sampling period. For continuous variables, linear interpolation, previous value preservation, or sliding alignment can be used; for switch quantities, state quantities, and event quantities, the most recent state preservation or time-series mapping methods can be used. Issues such as inconsistent units, abnormal formats, duplicate records, null records, and illegal characters can also be normalized in this step.

[0023] After the above processing, multi-source heterogeneous time-series process data with a unified time base, unified variable identifiers, and a unified unit system are obtained. This time-series process data can include continuous process variables, as well as event information, shift information, process section identifiers, and image index information, thus providing a unified data foundation for subsequent label construction and sample generation.

[0024] S2. Calculate the material residence time and total delay range based on time-series process data and flotation process mechanism. Within the delay range, time-align the offline test grade data with the corresponding high-frequency process data to obtain the process time window and its feature vector corresponding to each test grade sample, and form feature-label samples.

[0025] In lead-zinc flotation production, offline laboratory test results typically correspond to the comprehensive results of the slurry after passing through multiple process units over a certain period, rather than the instantaneous process state at the moment of the test record. Therefore, there is usually a time delay between offline test grade data and high-frequency process data, caused by material transportation, process reactions, sampling and testing, and laboratory analysis. Directly correlating test results with process data at the same moment can easily lead to label misalignment, affecting sample accuracy and subsequent model training effectiveness.

[0026] To address this issue, this application calculates the material residence time based on the flotation process mechanism and determines the total delay range accordingly. Specifically, it first estimates the material residence time along the path from the location of the relevant process variables to the sampling location, taking into account the effective volume of each unit, slurry flow rate, and operating conditions. For series process units, the total material residence time can be expressed as: in, This represents the total material residence time. Let j be the effective volume of the j-th process unit. This is the volume utilization factor. Let be the slurry flow rate corresponding to the j-th process unit, and n be the number of process units traversed from the location where the process variable is applied to the location where the sample is taken.

[0027] In actual production, in addition to the residence time of materials in the process, offline test data may also be affected by sampling, sample delivery, sample preparation, laboratory analysis, and result entry. Therefore, sampling delay, testing delay, and data entry delay are further introduced on top of the material residence time to form a total delay range. The total delay can be expressed as: in, For sampling and delivery time, For laboratory analysis time, This refers to the time the result was entered.

[0028] Considering that production load, liquid level fluctuations, changes in ore properties, and changes in circuit status can cause some fluctuations in delay, delay tolerance can be further set based on empirical values, historical statistical results, or calibration results to obtain the total delay range for time alignment.

[0029] For each offline test sample, based on its recording time, the corresponding process data is extracted forward within the total delay range, and this time interval is used as the process time window corresponding to that test sample. Compared with the traditional single-moment value method, this application, by introducing a process time window, better conforms to the actual mechanism of the flotation process: "input action - process transfer - result display".

[0030] After obtaining the process time window, features are extracted from the data within the window to form feature vectors corresponding to the test labels. Features may include statistical features, trend features, fluctuation features, process combination features, and image features. For example, the mean, maximum, minimum, standard deviation, range of change, and upward or downward trend of process variables within the window can be extracted, as well as the coupling relationship features between reagents, liquid levels, airflow, and concentrations. For image data, foam images can be synchronized with the corresponding process time window according to timestamps, and visual features such as foam color, texture, size distribution, stability, or depth can be extracted.

[0031] Finally, the feature vectors extracted from the process time window are paired with the corresponding offline test labels to form feature-label samples. The labels can be lead concentrate grade, zinc concentrate grade, tailings grade, recovery rate, or a combination of labels consisting of multiple process indicators.

[0032] This step achieves mechanistic alignment between offline test results and high-frequency process data, significantly reducing sample label mismatch and improving the accuracy and usability of sample construction.

[0033] S3. Extract features from the process variables and image features of the time-series process data and feature-label samples. Based on the extraction results, identify the working conditions and generate working condition labels. Associate the working condition labels as metadata with the corresponding time-series data and feature-label samples to form a data package.

[0034] The lead-zinc flotation process is affected by various factors, including changes in ore properties, reagent formulation adjustments, equipment operating conditions, and on-site operational interventions, resulting in significantly different operating conditions at different times. Indiscriminately mixing samples from different operating conditions for subsequent modeling can easily lead to mixed model patterns, increased prediction bias, and decreased generalization ability. Therefore, this application, after constructing feature-label samples, further identifies the operating conditions and generates operating condition labels.

[0035] In this embodiment, the features used for operating condition identification can be derived from time-series process data and feature-label samples, including but not limited to: raw ore property features, slurry concentration features, reagent formulation features, liquid level and air volume features, equipment load features, foam image features, event record features, and combinations of the above features.

[0036] Operating condition identification can be achieved using rule-based identification, clustering identification, supervised classification identification, or a combination thereof. For example, obvious operating conditions can be identified based on process experience rules: when the raw ore grade fluctuates significantly and the reagent addition strategy is adjusted synchronously, it can be identified as a minerality fluctuation condition; when the liquid level, air volume, and foam state continuously deviate from the normal range, it can be identified as a foam abnormality condition; when there is a mismatch in current, flow rate, or rotational speed, it can be identified as an equipment disturbance condition. For implicit operating conditions that are difficult to cover by rules, samples can be grouped through clustering, and then process experts can interpret and name the clustering results. When manually labeled samples are available, supervised learning models can also be used to automatically classify operating conditions.

[0037] The identified operating condition labels may include, but are not limited to: normal steady-state operating conditions, ore property fluctuation operating conditions, reagent adjustment operating conditions, liquid level fluctuation operating conditions, foam abnormal operating conditions, equipment abnormal operating conditions, and start-up and shutdown transition operating conditions.

[0038] After generating work condition labels, these labels are used as metadata and associated with corresponding time-series data, feature-label samples, image indexes, shift information, operation records, and event information to form a data package. In this way, each sample not only contains input features and target labels but also information about its production scenario. This data package structure facilitates subsequent sample selection by work condition, work condition-specific modeling, work condition switching analysis, and model interpretation and traceability.

[0039] S4. Based on the pre-built process mechanism rule base, perform data quality checks on the data packets, identify missing values ​​and / or outliers, estimate and repair repairable missing values ​​and / or outliers, and obtain a feature-label sample set with operating condition metadata.

[0040] In lead-zinc flotation production, data acquisition is often affected by factors such as sensor malfunctions, communication interruptions, equipment interference, manual input errors, and system anomalies, resulting in missing values, jump values, stuck values, out-of-bounds values, and logically conflicting values ​​in the raw data. If these problems are not addressed, the constructed sample set will fail to accurately reflect the process mechanism, and will also affect the effectiveness and stability of subsequent model training.

[0041] Therefore, this application pre-constructs a process mechanism rule base and performs quality checks on the data packet generated in step S3 based on this rule base. The process mechanism rule base may include variable physical boundary rules, process logic rules, mass balance rules, metal quantity balance rules, rate of change rules, and operating condition consistency rules, etc.

[0042] Among them, physical boundary rules are used to limit the reasonable range of variables such as liquid level, flow rate, grade, concentration, drug dosage, and current; process logic rules are used to constrain the logical consistency between shutdown state, valve state, equipment operating state and process variables; change rate rules are used to identify abnormal jumps that do not conform to the actual process in a short period of time; and operating condition consistency rules are used to verify whether the operating condition label and the corresponding sample features match each other.

[0043] In a preferred embodiment, the rule base also includes metal quantity balance constraints based on process mechanisms to assist in identifying sample anomalies. This relationship can be expressed as: Where F is the flow rate of the raw ore. C represents the grade of the raw ore, and C represents the flow rate of the concentrate. T represents the concentrate grade, and T represents the tailings flow rate. This refers to the grade of the tailings.

[0044] In practical applications, the above relationship is used to check whether the sample meets the basic process balance. When the balance deviation exceeds the preset threshold, the corresponding sample can be judged as an abnormal sample or marked as a low confidence sample for subsequent repair or removal.

[0045] This application does not delete all identified missing and outlier values, but rather classifies them according to variable attributes, duration of absence, degree of anomaly, operating condition type, and their impact on label and sample integrity. For example, for continuous variables that are short-term missing or have local jumps, linear interpolation, smoothing of values ​​before and after, moving average, or estimation of similar samples under the same working conditions can be used for repair; for variables that are strongly correlated with other variables, the relationship between related variables can also be used for estimation and repair; for missing frames in image data, adjacent keyframe replacement, feature aggregation compensation, or missing marker methods can be used for processing. For anomalous data that is missing key labels, has been missing for extended periods, seriously violates process logic, or cannot be explained by the rule base, it is preferable to directly remove the corresponding samples. During the data repair process, repair trace information is recorded, including repaired fields, repair methods, repair time, and states before and after repair, for subsequent quality tracking and model analysis.

[0046] After the above quality checks, repairs, and removals, a feature-labeled sample set with operating condition metadata is obtained. Compared with the original sample set, this sample set is improved in terms of completeness, consistency, reliability, and modelability, making it more suitable for subsequent intelligent model training and actual industrial deployment.

[0047] S5. Conduct a quality assessment of the sample set, generate a dataset quality evaluation result, and register the sample set, along with its working condition screening conditions, data cleaning and repair strategies, and quality evaluation results, as a dataset version for subsequent training and use of intelligent models.

[0048] To ensure that the dataset is reusable, comparable, and traceable, this application further conducts quality assessment of the sample set after completing sample construction and quality governance, and establishes a dataset version management mechanism.

[0049] In this implementation, dataset quality assessment can be conducted from aspects such as completeness, consistency, anomalies, operational condition coverage, sample balance, and usability. Completeness primarily reflects the degree of missing samples and variables; consistency primarily reflects whether the samples meet process rules and mechanistic constraints; anomalies primarily reflect the proportion of abnormal samples and the remaining risks after correction; operational condition coverage primarily reflects whether the samples cover the main production scenarios required for the target task; sample balance primarily reflects whether the sample distribution across different operational conditions, label intervals, and time periods is balanced; and usability comprehensively reflects whether the dataset meets the requirements for subsequent model training, validation, and invocation.

[0050] To facilitate a unified quantification of dataset quality, in one embodiment, a comprehensive scoring method is used for evaluation, the expression of which is: Where Q is the overall quality score of the dataset. As a completeness indicator, As an anomaly rate indicator, As a consistency indicator, For operating condition coverage indicators, As an indicator of sample balance, to Each indicator has a corresponding weight, and the sum of all weights is 1.

[0051] In practical applications, the weights of each indicator can be adjusted according to different model tasks. For example, for process indicator prediction tasks, the weights of indicators related to completeness, consistency, and label credibility can be appropriately increased; for anomaly diagnosis tasks, the weights of anomaly sample coverage and operating condition coverage can be appropriately increased.

[0052] After completing the quality assessment, the obtained sample set, along with its operating condition screening criteria, time alignment parameters, feature extraction configuration, data cleaning and repair strategies, quality evaluation results, sample quantity, and operating condition distribution information, are registered as a dataset version. Each dataset version preferably corresponds to a unique version number for subsequent model training, deployment, and issue tracing.

[0053] In subsequent model development and application, the corresponding dataset version can be directly specified for training, validation, and testing, thereby ensuring consistency in data source and processing logic. When model performance changes or anomalies occur in the production environment, the data source, sample composition, operating condition range, and cleaning strategy can be quickly traced back based on the dataset version bound to the model, improving model management efficiency and the reliability of industrial applications.

[0054] The following uses a flotation production line of a lead-zinc ore beneficiation plant as an example to illustrate the method of this application.

[0055] First, production-related data for a continuous operating cycle of the production line is collected. Process control data comes from the industrial control system, laboratory test data from the laboratory management system, foam image data from industrial cameras installed above the flotation cells, and operation and alarm records from the production management system. This data undergoes time unification, variable mapping, unit conversion, and format standardization to form unified multi-source heterogeneous time-series process data.

[0056] Subsequently, combining the flotation process structure, effective volume of the flotation cell, slurry flow rate, and laboratory analysis process of the production line, the material residence time and total delay range corresponding to the offline test results were calculated. Based on this total delay range, the corresponding process time window was matched forward for each offline test record, and statistical features, trend features, process combination features, and foam image features were extracted from this window to construct a feature-label sample.

[0057] Subsequently, based on ore properties, reagent formulation, liquid level, air volume, foam status, equipment load, and alarm events, the operating conditions of the samples are identified, operating condition tags are generated, and the operating condition tags are associated with the samples and corresponding time segments to form data packets.

[0058] The missing values, outliers, and logically conflicting data in the data package are then identified using a process mechanism rule base. For short-term missing and locally anomalous variables, interpolation, smoothing, or estimation using samples from the same operating condition are used for repair. For data with severely distorted key variables, missing labels, or data that clearly violates process logic, the corresponding samples are removed. Furthermore, the sample set is comprehensively scored based on quality evaluation indicators, and this sample set, along with its time alignment parameters, operating condition selection criteria, data cleaning strategy, and evaluation results, is registered as a unique version of the dataset.

[0059] This dataset can be further used for training and application of models for lead concentrate grade prediction, zinc concentrate recovery rate prediction, reagent optimization control, and abnormal operating condition identification.

[0060] A large lead-zinc ore beneficiation plant implements the method of this application: Deployment: Deploy this data management suite on the central data platform to access approximately 2,000 data points across the entire process.

[0061] Implementation: The system automatically identifies and specially marks the 3-hour "transitional state" data during the raw ore switching period. When training the concentrate grade prediction model, engineers selectively exclude data from the "transitional state" and "equipment cleaning" periods, using only a high-quality subset of data from the "stable operating state".

[0062] Results: The neural network model trained on the dataset constructed using the new method showed a significant reduction in prediction accuracy (compared to offline testing), with the root mean square error decreasing from 0.85% to 0.58% compared to the model trained on the old dataset. Simultaneously, by analyzing the "transition state" dataset, process engineers successfully located a previously undetected problem of excessive reagent addition.

[0063] A computer-readable storage medium stores a control program for a lead-zinc flotation process dataset construction system. When the control program for the lead-zinc flotation process dataset construction system is executed by a processor, it implements the steps of the lead-zinc flotation process dataset construction method described above.

[0064] In the various embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0065] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0066] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a dataset for intelligent optimization of lead-zinc flotation processes, characterized in that, Includes the following steps: Acquire production-related data during the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data; Based on the time-series process data and flotation process mechanism, the material residence time and total delay range are calculated. Within the delay range, the offline test grade data and the corresponding high-frequency process data are time-aligned to obtain the process time window and its feature vector corresponding to each test grade sample, and feature-label samples are formed. Feature extraction is performed on the process variables and image features of the time-series process data and feature-label samples. Based on the extraction results, working condition identification is performed to generate working condition labels. The working condition labels are then used as metadata and associated with the corresponding time-series data and feature-label samples to form a data package. The data packets are subjected to data quality checks based on a pre-built process mechanism rule base, which identifies missing values ​​and / or outliers. Repairable missing values ​​and / or outliers are estimated and repaired to obtain a feature-label sample set with operating condition metadata. The sample set is quality assessed to form a dataset quality evaluation result. The sample set, along with its working condition screening conditions, data cleaning and repair strategies, and quality evaluation result, is registered as a dataset version for subsequent training and use of intelligent models.

2. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The production-related data includes at least real-time process parameters from the distributed control system and / or programmable logic controller, online grade and / or particle size data from the online analyzer, foam images and their image feature data acquired by the machine vision device, and offline test grade data of concentrate, tailings and middlings recorded by the laboratory testing system.

3. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The calculation of material residence time and total delay range based on time-series process data and flotation process mechanism includes: The average residence time of the material in each flotation unit is estimated based on the effective volume and slurry flow rate of each unit, and the residence time of each unit is accumulated along the material flow direction to obtain the material residence time of the process section. By combining the slurry transport time in the pipeline with the detection delay of the online analyzer and the testing process, the total delay range of the testing sampling point relative to the upstream process measuring point is determined.

4. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The time alignment of offline laboratory sample data with corresponding high-frequency process data within the delay range includes: Within the total delay range, delay search and / or dynamic time warping are performed on the time series of laboratory grades and the time series of related process variables to determine the best matching process time for each laboratory grade sample. A process time window of a preset length is constructed with the optimal matching process time as the center, and statistics are calculated on the process variables and image features within the window to generate the feature vector.

5. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The process of extracting features from time-series process data and feature-labeled samples, and then identifying operating conditions, includes: Based on the sliding time window, the mean, variance, rate of change of each process variable, as well as the statistical characteristics of equipment start-up and shutdown status, alarm frequency and foam image, are calculated to construct the operating condition feature vector. Cluster analysis and / or classification identification are performed on the operating condition feature vectors to distinguish between stable operating state, raw ore property fluctuation dynamics, reagent adjustment transition state, equipment abnormal state and / or maintenance shutdown state, and corresponding operating condition labels are generated accordingly.

6. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The estimation and repair of repairable missing values ​​and / or outliers includes: Based on data that meet the process mechanism rules and are in a stable operating condition, process variables and / or image features that are correlated with the target variable are selected to construct training samples; The target variable is predicted and estimated using a random forest regression model and / or a gradient boosting tree model. The prediction results are used as repair values ​​for missing values ​​and soft outliers. When multiple target variables are repaired simultaneously, the repair results are adjusted in conjunction with material balance and / or metal balance constraints.

7. The method for constructing a dataset for lead-zinc flotation process according to claim 1, characterized in that, The quality assessment of the feature-label sample set with operating condition metadata includes: Completeness index is calculated based on the proportion of missing and outlier values; consistency index is calculated based on the deviation of material balance and / or metal balance; accuracy index is calculated based on the deviation between online analysis results and offline test results; and timeliness index is calculated based on data update delay and availability. Each indicator is weighted according to a preset weight to obtain a comprehensive score for the dataset quality, which is used as part of the dataset quality evaluation result.

8. A data set construction system for lead-zinc flotation process, applicable to the data set construction method for lead-zinc flotation process as described in any one of claims 1-7, characterized in that, include: The data acquisition and standardization module is used to acquire production-related data in the lead-zinc flotation process, unify the time base and variable identifiers of the production-related data, and form multi-source heterogeneous time-series process data. The time alignment and sample construction module is used to calculate the material residence time and total delay range based on the time process data and flotation process mechanism. Within the delay range, the offline test grade data and the corresponding high-frequency process data are time aligned to obtain the process time window and its feature vector corresponding to each test grade sample, and feature-label samples are formed. The working condition identification and data packet construction module is used to extract features from the process variables and image features of the time-series process data and feature-label samples, perform working condition identification based on the extraction results, generate working condition labels, and associate the working condition labels as metadata with the corresponding time-series data and feature-label samples to form a data packet. The data cleaning and repair module is used to perform data quality checks on the data packets based on a pre-built process mechanism rule base, identify missing values ​​and / or outliers, estimate and repair repairable missing values ​​and / or outliers, and obtain a feature-label sample set with operating condition metadata. The quality assessment and version management module is used to assess the quality of the feature-label sample set with working condition metadata, generate a dataset quality evaluation result, and register the sample set, along with its working condition screening conditions, data cleaning and repair strategies, and quality evaluation result, as a dataset version for subsequent training and use of intelligent models.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a control program for a lead-zinc flotation process dataset construction system, which, when executed by a processor, implements the steps of the lead-zinc flotation process dataset construction method as described in any one of claims 1 to 7.

Images