A Smart Control Method and System for Jigging Process Parameters Based on Multi-Algorithm Fusion

By using a multi-algorithm fusion approach, support vector machines and temporal neural network models are used to identify the batch conditions of raw coal and generate continuous phase parameters. This solves the problem of discontinuous cross-phase feature transmission in intelligent jigging control and improves the stability and accuracy of the jigging process.

CN122828822BActive Publication Date: 2026-11-10TIANJIN DETONG ELECTRIC
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611281106.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-10
Estimated Expiration
2046-08-24

AI Technical Summary

Technical Problem

Existing intelligent jigging control methods fail to fully characterize the state transmission relationship between the valve opening phase and the valve closing phase, making it difficult for phase parameters to maintain continuous transmission of cross-phase characteristics under changing operating conditions, thus affecting the stability and fine control of the jigging process.

Method used

A multi-algorithm fusion approach is adopted, which uses a support vector machine condition classifier and a temporal neural network model to identify the batch conditions of raw coal and divide the phase feature sequence to generate continuous phase parameters. Combined with the jig controller, the model is executed and corrected to achieve stable transmission of cross-phase features.

Benefits of technology

It improves the continuity of phase parameters, reduces state interruptions and parameter abrupt changes caused by the transition between valve opening and closing phases, and enhances the operational stability, parameter adjustment accuracy, and product quality consistency of the jigging process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122828822B_ABST
    Figure CN122828822B_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent control method and system for jigging process parameters based on multi-algorithm fusion, relating to the field of jigging coal preparation technology. The method includes: real-time identification of the batch feature dataset of raw coal using a support vector machine (SVM) condition classifier, outputting batch condition categories; selecting continuous features corresponding to the raw coal batches according to the batch condition categories; dividing the continuous features into phase-sequence sequences using the opening and closing times of the air valves as the pulsation phase boundaries; performing time-series deduction on the phase-sequence sequences using a temporal neural network model to generate phase parameters for each pulsation phase; concatenating the phase parameters of each pulsation phase to form a composite pulsation waveform; converting the composite pulsation waveform into a jigging process parameter adjustment scheme and sending it to the jigging machine controller for execution, generating batch execution result data. This invention improves the continuity of phase parameters through pulsation phase boundary gating.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of jigging coal preparation technology, and in particular to an intelligent control method and system for jigging process parameters based on multi-algorithm fusion. Background Technology

[0002] The jigging coal preparation process is gradually shifting from fixed-value regulation relying on manual experience to intelligent control driven by multi-source sensing, condition identification, and parameter prediction. Monitoring methods such as online ash content detection, electronic belt weighing, valve feedback, and bed sensing can continuously characterize raw coal properties, production load, product quality, and fluctuating operating status. Feature engineering, support vector machines, and neural networks have been used for complex condition classification, temporal feature modeling, and process parameter optimization. Meanwhile, phase analysis focusing on the valve opening and closing cycle is gaining increasing attention. By distinguishing between the valve opening and closing phases, establishing the temporal relationship between bed movement, material stratification, and discharge regulation, and converting predicted parameters into control commands, this has become an important research direction for improving the automation and precision of the jigging process.

[0003] Existing intelligent jigging control methods mostly use fixed time windows or complete pulsation cycles as the unified analysis object, and usually directly output the parameter setpoints for the entire batch or cycle, failing to fully characterize the state transmission relationship between the valve opening phase and the valve closing phase. Since bed thickness, valve action, and material settling have significant phase dependencies, if adjacent pulsation phases are independently deduced, the bed change trend formed at the end of the previous phase is difficult to be continuously transmitted to the next phase, which can easily cause abrupt changes in phase parameters at boundary positions and weaken the temporal consistency between the parameter prediction results and the actual execution process of the jigging machine. Therefore, how to maintain the continuous transmission of cross-phase characteristics and generate stable phase parameters under changing operating conditions has become the main technical problem in intelligent control of jigging processes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent control method for jigging process parameters based on multi-algorithm fusion to solve the problem of maintaining continuous transmission of cross-phase characteristics and generating stable phase parameters under changing operating conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent control method for jigging process parameters based on multi-algorithm fusion, comprising, Acquire multi-source data of the jigging process, perform feature engineering on the multi-source data, divide the raw coal into batches, and generate a raw coal batch feature dataset; The raw coal batch feature dataset is identified in real time by a support vector machine working condition classifier, and the batch working condition category is output. The continuous features corresponding to the raw coal batch are selected according to the batch working condition category. The opening and closing times of the air valve are used as the pulsation phase boundary. The continuous features are divided according to time sequence to form a phase-sequence feature sequence. The phase parameters of each pulsation phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model. The phase parameters of each pulsation phase are spliced ​​together to form a composite pulsation waveform. The composite pulsation waveform is then converted into a jigging process parameter adjustment scheme and sent to the jigging machine controller for execution, generating batch execution result data. The product ash content, clean coal yield, and actual phase parameters are read from the batch execution result data. The actual phase parameters are compared with the phase parameters one by one, and the batch operating condition category is verified based on the product ash content and clean coal yield to form batch verification data. The batch verification data is used to correct the support vector machine operating condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

[0007] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, wherein: the generation of the raw coal batch feature dataset specifically involves, Abnormal data segments in the multi-source data are removed, and linear interpolation is used to fill the data gaps formed after the removal of abnormal data segments to obtain complete multi-source data. Smoothing filtering is applied to the complete multi-source data to form a preprocessed dataset. The preprocessed dataset is divided into fixed time windows. Statistical operations are performed on the preprocessed data within each time window to form statistical features. The preprocessed data within each time window are staggered sequentially along the sampling time. The covariance of the overlapping data segments before and after the staggering is calculated. The covariance calculation results are arranged according to the staggering order to form time-series features. The statistical features and time-series features are concatenated to form window features. Mark the locations of raw coal ash content changes and production load changes in chronological order along the window features, and delineate the raw coal batch boundaries based on the locations of raw coal ash content changes and production load changes. By segmenting the window features of raw coal batch boundaries, the window features between adjacent raw coal batch boundaries are classified into the same raw coal batch, generating a raw coal batch feature dataset.

[0008] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, the multi-source data includes raw coal ash content, product ash content, clean coal yield, production load, and real-time operating data collected from the jig's air valves and bed sensors.

[0009] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, the output batch operating condition category specifically includes: The raw coal batch feature dataset is spatially mapped by the radial basis kernel function of the support vector machine working condition classifier to form the raw coal batch mapped feature set. Based on the classification hyperplane in the support vector machine working condition classifier, the raw coal batch mapping features in the raw coal batch mapping feature set are divided, and the raw coal batch mapping features located in the same classification region are classified into the same working condition category, thus forming the raw coal batch working condition classification result. Based on the working condition category classification results recorded in the raw coal batch working condition classification results, the working condition category corresponding to each raw coal batch is marked as the batch working condition category.

[0010] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, wherein: the formation of the phase-sequence segmentation feature sequence specifically comprises, Based on the batch working condition category, continuous features corresponding to each raw coal batch are selected from the raw coal batch feature dataset, and the sampling order of continuous features within the raw coal batch is retained; Among the selected continuous features, the opening and closing times of the air valves within the same batch of raw coal are alternately arranged according to the sampling order to form an air valve edge sequence; According to the time sequence of the valve edge sequence, the time from the opening of the valve to the closing of the next valve is defined as the valve opening phase, and the time from the closing of the valve to the opening of the next valve is defined as the valve closing phase. The start and end times of the valve opening phase and the valve closing phase are combined to form a pulsating phase boundary pair. Using the start and end times of the pulsating phase boundary pairs as the intercept positions, continuous features corresponding to the valve opening phase and valve closing phase are intercepted respectively to obtain phase-segmented feature segments; Based on the time sequence of the pulsating phase boundary pairs, the phase segment corresponding to the valve opening phase is sequentially connected with the phase segment corresponding to the next valve closing phase to form a phase segment feature sequence.

[0011] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, wherein: the generation of phase parameters for each pulsation phase specifically comprises, The phase-series features are derived by time-by-time using a temporal neural network model according to the sampling order. Based on the pulsating phase boundary, the timing deduction results of the valve opening phase and valve closing phase are extracted from the phase timing characteristics to form phase-specific deduction characteristics; Starting with the first phase-separated feature, the phase-separated features with adjacent time are gated and connected at the pulsating phase boundary to form a cross-phase inference sequence. The phase inference results are generated by continuing the inference of the cross-phase inference sequence through a time-series neural network model. Using the end time of each pulsation phase as the index position, the deduction features located at the index position in the phase deduction results are used as phase characterization features; The output layer of the temporal neural network model performs parameter mapping on the phase representation features to generate phase parameters for each pulsating phase.

[0012] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, wherein: the formation of batch execution result data specifically includes, The phase parameters of each pulsation phase are arranged in chronological order to form a phase parameter sequence; Using the boundary position between adjacent pulsating phases in the phase parameter sequence as the reconstruction position, monotonically piecewise cubic Hermite interpolation is used to set continuous transition parameter values ​​between the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase, forming a boundary reconstruction transition segment; Based on the parameter range and adjustment rate allowed by the jig controller, continuous transition parameter values ​​above the upper limit of the parameter range are replaced with the upper limit value of the parameter range, and continuous transition parameter values ​​below the lower limit of the parameter range are replaced with the lower limit value of the parameter range. The rate of change of adjacent continuous transition parameter values ​​is limited to the allowed adjustment rate, forming an execution constraint transition segment. The execution constraint transition segment is embedded into the corresponding boundary position in the phase parameter sequence, and the phase parameters and execution constraint transition segments of each pulsating phase are concatenated along the time sequence of the pulsating phase to form a composite pulsating waveform; According to the start execution time of each pulsation phase in the composite pulsation waveform, the phase parameters and execution constraint transition segments of each pulsation phase are converted into parameter adjustment instructions that record the parameter name, target value, adjustment rate and start execution time. The parameter adjustment instructions are arranged according to the start execution time to form a jigging process parameter adjustment scheme. According to the start execution time in the jigging process parameter adjustment plan, the parameter adjustment instructions are sequentially sent to the jigging machine controller. The jigging machine controller adjusts the jigging machine operating parameters corresponding to the parameter names according to the target value and adjustment rate. At the junction of adjacent pulse phases, the actual phase parameter recorded at the end of the previous pulse phase is used as the adjustment starting point of the next pulse phase, forming a batch parameter execution record. The product ash content and clean coal yield of raw coal batches are retrieved from multi-source data. The product ash content, clean coal yield and batch parameter execution records are merged to form batch execution result data.

[0013] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, wherein: the formation of the intelligent control result for jigging process parameters specifically involves... Using the pulse phase and parameter name as the reference position, the actual phase parameter in the batch execution result data is compared with the phase parameter item by item. The reference position where the actual phase parameter and the phase parameter are inconsistent is marked as the parameter deviation position, thus forming a phase parameter comparison record. According to the raw coal batch, the product ash content and clean coal yield in the batch execution result data are included in the phase parameter comparison record to form the batch quality record; Based on the product ash content and clean coal yield in the batch quality record, retrieve the corresponding working condition category label from the historical labeled data sample, verify the working condition category label with the batch working condition category item by item, replace the inconsistent batch working condition category with the working condition category label, and form a batch working condition verification record. The batch quality records and batch operating condition verification records are merged according to the raw coal batches to form batch verification data; Based on whether the batch operating condition category has been replaced and whether the parameter deviation exists, the batch verification data that only has the batch operating condition category replaced is classified as category correction data, the batch verification data that only has the parameter deviation is classified as parameter correction data, and the batch verification data that has both the batch operating condition category replaced and the parameter deviation is classified as joint correction data. The category correction data, parameter correction data and joint correction data are combined to form the branched batch verification data. Using the parameter deviation position in the parameter correction data and joint correction data as the filtering condition, batch verification data with the same parameter deviation position are selected from the historical batch verification data. The selected batch verification data and the branch batch verification data are arranged according to the formation time to generate a correction sample sequence. The classification hyperplane of the support vector machine condition classifier is updated by combining the raw coal batch feature dataset corresponding to the classification correction data and joint correction data in the correction sample sequence with the verified batch condition category, thus obtaining the updated support vector machine condition classifier. The network weights of the temporal neural network model are updated by correcting the phase feature sequences corresponding to the parameter correction data and joint correction data in the correction sample sequence and the actual phase parameters. The updated support vector machine condition classifier and the updated temporal neural network model are used as the intelligent control results of jigging process parameters.

[0014] As a preferred embodiment of the intelligent control method for jigging process parameters based on multi-algorithm fusion described in this invention, the continuous transition parameter value is an intermediate parameter value set in chronological order between the end parameter value of the previous pulsation phase and the beginning parameter value of the next pulsation phase, based on monotonically piecewise cubic Hermite interpolation. The end parameter value of the previous pulsation phase is the last valid jigging process parameter value corresponding to the end time of the previous pulsation phase. The starting parameter value of the next pulsation phase is the first effective jigging process parameter value corresponding to the beginning moment of the next pulsation phase.

[0015] Secondly, this invention provides an intelligent control system for jigging process parameters based on multi-algorithm fusion, comprising, The acquisition module is used to acquire multi-source data of the jigging process, perform feature engineering on the multi-source data, divide the raw coal into batches, and generate a raw coal batch feature dataset. The deduction module is used to identify the working conditions of raw coal batch feature dataset in real time through a support vector machine working condition classifier, output the working condition category of the batch, select the continuous features corresponding to the raw coal batch according to the working condition category, and divide the continuous features according to the time sequence of the pulsation phase, using the opening and closing time of the air valve as the boundary of the pulsation phase, forming a phase-sequence feature sequence. The phase parameters of each pulsation phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model. The execution module is used to splice the phase parameters of each pulsating phase to form a composite pulsating waveform, convert the composite pulsating waveform into a jigging process parameter adjustment scheme and send it to the jigging machine controller for execution, and generate batch execution result data. The calibration module is used to read the product ash content, clean coal yield and actual phase parameters from the batch execution result data, compare the actual phase parameters with the phase parameters one by one, and verify the batch operating condition category according to the product ash content and clean coal yield to form batch calibration data. The batch calibration data is used to correct the support vector machine operating condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

[0016] The beneficial effects of this invention are as follows: By gating the connection at the pulsating phase boundary, the continuity of phase parameters is improved. The bed change trend formed at the end of the previous pulsating phase is continuously preserved, and the characteristics at the beginning of the next pulsating phase participate in continuous deduction according to the state update ratio, reducing state interruptions and parameter abrupt changes caused by the transition between the valve opening and closing phases; the phase parameters maintain smooth connection at the pulsating boundary position, the timing coordination relationship between jigging frequency, target bed thickness, and discharge speed is more stable, the continuity and executability of the composite pulsating waveform are improved simultaneously, control command jumps and frequent actuator corrections are suppressed, and the bed movement process is consistent with the parameter adjustment process, thereby improving the operational stability, parameter adjustment accuracy, and product quality consistency of the jigging sorting process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an intelligent control method for jigging process parameters based on multi-algorithm fusion.

[0019] Figure 2 This is a schematic diagram of an intelligent control system for jigging process parameters based on multi-algorithm fusion.

[0020] Figure 3 A flowchart for generating batch execution result data.

[0021] Figure 4 This is a flowchart for batch verification and model self-calibration.

[0022] Figure 5 This is a diagram illustrating the effect of pulse phase gating connection and composite pulse waveform reconstruction. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides an intelligent control method for jigging process parameters based on multi-algorithm fusion, comprising the following steps: S1. Obtain multi-source data of the jigging process, perform feature engineering on the multi-source data, divide the raw coal into batches, and generate a raw coal batch feature dataset.

[0027] S1.1 An online ash content analyzer is installed on the raw coal feeding conveyor belt to continuously acquire the ash content of the raw coal; the quality of the raw coal being washed is recorded and the production load is generated by an electronic belt scale on the feeding conveyor belt; the opening and closing times and operating frequencies of the air valves in the jig are collected from the feedback signals of the air valves; the bed thickness and bed fluctuation signals are collected by the bed sensor; and the raw coal ash content, production load, air valve operating parameters, and bed sensor values ​​are correlated according to a unified sampling time to form real-time control data; after the raw coal batches are sorted, the ash content of the product is acquired by an online ash content analyzer on the clean coal discharge conveyor belt. The clean coal yield is determined based on the ratio between the clean coal quality recorded by the electronic belt scale of the clean coal conveyor and the raw coal entering the washing process. The batch quality data is generated by associating the ash content of the product with the clean coal yield according to the batch identification of the raw coal. The real-time control data and the batch quality data are recorded separately according to the data formation stage to form multi-source data. The real-time control data participates in the raw coal batch division, batch working condition category identification and phase parameter deduction. The batch quality data does not participate in the formation of the window feature of the current raw coal batch. It only participates in the batch execution result data merging and batch working condition category verification after the raw coal batch has been sorted.

[0028] Among them, the ash content of raw coal is obtained by an online ash content detector on the raw coal feeding belt; the production load is obtained by an electronic belt scale on the feeding belt; the operating parameters of the air valve are obtained by the opening and closing feedback signal of the jig air valve; the values ​​collected by the bed sensor are obtained by the bed sensor; the ash content of the product is obtained by an online ash content detector on the clean coal discharge belt; the clean coal yield is determined according to the ratio between the quality of clean coal and the quality of raw coal fed into the washing machine in the same period; and the raw coal batch identification is formed according to the start time of the raw coal batch, the end time of the raw coal batch, and the material conveying time.

[0029] S1.2. Outlier data segments are removed from the multi-source data. Linear interpolation is used to fill the data gaps left after the outlier data segments are removed, resulting in complete multi-source data. Smoothing filtering is then applied to the complete multi-source data to form a preprocessed dataset. Specifically… The multi-source data are arranged according to the collected fields, and the raw coal ash content, product ash content, clean coal yield, production load, air valve operating parameters, and bed sensor collected values ​​are checked according to the sampling time. Sampling points with repeated sampling times, reversed sampling times, or sampling values ​​exceeding the upper and lower limits of the collected equipment are marked as invalid sampling points. Sampling points where the jump between adjacent sampling values ​​reaches the abnormal jump limit and subsequent sampling values ​​return to the fluctuation range before the jump are marked as abnormal jump points. Invalid sampling points and abnormal jump points with continuous time are merged into abnormal data segments, and abnormal data segments are removed from the time series of the corresponding collected fields.

[0030] Using the valid sampled values ​​at the front and back ends of the abnormal data segment as interpolation endpoints, filler values ​​are set according to the time ratio of the missing sampling time between the two interpolation endpoints. When there are valid sampled values ​​on both sides of the abnormal data segment and the duration of the data gap does not exceed the maximum fillable duration, linear interpolation is used to fill the data gap. When the abnormal data segment lacks any interpolation endpoint or the duration of the data gap exceeds the maximum fillable duration, the missing marker is retained and no filler value is set, forming complete multi-source data. Smoothing filtering is applied to the real-time control data in the complete multi-source data. Among them, the raw coal ash content, production load and bed sensor acquisition values ​​are smoothed using a central moving average, and the smoothing window duration is less than half of the shortest phase duration of the air valve. The sampled values ​​at the beginning and end of the sampling sequence are averaged using the valid sampled values ​​on one side. The air valve opening time and air valve closing time retain the original edge position and are not smoothed. Batch quality data retain the original detection results and formation time and are not involved in smoothing, forming a preprocessed dataset.

[0031] The abnormal jump limit is determined by adding three times the median absolute deviation to the median absolute difference between adjacent samples in the historical normal operation record; the upper and lower limits of the data acquisition equipment calibration are obtained by the calibration records of the raw coal ash content analyzer, electronic belt scale, air valve feedback component and bed sensor; the maximum fillable duration is determined by the continuous sample loss duration recorded in the data acquisition log and the continuity of the sampling sequence after fill; the shortest phase duration of the air valve is determined by the time difference between the opening time and the closing time of the air valve.

[0032] S1.3. Divide the preprocessed dataset into fixed time windows, perform statistical operations on the preprocessed data within each time window to form statistical features, sequentially stagger the preprocessed data within each time window along the sampling time, calculate the covariance of the overlapping data segments before and after the stagger, and arrange the covariance calculation results according to the stagger order to form time-series features. Concatenate the statistical features and time-series features to form window features. Specifically, The preprocessed dataset is arranged according to a uniform sampling time, and fixed time windows are continuously divided starting from the first valid sampling time. The length of the fixed time window is the shortest duration that can completely cover the opening of the primary air valve to the opening of the next air valve, and can accommodate no less than two valid sampling points for raw coal ash content and production load. The fixed time windows are connected end to end. When the remaining time at the end is less than one fixed time window, the remaining segment at the end is retained and does not participate in the formation of window features.

[0033] In this case, the fixed time window only divides the real-time control data in the preprocessed dataset, and the batch quality data is not written into the fixed time window and does not participate in the formation of statistical features, time series features and window features.

[0034] Collect fields In the The sampled values ​​arranged in the sampling order within a fixed time window are denoted as follows: Subscript comma comma The corresponding sampling time is denoted as Subscript comma The number of valid sampling points within a fixed time window is denoted as Subscript Fields to be collected In the The mean within a fixed time window is expressed by the following formula: ; Collection Fields In the The variance within a fixed time window is expressed by the following formula: ; Collection Fields In the The rate of change within a fixed time window is expressed by the following formula: ; In the formula, This represents the mean. Represents variance. To represent the rate of change, the maximum and minimum sampled values ​​within a fixed time window are recorded as the peak and valley values, respectively. The mean, variance, peak, valley values, and rate of change are arranged in the order of the collected fields to form statistical features.

[0035] Starting from the first sampled value within each fixed time window, the sampling sequence of the same acquisition field is staggered by one sampling point, two sampling points, and so on, up to the maximum stagger order. The maximum stagger order is one-quarter of the number of valid sampling points in the fixed time window, rounded down, so that the overlapping data segments before and after the stagger are retained at least three-quarters of the number of valid sampling points in the fixed time window.

[0036] Collection Fields In the The covariance within a fixed time window, staggered by the lth order, is expressed by the following formula: ; In the formula, This represents the covariance corresponding to the l-th order offset. This represents the mean of the overlapping data segments before they were staggered. represents the mean of the overlapping data segments after staggering, and l represents the staggering order.

[0037] Arrange the covariances corresponding to the same data collection field in ascending order of staggered order, and then connect all the covariances in the order of the data collection fields to form a time series feature.

[0038] The first-level arrangement is based on the order of the collected fields, and the second-level arrangement is based on the statistical feature name and staggered order. The statistical features are connected before the time-series features to form window features. When both raw coal ash content and production load have valid statistical features within a fixed time window, the window features are retained and the start and end times of the corresponding fixed time window are written. When either raw coal ash content or production load statistical feature is invalid, the fixed time window position is retained and an invalid window mark is set. The window features are used to identify the location of changes in raw coal ash content and production load along the time sequence.

[0039] The fixed time window length is determined by the opening and closing times of the air valve, the sampling time of raw coal ash content, and the sampling time of production load; the maximum staggered order is determined by the number of effective sampling points within the fixed time window.

[0040] S1.4. Mark the locations of raw coal ash content changes and production load changes along the time sequence of window features, and delineate the raw coal batch boundaries based on these locations. Segment the window features using the raw coal batch boundaries, grouping the window features between adjacent raw coal batch boundaries into the same raw coal batch, and generating a raw coal batch feature dataset. Specifically... The average raw coal ash content in adjacent window features is compared item by item in chronological order, and the absolute value of the difference between adjacent average raw coal ash content is recorded. The average production load in adjacent window features is compared item by item at the same time position, and the absolute value of the difference between adjacent average production load is recorded. When the absolute value of the difference between the average raw coal ash content reaches the raw coal ash content change threshold, the boundary position of two adjacent fixed time windows is marked as the candidate change position of raw coal ash content. If the raw coal ash content change threshold is not reached, the original time continuity is maintained. When the absolute value of the difference between the average production load reaches the production load change threshold, the corresponding boundary position is marked as the candidate change position of production load. If the production load change threshold is not reached, the original time continuity is maintained.

[0041] Read two consecutive valid window features from the candidate change positions of raw coal ash content and production load, and take the window feature before the candidate change position as the original level window feature. When both consecutive valid window features remain at the new level after the candidate change position, confirm the candidate change position of raw coal ash content as the raw coal ash content change position and the candidate change position of production load as the production load change position. When both consecutive valid window features return to the original level, cancel the corresponding candidate change position and retain the short-term fluctuation mark. When the change position of raw coal ash content and the change position of production load fall at the same fixed time window boundary, merge them into one raw coal batch boundary. When no complete window feature is formed between the two change positions, merge the two change positions and use the change position with earlier time as the raw coal batch boundary. When a complete window feature is formed between the two change positions, retain them as two raw coal batch boundaries respectively.

[0042] Using the start time of the first valid window feature and the end time of the last valid window feature as the outer boundary of the raw coal batch, all raw coal batch boundaries are arranged in chronological order. The first window feature after the boundary of a raw coal batch to the last window feature before the boundary of the next raw coal batch are grouped into the same raw coal batch. When the boundary of a raw coal batch falls into an invalid window, the raw coal batch boundary is moved to the starting position of the next valid window feature, while retaining the original changed position and the moved position. Each raw coal batch is assigned a raw coal batch number in chronological order, and the raw coal batch start time, raw coal batch end time, window feature, and anomaly marker are associated to generate a raw coal batch feature dataset. The raw coal batch feature dataset is used by the support vector machine working condition classifier for real-time working condition identification.

[0043] The threshold for raw coal ash content variation is determined by adding three times the median absolute deviation to the median absolute difference of the average ash content of raw coal in adjacent windows within a historically stable batch of raw coal. The threshold for production load variation is determined by adding three times the median absolute deviation to the median absolute difference of the average production load in adjacent windows within a historically stable batch of raw coal. An example value for the raw coal ash content variation threshold is 0.8 percentage points, and an example value for the production load variation threshold is 8% of the rated production load. Using this value can distinguish the continuous changes in the properties of raw coal and the amount of raw coal washed, and filter out fluctuations and short-term load disturbances.

[0044] S2. The raw coal batch feature dataset is identified in real time by a support vector machine working condition classifier, and the batch working condition category is output. The continuous features corresponding to the raw coal batch are selected according to the batch working condition category. The opening and closing times of the air valve are used as the pulsation phase boundary. The continuous features are divided according to the time sequence to form a phase-sequence feature sequence. The phase parameters of each pulsation phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model.

[0045] S2.1. Spatial mapping is performed on the raw coal batch feature dataset using the radial basis function of the support vector machine condition classifier to form a raw coal batch mapped feature set. Specifically, The window features in the raw coal batch feature dataset are mapped to the feature positions used in the training phase of the support vector machine (SVM) condition classifier according to their feature names and arrangement positions. For continuous numerical features, the mean and standard deviation of the SVM condition classifier training samples are used to unify the scale. The original values ​​of the valve status label and the anomaly label are maintained. Starting from the first valid window feature of each raw coal batch, the scaled window features are fed into the radial basis function (RBF) kernel function one by one. The RBF kernel function forms the kernel function value according to the feature distance between the window feature and the support vector. Then, the kernel function values ​​are organized according to the arrangement order of the support vectors in the SVM condition classifier. The raw coal batch number, the start and end times of the window, and the anomaly label are retained to form the raw coal batch mapping feature set. Window features with invalid window labels do not participate in spatial mapping, but their corresponding time positions are retained in the raw coal batch mapping feature set to prevent time misalignment in subsequent condition category assignments.

[0046] The mean and standard deviation of the training samples of the support vector machine (SVM) working condition classifier are formed by the effective feature values ​​corresponding to the same feature name in the historical labeled data samples.

[0047] S2.2. Based on the classification hyperplane in the support vector machine working condition classifier, the raw coal batch mapping features in the raw coal batch mapping feature set are divided. Raw coal batch mapping features located in the same classification region are classified into the same working condition category, forming the raw coal batch working condition classification result. Specifically, The raw coal batch mapping features are substituted into the classification hyperplane corresponding to each working condition category in the support vector machine working condition classifier. A classification score is formed according to the support vector weights, kernel function values, and bias terms of the classification hyperplane. The working condition category with the highest classification score is taken as the working condition category of the current raw coal batch mapping feature, and the working condition category, along with the raw coal batch number and the start and end times of the window, is written into the raw coal batch working condition classification result. When there are ties for the highest classification score, the support vector with the smallest distance to the current raw coal batch mapping feature is found from the support vectors corresponding to the ties, and the working condition category to which the support vector with the smallest distance belongs is written into the raw coal batch working condition classification result. When the minimum distances corresponding to the ties are still consistent, the working condition category corresponding to the previous effective raw coal batch mapping feature with the adjacent time is adopted, while retaining the classification boundary marker, to form the raw coal batch working condition classification result.

[0048] The training process of the Support Vector Machine (SVM) work condition classifier is as follows: Window features from historical labeled data samples are used as training features, and the work condition categories recorded in the historical labeled data samples are used as training labels. The training features are scaled uniformly as in S2.1. Kernel function values ​​are formed between training features using the radial basis function (RBF) kernel. The support vector weights, bias terms, penalty coefficients, and RBF kernel width are adjusted using hinge loss and classification margin constraints to ensure that training features for different work condition categories are located in the classification regions of their corresponding classification hyperplanes. Cross-validation is used to compare the classification accuracy corresponding to different penalty coefficients and RBF kernel widths. Training ends when the cross-validation accuracy in this round is no higher than the previous round and the optimal parameter combination remains consistent. The support vectors, support vector weights, bias terms, and classification hyperplanes are saved to form the SVM work condition classifier.

[0049] In this embodiment, historical labeled data samples are classified into operating condition categories according to the changes in raw coal ash content, production load, bed thickness fluctuation, and valve operation stability. Historical raw coal batches whose raw coal ash content and production load remain within the fluctuation range corresponding to historically stable raw coal batches, and whose bed thickness fluctuation amplitude and valve operation frequency remain stable, are marked as normal operating conditions. Historical raw coal batches whose raw coal ash content and production load deviate continuously from the fluctuation range corresponding to historically stable raw coal batches, but whose bed thickness fluctuation amplitude and valve operation frequency are still within the allowable operating range of the jig, are marked as transitional operating conditions. Historical raw coal batches whose bed thickness fluctuation amplitude, valve operation frequency, and production load are close to the allowable operating boundary of the jig are marked as boundary operating conditions.

[0050] Based on the formation time of raw coal batches, the first 80% of the historical labeled data samples were used as training samples, and the last 20% were used as validation samples. It was ensured that there were corresponding records for normal working conditions, transitional working conditions, and boundary working conditions in both training and validation samples. The support vector machine working condition classifier used five-fold cross-validation to compare the classification accuracy corresponding to different penalty coefficients and radial basis function widths. The penalty coefficients were selected from 0.1, 1, 10, 100, and 1000, and the radial basis function widths were selected from 0.001, 0.01, 0.1, 1, and 10, respectively. The above-mentioned cross-order-of-magnitude values ​​for the penalty coefficients and radial basis function widths were used to balance the classification interval constraint strength and feature space mapping scale within a parameter range covering multiple orders of magnitude, and to select parameter combinations that fit the distribution of historical labeled data samples through cross-validation.

[0051] The optimal parameter combination is the one with the highest average classification accuracy in five-fold cross-validation. When the average classification accuracy of multiple parameter combinations is consistent, the parameter combination with the smallest penalty coefficient is selected. When the penalty coefficients are still consistent, the parameter combination with the smallest radial basis function width is selected. Training ends when two consecutive rounds of cross-validation produce the same optimal parameter combination and the change in average classification accuracy in this round relative to the average classification accuracy in the previous round does not exceed one-thousandth.

[0052] The judgment ranges corresponding to normal operating conditions, transitional operating conditions, and boundary operating conditions are determined jointly by the historical raw coal batch feature dataset, the allowable operating parameter range of the jig, and the production personnel's review records; the historical labeled data samples are generated by the historical raw coal batch feature dataset and the operating condition categories recorded in the production log according to the raw coal batch number and production time.

[0053] S2.3. Based on the working condition category classification results recorded in the raw coal batch working condition classification, mark the working condition category corresponding to each raw coal batch as the batch working condition category, specifically as follows: From the batch working condition classification results of raw coal, all valid working condition categories corresponding to the same raw coal batch number are collected. The number of valid window features covered by each working condition category is recorded, and the working condition category with the most valid window features is marked as the batch working condition category of the corresponding raw coal batch. When the number of valid window features covered by two working condition categories is the same, the valid window features with later time in the same raw coal batch are read, and the working condition category corresponding to the valid window features with later time is marked as the batch working condition category. When the raw coal batch only contains classification boundary markers, the working condition category corresponding to the classification boundary markers is written into the batch working condition category and the classification boundary markers are retained to form a batch working condition category that corresponds to each item of the raw coal batch number.

[0054] S2.4. Based on the batch operating condition category, select the continuous features corresponding to each raw coal batch from the raw coal batch feature dataset, and retain the sampling order of the continuous features within the raw coal batch. Specifically, Using the batch working condition category as the selection position for continuous features, the names of continuous features reserved for the same working condition category during the training phase of the time-series neural network model are read. Corresponding continuous features are selected sequentially from the raw coal batch feature dataset, and the continuous features are arranged according to the sampling time within the raw coal batch. When the continuous feature name exists in the raw coal batch feature dataset, the correspondence between the continuous feature name, sampled value, sampling time, and anomaly marker is maintained. When the sampled value corresponding to the continuous feature name has a missing marker, the sampling position and missing marker are retained, and adjacent continuous features are not used for replacement. When the continuous feature name does not appear in the raw coal batch feature dataset, the missing marker of the continuous feature is retained, thus completing the selection of continuous features corresponding to each raw coal batch.

[0055] The continuous feature names corresponding to each working condition category are formed by masking continuous features one by one during the training phase of the time-series neural network model and comparing the prediction errors of phase parameters before and after masking. Continuous feature names whose prediction errors of phase parameters increase after masking are retained, while continuous feature names whose prediction errors of phase parameters do not increase after masking are not included in the time-series deduction of the corresponding working condition category.

[0056] S2.5. Among the selected continuous features, the opening and closing times of the air valves within the same batch of raw coal are alternately arranged according to the sampling order to form an air valve edge sequence, specifically, The feedback signals of the air valve status are expanded according to the sampling time sequence within the raw coal batch. The first sampling time when the air valve status switches from closed to open is recorded as the air valve opening time, and the first sampling time when the air valve status switches from open to closed is recorded as the air valve closing time. Starting from the first complete state switch position, all valid edges are arranged in the order of air valve opening time, air valve closing time, and the next air valve opening time. When two consecutive air valve opening times occur, the air valve opening time with the earlier time is retained, and the air valve opening time with the later time is marked as a repeated opening edge. When two consecutive air valve closing times occur, the air valve closing time with the earlier time is retained, and the air valve closing time with the later time is marked as a repeated closing edge. Air valve opening times that lack corresponding closing edges and air valve closing times that lack corresponding opening edges are marked as incomplete edges, forming an air valve edge sequence.

[0057] For example, when the states of the air valve are recorded sequentially as closed, closed, open, open, closed, only the sampling time when the closed state first switches to the open state is recorded as the air valve opening time, and the sampling time when the open state first switches to the closed state is recorded as the air valve closing time. The sampling positions that continuously maintain the same state do not form new edges.

[0058] S2.6. According to the time sequence of the valve edge sequence, the time from the valve opening moment to the next valve closing moment is designated as the valve opening phase, and the time from the valve closing moment to the next valve opening moment is designated as the valve closing phase. The start and end times of the valve opening phase and the valve closing phase form a pulsating phase boundary pair, specifically, Using the effective valve opening time in the valve edge sequence as the starting position of the valve opening phase, the first effective valve closing time is searched for. The valve opening time and the first effective valve closing time are recorded as the starting and ending times of the valve opening phase, respectively. Using the same valve closing time as the starting position of the valve closing phase, the first effective valve opening time is searched for. The valve closing time and the first effective valve opening time are recorded as the starting and ending times of the valve closing phase, respectively. When both the valve opening phase and the immediately following valve closing phase have complete start and end times, the two sets of start and end times are written into the same pulsating phase boundary pair according to the time sequence. When the starting position of the valve edge sequence is the valve closing time, the section from the starting position to the first valve opening time retains the incomplete valve closing phase mark. When the ending position of the valve edge sequence lacks the next edge, the end section retains the incomplete phase mark and is not included in the complete pulsating phase boundary pair, forming a pulsating phase boundary pair.

[0059] S2.7. Using the start and end times of the pulsating phase boundary pairs as the intercept positions, extract the continuous features corresponding to the valve opening phase and valve closing phase respectively to obtain phase-segmented feature segments, specifically, From the continuous features, find the continuous features whose sampling time falls between the start and end times of the valve opening phase. Include the continuous features corresponding to the start time of the valve opening phase into the valve opening phase, and transfer the continuous features corresponding to the end time of the valve opening phase into the immediately following valve closing phase. Assign the continuous features whose sampling time falls between the start and end times of the valve closing phase to the valve closing phase, and transfer the continuous features corresponding to the end time of the valve closing phase into the next valve opening phase. If there is no next valve opening phase for the last complete valve closing phase, retain the continuous features corresponding to the end time of the last complete valve closing phase in the last complete valve closing phase. If there are missing markers within a phase, maintain the original sampling position and missing markers, and do not change the start and end times of the phase, thus obtaining the phase-specific feature segments corresponding to the valve opening and valve closing phases.

[0060] S2.8. Based on the time sequence of the pulsating phase boundary pairs, the phase segment characteristic segments corresponding to the valve opening phase and the next valve closing phase are sequentially connected in series to form a phase segment characteristic sequence, specifically as follows: The valve-opening phase segment of the same pulsating phase boundary pair is placed first, followed by the valve-closing phase segment, and the two segments are connected according to the sampling time. The connection position retains only one boundary sampling position to prevent the same sampling time from repeatedly entering the phase segment feature sequence. When adjacent pulsating phase boundary pairs are continuous in time, the end position of the previous valve-closing phase and the start position of the next valve-opening phase are arranged continuously. When there are incomplete phase marks between adjacent pulsating phase boundary pairs, the connection between the two is interrupted at the position corresponding to the incomplete phase mark, and the arrangement is restarted from the first complete valve-opening phase. Each continuous arrangement segment is organized according to the raw coal batch number and time sequence, and the valve-opening phase mark, valve-closing phase mark and pulsating phase boundary are retained to form a phase segment feature sequence.

[0061] S2.9. The phase-series feature sequences are analyzed hourly according to the sampling order using a time-series neural network model to form phase-series features. Specifically, The first valid sampling position in the phase-sequence feature sequence is taken as the starting point for the derivation of the temporal neural network model. The continuous features at the same sampling position are arranged in the order of the continuous feature names determined in S2.4. The continuous features at the current sampling position are converted into hidden states through the sampling position derivation layer of the temporal neural network model, and the hidden states formed at the previous sampling position are passed to the current sampling position. This allows the hidden state formed at the current sampling position to simultaneously represent the current continuous features and the change relationship of the continuous features before the current sampling position.

[0062] The hidden state is updated item by item along the sampling time from the first valid sampling position. When the current sampling position has a missing label, the hidden state of the previous valid sampling position is maintained, the missing sampling value is not used to update the hidden state, and the missing label is appended to the corresponding hidden state position. When the previous valid sampling position does not exist, the initial hidden state determined during the training phase of the temporal neural network model is used as the previous hidden state of the current sampling position.

[0063] The hidden states formed at each sampling location are arranged sequentially according to the raw coal batch number, pulsation phase marker, and sampling time to form phase temporal features. The feature dimension of the phase temporal features is consistent with the hidden state dimension of the temporal neural network model, and is used to extract phase-by-phase deduction features at the pulsation phase boundary position.

[0064] S2.10. Based on the pulsating phase boundary, extract the timing deduction results of the valve opening phase and valve closing phase from the phase timing characteristics to form phase-specific deduction characteristics, specifically, The start and end times of the valve opening phase, the start and end times of the valve closing phase, and the end times of the valve closing phase are recorded in the phase timing features. Phase timing features between the start and end times of the valve opening phase are assigned to the valve opening phase, and phase timing features between the start and end times of the valve closing phase are assigned to the valve closing phase. The boundary sampling position is retained once according to the assignment method adopted in S2.7 to avoid the same phase timing feature from entering two adjacent pulsating phases at the same time. When the sampling time of the phase timing feature is consistent with the sampling time of the pulsating phase boundary, it is directly truncated. When the phase timing feature does not fall at the end time of the pulsating phase, the nearest valid phase timing feature before the end time of the pulsating phase is used as the end feature to form the phase sub-phase deduction features corresponding to the valve opening phase and the valve closing phase.

[0065] S2.11. Taking the first phase-separated feature as the starting point, gating and connecting temporally adjacent phase-separated features at the pulsating phase boundary to form a cross-phase deduction sequence, specifically, The first effective hidden state of the first phase-deduction feature is used as the initial hidden state. The phase-deduction feature is advanced along the time sequence to the end of the first pulsating phase. The hidden state at the end of the previous phase-deduction feature and the hidden state at the beginning of the next phase-deduction feature are fed into the gating structure of the temporal neural network model. The gating structure determines the retention coefficient of the previous phase state based on the change relationship between the two hidden states, and determines the update coefficient of the next phase state based on the retention coefficient of the previous phase state.

[0066] Both the previous phase state retention coefficient and the subsequent phase state update coefficient are limited to between zero and one, and the sum of the previous phase state retention coefficient and the subsequent phase state update coefficient is equal to one. The previous phase state retention coefficient controls the retention ratio of the hidden state at the end of the previous pulse phase, and the subsequent phase state update coefficient controls the addition ratio of the hidden state at the beginning of the subsequent pulse phase.

[0067] The hidden state components in the terminal hidden state of the previous phase-deduction feature are weighted according to the previous phase state retention coefficient, and the corresponding hidden state components in the head hidden state of the next phase-deduction feature are weighted according to the subsequent phase state update coefficient. The two weighted results corresponding to the same hidden state position are merged one by one to form the boundary connection hidden state, and the original head hidden state of the next phase-deduction feature is replaced by the boundary connection hidden state.

[0068] Starting from the second sampling position of the features derived from the next phase, the updated hidden states are arranged, and the gating connection is repeated at the boundary positions of subsequent pulsating phases. When the current pulsating phase has a missing marker, the retention coefficient of the previous phase state is increased and the update coefficient of the subsequent phase state is decreased. After the current pulsating phase recovers valid sampling, the retention coefficient of the previous phase state and the update coefficient of the subsequent phase state are re-determined based on the recovered valid hidden states. The hidden states that have completed the gating connection are arranged in the pulsating phase time order to form a cross-phase derivation sequence.

[0069] For example, when the valve opening phase transitions to the valve closing phase, the gating structure retains the bed thickness change trend at the end of the valve opening phase according to the previous phase state retention coefficient, and adds the valve closing state and bed settlement change at the beginning of the valve closing phase according to the subsequent phase state update coefficient, so that the valve closing phase can inherit the state change formed by the valve opening phase.

[0070] S2.12. The cross-phase extrapolation sequence is continuously extrapolated using a time-series neural network model to form phase extrapolation results. Using the end time of each pulsating phase as the index position, the extrapolation features located at the index positions in the phase extrapolation results are used as phase representation features. Specifically, The boundary connection hidden state in the cross-phase extrapolation sequence is used as the state transfer position between adjacent pulsating phases. Through the cross-phase continuation extrapolation layer of the time-series neural network model, the hidden state is updated sequentially from the first complete pulsating phase to the last complete pulsating phase. This allows the extrapolation process of the next pulsating phase to inherit the bed changes and valve action effects retained at the end of the previous pulsating phase, thus forming the phase extrapolation result.

[0071] The cross-phase continuation deduction layer processes the hidden states formed by S2.9 and the boundary connection hidden states formed by S2.11. The cross-phase continuation deduction layer does not process the original continuous features in the phase feature sequence again. The dimension of the hidden states formed by the sampling position deduction layer, the dimension of the boundary connection hidden states and the dimension of the hidden states used by the cross-phase continuation deduction layer are consistent, so that the phase time series features can be directly inherited to the cross-phase deduction sequence.

[0072] The phase deduction results are found for the end time of each recorded pulsating phase using the pulsating phase boundary. When a corresponding deduction feature exists at the end time, the corresponding deduction feature is used as the phase representation feature. When there is no completely consistent sampling position at the end time, the nearest valid deduction feature before the end time is used as the phase representation feature. When there is no valid deduction feature inside the pulsating phase, the boundary connection hidden state formed by the start position of the pulsating phase is used as the phase representation feature, and the phase missing mark is retained to form a phase representation feature that corresponds to each pulsating phase item by item.

[0073] S2.13. The phase representation features are parameter-mapped through the output layer of the temporal neural network model to generate the phase parameters for each pulsating phase. Specifically, The phase representation features of each pulsation phase to be executed are sent to the output layer of the temporal neural network model in pulsation phase time order. The output layer sets independent parameter output positions according to the pulsation phase type. The parameter output positions corresponding to the valve opening phase include valve opening degree, valve opening duration, air inlet pressure, jigging frequency and target bed thickness. The parameter output positions corresponding to the valve closing phase include valve closing duration, exhaust duration, discharge speed, jigging frequency and target bed thickness.

[0074] The phase characteristics of the valve opening phase are converted into parameter values ​​corresponding to the valve opening degree, valve opening duration, air intake pressure, jigging frequency, and target bed thickness by using output layer weights and bias terms. This allows the phase parameters of the valve opening phase to define the air intake process and the bed lifting process. Similarly, the phase characteristics of the valve closing phase are converted into parameter values ​​corresponding to the valve closing duration, exhaust duration, discharge rate, jigging frequency, and target bed thickness by using output layer weights and bias terms. This allows the phase parameters of the valve closing phase to define the bed settling process and the discharge process.

[0075] Each parameter value is associated with the raw coal batch number, pulsation phase number, pulsation phase type, parameter name, predicted formation time, start execution time, and pulsation phase end time. When the parameter value is within the allowable range of the corresponding jig operating parameters, the output value is retained. When the parameter value exceeds the allowable range of the corresponding jig operating parameters, the original output value and the out-of-bounds flag are retained. Out-of-bounds parameter values ​​are subject to execution constraint processing in S3 to generate the phase parameters for each pulsation phase.

[0076] The phase parameters of each pulsation phase are arranged according to the raw coal batch number and the pulsation phase time sequence. The parameter names and parameter values ​​corresponding to the same pulsation phase are grouped into the same phase parameter record, and the prediction formation time, start execution time, pulsation phase end time and boundary crossing mark are retained to form a sub-phase prediction parameter set. The sub-phase prediction parameter set is used to form a phase parameter sequence according to the pulsation phase time sequence and is used to compare it item by item with the actual phase parameters in the batch execution result data.

[0077] The training process of the temporal neural network model is as follows: the historical phase feature sequence is used as the training feature, and the valve opening degree, valve opening duration, air intake pressure, jigging frequency and target bed thickness corresponding to the valve opening phase in the historical batch parameter execution record, as well as the valve closing duration, exhaust duration, discharge speed, jigging frequency and target bed thickness corresponding to the valve closing phase, are used as the actual phase parameters. The actual phase parameters are used as the training target, and the training samples are constructed according to the pulsating phase time sequence.

[0078] The hidden state is updated along the sampling time using a temporal neural network model, and the pre-phase state retention coefficient and post-phase state update coefficient are trained at the boundary of adjacent pulsating phases. The mean square error between the phase parameters output by the temporal neural network model and the actual phase parameters corresponding to the same parameter names is used as the parameter prediction loss. At the same time, the abrupt change amplitude of the hidden state on both sides of the boundary of adjacent pulsating phases is included in the training loss. The hidden state weights, gating weights, output layer weights and bias terms of the temporal neural network model are updated by backpropagation along the time direction.

[0079] In this embodiment, the temporal neural network model adopts a two-layer gated loop structure, the hidden state dimension is set to 64, the training batch contains 32 training samples, the initial learning rate is set to 0.1%, and the maximum number of training rounds is set to 200. After each round of training, the parameter prediction loss of the validation samples is calculated. Training ends when the parameter prediction loss of ten consecutive rounds of training is not lower than the previous lowest parameter prediction loss, and the network weights corresponding to the round with the lowest parameter prediction loss are retained to form a temporal neural network model.

[0080] During the training of the temporal neural network model, continuous features are masked one by one, and the parameter prediction loss of the validation samples before and after masking is compared. If the parameter prediction loss increases after masking the corresponding continuous features, the name of the corresponding continuous feature is retained. If the parameter prediction loss does not increase after masking the corresponding continuous features, the name of the corresponding continuous feature is not included in the continuous feature names adopted in the same working condition category, thus forming the continuous feature names corresponding to each working condition category adopted in S2.4.

[0081] Among them, the historical phase feature sequence is generated by dividing the historical raw coal batch feature dataset according to the historical air valve opening time and the historical air valve closing time; the historical batch parameter execution record is formed by the parameter name, actual parameter value, actual start time and actual end time saved by the jig controller.

[0082] S3. The phase parameters of each pulsation phase are spliced ​​together to form a composite pulsation waveform. The composite pulsation waveform is converted into a jigging process parameter adjustment scheme and sent to the jigging machine controller for execution, forming batch execution result data.

[0083] S3.1 Arrange the phase parameters of each pulsation phase according to the time sequence of each pulsation phase to form a phase parameter sequence, specifically as follows: Phase parameters for each pulsating phase are retrieved from the phase prediction parameter set according to the raw coal batch number. The valve opening phase and valve closing phase are arranged in chronological order, with the start time of the pulsating phase as the arrangement position. The phase parameters in the same pulsating phase are set according to their parameter names to maintain the name correspondence between different pulsating phases.

[0084] For a single phase parameter value formed within the same pulse phase, the corresponding phase parameter value is used as the starting parameter value of the corresponding parameter name at the start of the pulse phase execution time, and the corresponding phase parameter value is maintained during the execution period of the corresponding pulse phase. The last valid parameter value before the end of the pulse phase is determined as the ending parameter value. The ending parameter value of the previous pulse phase refers to the last valid parameter value formed before the end of the previous pulse phase. The starting parameter value of the next pulse phase refers to the first valid parameter value used at the start of the execution time of the next pulse phase.

[0085] When adjacent pulsation phases have the same parameter name, the end parameter value of the previous pulsation phase and the start parameter value of the next pulsation phase are recorded sequentially according to the boundary time of the adjacent pulsation phases; when the next pulsation phase lacks a corresponding parameter name, the parameter missing mark is retained, and the parameter value corresponding to other parameter names is not used to fill the missing position.

[0086] The raw coal batch number, pulsation phase number, pulsation phase type, pulsation phase start and end time, parameter name, starting parameter value, ending parameter value, and anomaly mark are arranged in chronological order to form a phase parameter sequence. The phase parameter sequence is used to locate the parameter boundary position of adjacent pulsation phases and to provide the ending parameter value of the previous pulsation phase and the starting parameter value of the next pulsation phase for the boundary reconstruction transition section.

[0087] S3.2. Using the boundary position between adjacent pulsating phases in the phase parameter sequence as the reconstruction position, monotonically piecewise cubic Hermitian interpolation is used to set continuous transition parameter values ​​between the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase, forming a boundary reconstruction transition segment; wherein, the continuous transition parameter values ​​are intermediate parameter values ​​set in chronological order between the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase based on monotonically piecewise cubic Hermitian interpolation, specifically, Using the last valid parameter value of the same parameter name in the previous pulse phase and the first valid parameter value in the next pulse phase as interpolation endpoints, the time interval between the sampling times corresponding to the two interpolation endpoints is taken as the boundary reconstruction interval. When the parameter value at the end of the previous pulse phase is consistent with the parameter value at the beginning of the next pulse phase, the same parameter value is maintained within the boundary reconstruction interval. When the parameter value at the end of the previous pulse phase is inconsistent with the parameter value at the beginning of the next pulse phase, continuous transition parameter values ​​are set along the boundary reconstruction interval.

[0088] The normalized position of the sampling time within the boundary reconstruction interval is denoted as... The continuous transition parameter value is expressed by the following formula: ; In the formula, Indicates the sampling time The corresponding continuous transition parameter values, Indicates the sampling time within the boundary reconstruction interval. Indicates the interpolation parameter number. Indicates the sampling time The normalized position within the boundary reconstruction interval Indicates the relationship with the first The standard cubic basis functions corresponding to each interpolation parameter term This indicates the interpolation parameter items arranged according to their serial numbers. The interpolation parameter items correspond in sequence to the parameter value at the end of the previous pulsation phase, the parameter change caused by the slope of the previous interpolation endpoint within the boundary reconstruction interval, the parameter value at the beginning of the next pulsation phase, and the parameter change caused by the slope of the next interpolation endpoint within the boundary reconstruction interval.

[0089] When the slope of the previous parameter interval is in the same direction as the slope of the boundary reconstruction interval, the harmonic mean of the slopes of two adjacent segments is used to determine the slope of the previous interpolation endpoint; when the two slopes are not in the same direction, the slope of the previous interpolation endpoint is set to zero; when either slope is zero, the slope of the previous interpolation endpoint is set to zero. When the slope of the boundary reconstruction interval is in the same direction as the slope of the next parameter interval, the harmonic mean of the slopes of two adjacent segments is used to determine the slope of the next interpolation endpoint; when the two slopes are not in the same direction, the slope of the next interpolation endpoint is set to zero. Set to zero; when any change slope is zero, set the change slope of the next interpolation endpoint to zero; when the boundary position lacks the previous parameter interval, set the change slope of the previous interpolation endpoint to the change slope of the boundary reconstruction interval; when the boundary position lacks the next parameter interval, set the change slope of the next interpolation endpoint to the change slope of the boundary reconstruction interval; when any interpolation endpoint of adjacent pulsation phases has a parameter missing mark, do not perform monotonically piecewise cubic Hermite interpolation, and retain the boundary reconstruction missing mark; arrange the continuous transition parameter values ​​formed by the effective boundary positions according to the sampling time to form the boundary reconstruction transition segment.

[0090] S3.3. Based on the allowable parameter range and adjustment rate of the jig controller, continuous transition parameter values ​​above the upper limit of the parameter range are replaced with the upper limit value, and continuous transition parameter values ​​below the lower limit of the parameter range are replaced with the lower limit value. The rate of change of adjacent continuous transition parameter values ​​is limited to the allowable adjustment rate, forming an execution constraint transition segment. Specifically, From the jig controller parameter configuration record, read the lower limit of the parameter range, the upper limit of the parameter range, and the allowable adjustment rate according to the parameter name. Match the parameter name with the continuous transition parameter value in the boundary reconstruction transition section one by one. When the continuous transition parameter value is between the lower limit and the upper limit of the parameter range, retain the continuous transition parameter value. When the continuous transition parameter value is higher than the upper limit of the parameter range, limit the continuous transition parameter value to the upper limit of the parameter range. When the continuous transition parameter value is lower than the lower limit of the parameter range, limit the continuous transition parameter value to the lower limit of the parameter range. When no parameter configuration record with the same name is found, retain the position of the corresponding continuous transition parameter value and add a parameter constraint missing mark.

[0091] Arrange the continuous transition parameter values ​​that have completed the parameter range limitation according to the sampling time. Starting from the first valid continuous transition parameter value, verify the parameter change status between adjacent sampling times. If the change rate corresponding to the adjacent continuous transition parameter value does not exceed the allowable adjustment rate, retain the current continuous transition parameter value. If the change rate corresponding to the adjacent continuous transition parameter value exceeds the allowable adjustment rate, starting from the previous valid continuous transition parameter value, limit the continuous transition parameter value of the current sampling time to the parameter value boundary that the allowable adjustment rate can reach along the direction where the current continuous transition parameter value is located. If adjacent sampling times are repeated or the time order is reversed, retain the previous valid continuous transition parameter value and add a time order anomaly mark. Arrange the continuous transition parameter values ​​that have completed the parameter range limitation and adjustment rate limitation, along with the parameter name, sampling time, and anomaly mark, to form the execution constraint transition segment.

[0092] The allowable parameter range and adjustment rate of the jig controller are obtained from the jig controller parameter configuration table and actuator calibration record.

[0093] S3.4. Embed the execution constraint transition segment into the corresponding boundary position in the phase parameter sequence, and concatenate the phase parameters and execution constraint transition segments of each pulsating phase along the time sequence of the pulsating phase to form a composite pulsating waveform, specifically, Using the raw coal batch identifier, parameter name, previous pulse phase identifier, next pulse phase identifier, and boundary time recorded in the execution constraint transition section as positioning fields, the boundary position with the same fields is searched in the phase parameter sequence. When the positioning fields form a unique correspondence, the original parameter jump connection of the boundary position is replaced with the execution constraint transition section, and the beginning of the execution constraint transition section is connected to the end of the previous pulse phase, and the end of the execution constraint transition section is connected to the beginning of the next pulse phase. When no corresponding boundary position is found or a duplicate boundary position is found, the original boundary position is retained, and an abnormal transition section embedding mark is added.

[0094] Following the time sequence of the pulsating phases recorded in the phase parameter sequence, the arrangement order of the phase parameters within each pulsating phase is maintained, and the execution constraint transition segment that has completed positioning is placed between the corresponding adjacent pulsating phases; when the endpoint of the execution constraint transition segment is consistent with the boundary parameter value in the phase parameter sequence, the duplicate boundary parameter nodes are merged; when the endpoint of the execution constraint transition segment is inconsistent with the boundary parameter value, the endpoint of the execution constraint transition segment is retained, and a parameter connection anomaly mark is added at the boundary position; the phase parameters, execution constraint transition segments, and anomaly marks of each pulsating phase are arranged continuously according to the sampling time to form a composite pulsating waveform.

[0095] like Figure 5 As shown, based on the valve edge sequence formed by the alternating opening and closing times of the valves, the continuous operation process is divided into sequentially connected valve opening and closing phases, and the pulsating phase boundary is determined by the start and end times of each phase. For each pulsating phase, the time-series neural network model forms corresponding sub-phase deduction features, and reads the end hidden state of the previous sub-phase deduction feature and the beginning hidden state of the next sub-phase deduction feature at the boundary of adjacent pulsating phases. The retention ratio of the previous phase state and the update ratio of the next phase state are determined by the pulsating phase boundary gating connection, forming a cross-phase continuous feature. On this basis, a phase parameter sequence is formed according to the time sequence of the pulsating phases, and a boundary reconstruction transition segment is set at the boundary of adjacent phase parameters based on the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase. Then, combined with the parameter range and adjustment rate allowed by the jig controller, the boundary reconstruction transition segment is subjected to execution constraint processing to form an execution constraint transition segment, thereby obtaining a composite pulsating waveform.

[0096] To further explain, Figure 5 The lower composite pulsating waveform reconstruction effect comparison area uses execution time as the horizontal variation benchmark and the parameter target value as the vertical variation benchmark, and is divided into sections from t0 to t... n The time sequence is set with multiple continuous pulsating phases; t1, t2, t3 and t4 correspond to the pulsating phase boundaries between adjacent valve opening and closing phases. The vertical dashed lines only indicate the time boundary positions of adjacent pulsating phases and do not indicate the parameter transition intervals. The black phase parameter curves corresponding to the traditional splicing method extend continuously within each pulsating phase according to the current phase parameter change trend. After reaching the phase boundaries corresponding to t1, t2, t3 and t4, the parameter value at the end of the previous pulsating phase is directly switched to the parameter value at the beginning of the next pulsating phase. Therefore, obvious vertical parameter jumps are formed at the phase boundary positions, indicating that there is a lack of continuous parameter transition process between adjacent pulsating phases.

[0097] The blue phase parameter curves corresponding to the present invention are connected along the same pulsating phase boundary. The boundary reconstruction transition section is not the pulsating phase boundary itself, but a continuous parameter change interval formed near the boundary of adjacent phase parameters, using the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase as connection endpoints. In the figure, the blue marked area near t1 indicates the boundary reconstruction transition section. The original directly jumpable parameter connection is reconstructed into a continuous change trajectory within this interval, thereby eliminating the abrupt jump between the phase parameters. The execution constraint transition section is based on the boundary reconstruction transition section, further based on the parameter range and adjustment speed allowed by the jig controller. The executable parameter segment is formed after the rate limits the continuous transition parameter values. The green dashed area between t2 and t3 in the figure represents the corresponding constraint effect. The boundary reconstruction transition segment solves the problem of continuous connection between adjacent pulsating phase parameters, and the execution constraint transition segment further restricts the range and rate of change of the reconstructed parameters. After the two types of processing work together, the blue phase parameter curve maintains continuous change at each phase transition position, and corresponds to the formation relationship of the upper phase parameter sequence, boundary reconstruction transition segment, execution constraint transition segment to the composite pulsating waveform. This intuitively shows that the composite pulsating waveform is formed by the phase parameters after continuous boundary reconstruction and execution condition constraints.

[0098] S3.5. Based on the start execution time of each pulsation phase in the composite pulsation waveform, convert the phase parameters and execution constraint transition segments of each pulsation phase into parameter adjustment instructions that record the parameter name, target value, adjustment rate, and start execution time. Arrange the parameter adjustment instructions according to the start execution time to form a jigging process parameter adjustment scheme, specifically as follows: The composite pulsating waveform is divided into a phase parameter segment and an execution constraint transition segment according to the parameter name and pulsation phase. The parameter name, phase parameter value and pulsation phase start execution time are read from the phase parameter segment. The phase parameter value is recorded as the target value and the pulsation phase start execution time is recorded as the start execution time. The allowable adjustment rate is read from the jig controller parameter configuration record according to the parameter name, and the most recent valid actual phase parameter before the start execution time is read.

[0099] The required adjustment rate is determined based on the most recent valid actual phase parameter, the target value, and the duration between the start time of the pulsating phase and the target arrival time. If the required adjustment rate does not exceed the allowable adjustment rate, the required adjustment rate is recorded as the adjustment rate of the parameter adjustment command corresponding to the phase parameter segment. If the required adjustment rate exceeds the allowable adjustment rate, the allowable adjustment rate is recorded as the adjustment rate, and the target arrival time is delayed according to the allowable adjustment rate.

[0100] If no valid actual phase parameter exists before the start of execution, the default adjustment rate of the corresponding parameter name is read from the jig controller parameter configuration record, and the default adjustment rate is recorded as the adjustment rate; if no allowed adjustment rate or default adjustment rate is found, the adjustment rate missing mark is retained, and the corresponding record is not sent.

[0101] For the execution constraint transition segment, the parameter nodes are identified according to the control cycle of the jig controller. The continuous transition parameter values ​​corresponding to the parameter nodes are recorded as target values, the limited rate of change in the execution constraint transition segment is recorded as the adjustment rate, and the sampling time corresponding to the parameter nodes is recorded as the start execution time.

[0102] Records with complete parameter names, target values, adjustment rates, and start times are retained as parameter adjustment instructions; records with missing fields are not issued, and a field missing instruction marker is added; when duplicate parameter adjustment instructions with the same parameter name appear at the same start time, duplicate parameter adjustment instructions with the same target value and adjustment rate are merged; when target values ​​are inconsistent, the parameter adjustment instructions corresponding to the execution constraint transition section are retained, and a target value conflict marker is added; when target values ​​are consistent but adjustment rates are inconsistent, the adjustment rate with the smallest value is retained, and an adjustment rate conflict marker is added.

[0103] When different parameter names correspond to the same start execution time, the parallel execution relationship of parameter adjustment instructions is maintained; all valid parameter adjustment instructions are arranged according to their start execution time, and instruction field missing flags, target value conflict flags, and adjustment rate conflict flags are collected to form a jigging process parameter adjustment scheme.

[0104] The allowable adjustment rate and the default adjustment rate are obtained through the jig controller parameter configuration record and actuator calibration record; the most recent valid actual phase parameter is obtained through the parameter feedback record formed by the jig controller before the start of execution.

[0105] S3.6. Based on the start execution time in the jigging process parameter adjustment plan, parameter adjustment commands are sequentially sent to the jig controller. The jig controller adjusts the jig operating parameters corresponding to the parameter names according to the target values ​​and adjustment rates. At the boundary between adjacent pulse phases, the actual phase parameter recorded at the end of the previous pulse phase is used as the adjustment starting point for the next pulse phase, forming a batch parameter execution record. Specifically, The parameter adjustment instructions in the jigging process parameter adjustment scheme are entered into the execution sequence according to the start execution time. When the start execution time arrives, the parameter name, target value, and adjustment rate are sent to the jig controller. When the jig controller confirms that the parameter name is consistent with the adjustable jig operating parameters, it drives the corresponding jig operating parameters to change towards the target value according to the adjustment rate, and records the actual phase parameters according to the control cycle. If the parameter name cannot be identified, the target value field is missing, or the adjustment rate field is missing, the corresponding parameter adjustment instruction is not implemented, and a parameter adjustment instruction execution abnormality is recorded. If the parameter adjustment instruction is not sent at the start execution time, a delayed execution flag is recorded at the actual sending time.

[0106] The end time of the previous pulse phase is taken as the boundary time of the adjacent pulse phase. The actual phase parameters recorded at the boundary time are read and used as the starting point for adjusting the jig operation parameters corresponding to the next pulse phase. If the actual phase parameters are missing at the boundary time, the most recent valid actual phase parameters before the boundary time are used as the starting point for adjusting the next pulse phase, and the boundary parameter missing mark is retained. If there are no valid actual phase parameters before the boundary time, the adjustment of the parameters of the next pulse phase with the corresponding parameter name is paused, and the parameter execution interruption mark is registered.

[0107] The batch parameter execution record is formed by arranging the raw coal batch identifier, pulse phase identifier, parameter name, planned start time, actual issuance time, target value, adjustment rate, actual phase parameter, instruction execution status, and execution anomaly flag in chronological order.

[0108] The actual phase parameters, actual issuance time, and command execution status are obtained through the parameter feedback records and command execution logs of the jig controller.

[0109] S3.7. Retrieve the ash content and clean coal yield of raw coal batches from multi-source data, and merge the ash content, clean coal yield, and batch parameter execution records to form batch execution result data. Specifically, The product ash content record and clean coal yield record are retrieved from multi-source data according to the raw coal batch identifier. The product ash content record and clean coal yield record are then verified against the raw coal batch identifier in the batch parameter execution record. If the raw coal batch identifiers match, the product ash content and clean coal yield are included in the corresponding batch parameter execution record. If the raw coal batch identifiers do not match, the product ash content and clean coal yield are not included in the batch parameter execution record, and an abnormal batch classification mark is retained. If the product ash content record and clean coal yield record lack the raw coal batch identifier, a product quality data missing mark is retained, and the product ash content and clean coal yield of other raw coal batches are not used as substitutes.

[0110] The actual phase parameters, parameter names, pulsation phase identifiers, target values, adjustment rates, instruction execution status, and execution anomaly markers in the batch parameter execution records of the completed raw coal batch are collected according to the raw coal batch identifier. When there are duplicate records of product ash content and clean coal yield, the valid records with complete detection time and recording status are retained, and the duplicate record marker is registered. When both product ash content and clean coal yield are completed, a complete batch result record is formed. If any item is missing, a batch result record with a product quality data missing marker is formed. The complete batch result record and the batch result record with the anomaly marker are arranged in a unified manner to form the batch execution result data.

[0111] S4. Read the product ash content, clean coal yield and actual phase parameters from the batch execution result data, compare the actual phase parameters with the phase parameters one by one, and verify the batch working condition category according to the product ash content and clean coal yield to form batch verification data. Use the batch verification data to correct the support vector machine working condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

[0112] S4.1 Using the pulse phase and parameter name as the reference position, compare the actual phase parameter in the batch execution result data with the phase parameter item by item. Mark the reference position where the actual phase parameter and the phase parameter are inconsistent as the parameter deviation position, and form a phase parameter reference record. Specifically, A parameter comparison index is constructed using the raw coal batch identifier, the pulsation phase identifier, and the parameter name. The actual phase parameter corresponding to the end time of the pulsation phase is read from the batch execution result data, and the phase parameter with the same parameter comparison index is retrieved in the sub-phase prediction parameter set. When the parameter comparison index forms a unique correspondence, the actual phase parameter, phase parameter, pulsation phase identifier, and parameter name are arranged in the same comparison position. When the parameter comparison index does not form a correspondence, the parameter comparison index is retained and the parameter comparison is marked as missing. When the parameter comparison index corresponds to multiple phase parameters, the latest valid phase parameter with the formation time recorded in the sub-phase prediction parameter set is retained, and the duplicate parameter record is marked.

[0113] The deviation between the actual phase parameter and the phase parameter is verified against the parameter comparison tolerance. If the deviation does not exceed the parameter comparison tolerance, the comparison position is marked as the parameter consistent position. If the deviation exceeds the parameter comparison tolerance, the comparison position is marked as the parameter deviation position, and the deviation direction and deviation amount are recorded. If the actual phase parameter has a parameter execution interruption mark, a boundary parameter missing mark, or a parameter adjustment command execution abnormality, the corresponding comparison position is marked as the parameter execution abnormality position, and the parameter execution abnormality position is not used to judge the parameter deviation. The raw coal batch identifier, pulsation phase identifier, parameter name, actual phase parameter, phase parameter, parameter comparison status, and abnormal mark are arranged in the order of pulsation phase to form a phase parameter comparison record.

[0114] The parameter comparison tolerance is determined by the control resolution in the jig controller parameter configuration record and the measurement accuracy in the actuator feedback calibration record. The parameter comparison tolerance takes the record value with the larger value between the control resolution and the measurement accuracy.

[0115] S4.2. According to the raw coal batch, the product ash content and clean coal yield in the batch execution result data are included in the phase parameter comparison record to form a batch quality record. Specifically, The phase parameter comparison records are grouped according to the raw coal batch identifier. The ash content and clean coal yield corresponding to the same raw coal batch identifier are read from the batch execution result data, and the ash content and clean coal yield are recorded in the parameter comparison record of the same raw coal batch. When the raw coal batch identifier is consistent and the ash content and clean coal yield have valid records, the batch affiliation relationship between the product quality data and all parameter comparison positions is established. When there are missing markers for the ash content and clean coal yield, the existing product quality data is retained, and the missing markers for the product quality data are continued.

[0116] Collect the ash content, clean coal yield, parameter consistency position, parameter deviation position, and parameter execution anomaly position of the same batch of raw coal into the same batch record, and arrange the comparison position of each parameter according to the pulse phase identifier and parameter name to form a batch quality record.

[0117] S4.3. Based on the product ash content and clean coal yield in the batch quality record, retrieve the corresponding operating condition category label from the historical labeled data sample. Verify the operating condition category label against the batch operating condition category item by item, and replace the inconsistent batch operating condition category with the operating condition category label to form a batch operating condition verification record. Specifically, Using the product ash content and clean coal yield in the batch quality record as joint retrieval fields, historical labeled data samples with product ash content falling within the same product ash content range and clean coal yield falling within the same clean coal yield range are selected from the historical labeled data samples. When the selection result contains only one working condition category label, the working condition category label is used as the verification label corresponding to the raw coal batch. When the selection result contains multiple working condition category labels, the frequency of different working condition category labels in the selection result is counted, and the working condition category label with the most frequency is used as the verification label. When the frequency of occurrence is the same, the working condition category label recorded in the historical labeled data sample with the latest formation time and manual verification record is used. When no historical labeled data sample that meets both quality ranges is selected, the original batch working condition category is retained, and the working condition category label is marked as missing.

[0118] Verify the verification label against the batch operating condition category; if the verification label matches the batch operating condition category, retain the batch operating condition category and record that the operating condition category verification is consistent; if the verification label does not match the batch operating condition category, replace the batch operating condition category with the verification label, and record the batch operating condition category before replacement, the batch operating condition category after replacement, and the operating condition category replacement mark; if the product ash content and clean coal yield have a product quality data missing mark, do not perform operating condition category label retrieval, retain the batch operating condition category and record that the operating condition category verification is incomplete, and form a batch operating condition verification record.

[0119] Among them, the historical labeled data samples are generated from the historical raw coal batch records that have completed product quality testing and have been reviewed by production personnel. The historical labeled data samples record the raw coal batch identifier, product ash content, clean coal yield, working condition label, formation time and manual review status.

[0120] S4.4. Merge batch quality records and batch operating condition verification records according to raw coal batches to form batch verification data, specifically as follows: Using the raw coal batch identifier as the merging field, the product ash content, clean coal yield, and phase parameter comparison records in the batch quality record are linked to the batch operating condition category, verification label, and operating condition category replacement mark in the batch operating condition verification record to establish a batch-level correspondence. When the raw coal batch identifier forms a unique record in both the batch quality record and the batch operating condition verification record, the two records are merged into the same batch master record. When the batch quality record is missing, the batch operating condition verification record is retained and a batch quality record missing mark is added.

[0121] Attach parameter deviation position, parameter execution abnormal position, product quality data missing mark, working condition category replacement mark, and working condition category verification status to the corresponding batch master record. Arrange all batch master records according to the raw coal batch formation time to form batch verification data.

[0122] S4.5. Using whether the batch operating condition category has been replaced and whether the parameter deviation exists as the branching conditions, batch verification data that only has a batch operating condition category replacement is classified as category correction data, batch verification data that only has a parameter deviation is classified as parameter correction data, and batch verification data that has both a batch operating condition category replacement and a parameter deviation is classified as joint correction data. The category correction data, parameter correction data, and joint correction data are then combined to form the branched batch verification data. Specifically, Read the operating condition category replacement mark and parameter deviation position from the batch verification data. Batch verification data with the operating condition category replacement mark and no parameter deviation position are classified into the classification correction data. Batch verification data with no operating condition category replacement mark and at least one parameter deviation position are classified into the parameter correction data. Batch verification data with the operating condition category replacement mark and at least one parameter deviation position are classified into the joint correction data.

[0123] For batch verification data that do not have a working condition category replacement mark and do not have parameter deviation position, record the verification pass status and do not include them in the model correction branch; for batch verification data with missing product quality data mark, missing parameter comparison mark, and missing working condition category label mark, retain the verification abnormal status and do not include them in the model correction branch until the missing content is filled in; retain the branch type mark for the classification correction data, parameter correction data, and joint correction data respectively, and arrange them according to the formation time of the raw coal batch to form the branch batch verification data.

[0124] S4.6. Using the parameter deviation position in the parameter correction data and joint correction data as the filtering criteria, select batch verification data with consistent parameter deviation positions from the historical batch verification data. Arrange the selected batch verification data and the branch batch verification data according to their formation time to generate a calibration sample sequence. Specifically, The pulse phase identifier and parameter name recorded in the parameter correction data and joint correction data are combined into a parameter deviation position set. The batch verification data with completely consistent parameter deviation position sets are searched in the historical batch verification data. When the parameter deviation position set in the historical batch verification data and the current parameter deviation position set contain the same pulse phase identifier and the same parameter name, the corresponding historical batch verification data is retained. When the parameter deviation position set has missing items or additional parameter deviation positions, it is not included in the current screening results.

[0125] Records with duplicate raw coal batch identifiers, missing product quality data, missing parameter comparisons, and incomplete operating condition category verification are removed from the screening results. When raw coal batch identifiers are duplicated, the historical batch verification data with the latest formation time and complete verification fields is retained. When no historical batch verification data with the same parameter deviation position set is found, the current parameter correction data and joint correction data are retained, and the missing historical correction samples are marked. The valid historical batch verification data and branch batch verification data are arranged in chronological order of formation time, and the classification correction type, parameter correction type, and joint correction type are retained to generate a correction sample sequence.

[0126] Among them, the historical batch verification data is formed by saving the batch verification data of previous raw coal batches after field integrity checks and manual review.

[0127] S4.7. Update the classification hyperplane of the support vector machine (SVM) condition classifier by combining the raw coal batch feature dataset corresponding to the classification correction data and joint correction data in the correction sample sequence with the validated batch condition categories, thus obtaining the updated SVM condition classifier. Specifically, Classification correction data and joint correction data are selected from the correction sample sequence. The corresponding raw coal batch feature dataset is retrieved according to the raw coal batch identifier, and the replaced batch condition category in the batch condition verification record is used as the correction label of the corresponding raw coal batch feature dataset. When the feature field order of the raw coal batch feature dataset is consistent with the original training feature field order of the support vector machine condition classifier, the raw coal batch feature dataset is retained. When feature fields are missing, the feature field order cannot be recovered, or the correction label is missing, the classification training sample is marked as abnormal and does not participate in the classification hyperplane update.

[0128] New classification training samples are formed using a valid raw coal batch feature dataset and calibration labels. These new training samples are then combined with the existing support vector samples from the support vector machine (SVM) working condition classifier to form classification update samples. During the classification update process, the original feature scaling parameters and kernel function parameters remain unchanged. The calibration labels constrain the classification results, and the support vectors, kernel function coefficients, and bias terms are adjusted using a hinge loss with regularization constraints, causing misclassified samples to move to the region corresponding to the validated batch working condition category. After the classification update samples have completed training, the classification results are verified using valid batch validation data that did not participate in the update. Training continues as long as the validation misclassification rate continues to decrease, and ends when the validation misclassification rate stops decreasing. If the updated validation misclassification rate is higher than the pre-update validation misclassification rate, the classification parameters are restored to their original state. The validated support vectors, kernel function coefficients, bias terms, and classification hyperplane are saved to the SVM working condition classifier to obtain the updated SVM working condition classifier.

[0129] The updated training data for the support vector machine (SVM) working condition classifier includes the raw coal batch feature dataset corresponding to the classification correction data and joint correction data, the verified batch working condition categories, and the original support vector samples; the update loss adopts the hinge loss with regularization constraints; the update mechanism adopts iterative adjustment of support vectors, kernel function coefficients, and bias terms; the update completion condition is that the verification misclassification rate stops decreasing and is not higher than the verification misclassification rate before the update.

[0130] S4.8. Update the network weights of the time-series neural network model by comparing the phase-series feature sequences corresponding to the parameter correction data and joint correction data in the correction sample sequence with the actual phase parameters. Use the updated support vector machine condition classifier and the updated time-series neural network model as the intelligent control results of the jigging process parameters. Specifically, Parameter correction data and joint correction data are selected from the correction sample sequence. The corresponding phase sub-sequences are retrieved according to the raw coal batch identifier and the pulsation phase identifier. The actual phase parameters are attached to the phase end position of the phase sub-sequence according to the parameter name. When the phase sub-sequence is complete in time sequence and the actual phase parameters have valid records, the phase sub-sequence is used as the network update sample and the actual phase parameters are used as the phase parameter correction target. When the phase sub-sequence has an interruption mark, the interruption position is retained and the missing state mark is used to participate in the time sequence propagation. When the actual phase parameter has a parameter execution anomaly mark, the corresponding pulsation phase is not used to update the network weights.

[0131] The mean square error between the predicted phase parameters and the actual phase parameters corresponding to the phase-sequence features is used as the parameter correction error, and the hidden state transitions at the boundaries of adjacent pulsating phases are included in the boundary continuity error. The parameter correction error and the boundary continuity error are used together as the network update loss, and backpropagation is performed along the time direction of the phase-sequence features to adjust the network weights in the temporal neural network model that participate in the temporal state transmission and phase parameter output. After completing one training cycle, the network update loss is verified using valid correction samples that did not participate in the network weight adjustment. Training continues when the verification loss continues to decrease and ends when the verification loss stops decreasing. When the updated verification loss is higher than the previous verification loss, the network weights are restored to their previous state.

[0132] The verified network weights, updated sample formation time periods, and model versions are registered in the updated temporal neural network model. The updated support vector machine condition classifier and the updated temporal neural network model are then combined to form the intelligent control results of the jigging process parameters.

[0133] The training data for updating the temporal neural network model includes the phase-sequence feature sequences and actual phase parameters corresponding to the parameter correction data and joint correction data; the update loss consists of the parameter correction error and the boundary continuity error; the update mechanism uses backpropagation along the time direction to adjust the network weights; the update completion condition is that the validation loss stops decreasing and is not higher than the validation loss before the update.

[0134] This embodiment also provides an intelligent control system for jigging process parameters based on multi-algorithm fusion, including: The acquisition module is used to acquire multi-source data of the jigging process, perform feature engineering on the multi-source data, divide the raw coal into batches, and generate a raw coal batch feature dataset. The deduction module is used to identify the working conditions of raw coal batch feature dataset in real time through a support vector machine working condition classifier, output the working condition category of the batch, select the continuous features corresponding to the raw coal batch according to the working condition category, and divide the continuous features according to the time sequence of the pulsation phase, using the opening and closing time of the air valve as the boundary of the pulsation phase, forming a phase-sequence feature sequence. The phase parameters of each pulsation phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model. The execution module is used to splice the phase parameters of each pulsating phase to form a composite pulsating waveform, convert the composite pulsating waveform into a jigging process parameter adjustment scheme and send it to the jigging machine controller for execution, and generate batch execution result data. The calibration module is used to read the product ash content, clean coal yield and actual phase parameters from the batch execution result data, compare the actual phase parameters with the phase parameters one by one, and verify the batch operating condition category according to the product ash content and clean coal yield to form batch calibration data. The batch calibration data is used to correct the support vector machine operating condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

[0135] This embodiment also provides a computer device applicable to the intelligent control method for jigging process parameters based on multi-algorithm fusion, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent control method for jigging process parameters based on multi-algorithm fusion proposed in the above embodiment.

[0136] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0137] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent control method for jigging process parameters based on multi-algorithm fusion as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0138] In summary, this invention improves the continuity of phase parameters through gating the transition between pulsating phase boundaries. The bed change trend formed at the end of the previous pulsating phase is continuously preserved, and the characteristics at the beginning of the next pulsating phase participate in continuous deduction according to the state update ratio, reducing state interruptions and parameter abrupt changes caused by the transition between valve opening and closing phases. The phase parameters maintain smooth transition at the pulsating boundary position, the timing coordination relationship between jigging frequency, target bed thickness, and discharge speed is more stable, the continuity and executability of the composite pulsating waveform are improved simultaneously, control command jumps and frequent actuator corrections are suppressed, and the bed movement process is consistent with the parameter adjustment process, thereby improving the operational stability, parameter adjustment accuracy, and product quality consistency of the jigging sorting process.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent control of jigging process parameters based on multi-algorithm fusion, characterized in that: include, Multi-source data of the jigging process is acquired, feature engineering is performed on the multi-source data, and raw coal is divided into batches to generate raw coal batch feature datasets. Specifically, Abnormal data segments in the multi-source data are removed, and linear interpolation is used to fill the data gaps formed after the removal of abnormal data segments to obtain complete multi-source data. Smoothing filtering is applied to the complete multi-source data to form a preprocessed dataset. The preprocessed dataset is divided into fixed time windows. Statistical operations are performed on the preprocessed data within each time window to form statistical features. The preprocessed data within each time window are staggered sequentially along the sampling time. The covariance of the overlapping data segments before and after the staggering is calculated. The covariance calculation results are arranged according to the staggering order to form time-series features. The statistical features and time-series features are concatenated to form window features. Mark the locations of raw coal ash content changes and production load changes in chronological order along the window features, and delineate the raw coal batch boundaries based on the locations of raw coal ash content changes and production load changes. By segmenting the window features of raw coal batch boundaries, the window features between adjacent raw coal batch boundaries are classified into the same raw coal batch, generating a raw coal batch feature dataset. A support vector machine (SVM) condition classifier is used to identify the batch condition of raw coal in real time, outputting the batch condition category. The raw coal batch feature dataset is spatially mapped by the radial basis kernel function of the support vector machine working condition classifier to form the raw coal batch mapped feature set. Based on the classification hyperplane in the support vector machine working condition classifier, the raw coal batch mapping features in the raw coal batch mapping feature set are divided, and the raw coal batch mapping features located in the same classification region are classified into the same working condition category, thus forming the raw coal batch working condition classification result. According to the working condition category classification results of raw coal batches, the working condition category corresponding to each batch of raw coal is marked as the batch working condition category; According to the batch working condition category, the continuous features corresponding to the raw coal batch are selected, and the opening and closing time of the air valve is used as the boundary of the pulsating phase. The continuous features are divided according to the time sequence to form a phase-sequence feature sequence. The phase parameters of each pulsating phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model. The phase parameters of each pulsation phase are spliced ​​together to form a composite pulsation waveform. The composite pulsation waveform is then converted into a jigging process parameter adjustment scheme and sent to the jigging machine controller for execution, generating batch execution result data. The product ash content, clean coal yield, and actual phase parameters are read from the batch execution result data. The actual phase parameters are compared with the phase parameters one by one, and the batch operating condition category is verified based on the product ash content and clean coal yield to form batch verification data. The batch verification data is used to correct the support vector machine operating condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

2. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 1, characterized in that: The multi-source data includes raw coal ash content, product ash content, clean coal yield, production load, and real-time operating data collected from jig air valves and bed sensors.

3. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 2, characterized in that: The formation of the phase-sequence feature sequence specifically involves, Based on the batch working condition category, continuous features corresponding to each raw coal batch are selected from the raw coal batch feature dataset, and the sampling order of continuous features within the raw coal batch is retained; Among the selected continuous features, the opening and closing times of the air valves within the same batch of raw coal are alternately arranged according to the sampling order to form an air valve edge sequence; According to the time sequence of the valve edge sequence, the time from the opening of the valve to the closing of the next valve is defined as the valve opening phase, and the time from the closing of the valve to the opening of the next valve is defined as the valve closing phase. The start and end times of the valve opening phase and the valve closing phase are combined to form a pulsating phase boundary pair. Using the start and end times of the pulsating phase boundary pairs as the intercept positions, continuous features corresponding to the valve opening phase and valve closing phase are intercepted respectively to obtain phase-segmented feature segments; Based on the time sequence of the pulsating phase boundary pairs, the phase segment corresponding to the valve opening phase is sequentially connected with the phase segment corresponding to the next valve closing phase to form a phase segment feature sequence.

4. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 3, characterized in that: The phase parameters for generating each pulsation phase are specifically as follows: The phase-series features are derived by time-by-time using a temporal neural network model according to the sampling order. Based on the pulsating phase boundary, the timing deduction results of the valve opening phase and valve closing phase are extracted from the phase timing characteristics to form phase-specific deduction characteristics; Starting with the first phase-separated feature, the phase-separated features with adjacent time are gated and connected at the pulsating phase boundary to form a cross-phase inference sequence. The phase inference results are generated by continuing the inference of the cross-phase inference sequence through a time-series neural network model. Using the end time of each pulsation phase as the index position, the deduction features located at the index position in the phase deduction results are used as phase characterization features; The output layer of the temporal neural network model performs parameter mapping on the phase representation features to generate phase parameters for each pulsating phase.

5. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 4, characterized in that: The batch execution result data is specifically generated as follows: The phase parameters of each pulsation phase are arranged in chronological order to form a phase parameter sequence; Using the boundary position between adjacent pulsating phases in the phase parameter sequence as the reconstruction position, monotonically piecewise cubic Hermite interpolation is used to set continuous transition parameter values ​​between the end parameter value of the previous pulsating phase and the beginning parameter value of the next pulsating phase, forming a boundary reconstruction transition segment; Based on the parameter range and adjustment rate allowed by the jig controller, continuous transition parameter values ​​above the upper limit of the parameter range are replaced with the upper limit value of the parameter range, and continuous transition parameter values ​​below the lower limit of the parameter range are replaced with the lower limit value of the parameter range. The rate of change of adjacent continuous transition parameter values ​​is limited to the allowed adjustment rate, forming an execution constraint transition segment. The execution constraint transition segment is embedded into the corresponding boundary position in the phase parameter sequence, and the phase parameters and execution constraint transition segments of each pulsating phase are concatenated along the time sequence of the pulsating phase to form a composite pulsating waveform; According to the start execution time of each pulsation phase in the composite pulsation waveform, the phase parameters and execution constraint transition segments of each pulsation phase are converted into parameter adjustment instructions that record the parameter name, target value, adjustment rate and start execution time. The parameter adjustment instructions are arranged according to the start execution time to form a jigging process parameter adjustment scheme. According to the start execution time in the jigging process parameter adjustment plan, the parameter adjustment instructions are sequentially sent to the jigging machine controller. The jigging machine controller adjusts the jigging machine operating parameters corresponding to the parameter names according to the target value and adjustment rate. At the junction of adjacent pulse phases, the actual phase parameter recorded at the end of the previous pulse phase is used as the adjustment starting point of the next pulse phase, forming a batch parameter execution record. The product ash content and clean coal yield of raw coal batches are retrieved from multi-source data. The product ash content, clean coal yield and batch parameter execution records are merged to form batch execution result data.

6. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 5, characterized in that: The intelligent control results for forming jigging process parameters are specifically as follows: Using the pulse phase and parameter name as the reference position, the actual phase parameter in the batch execution result data is compared with the phase parameter item by item. The reference position where the actual phase parameter and the phase parameter are inconsistent is marked as the parameter deviation position, thus forming a phase parameter comparison record. According to the raw coal batch, the product ash content and clean coal yield in the batch execution result data are included in the phase parameter comparison record to form the batch quality record; Based on the product ash content and clean coal yield in the batch quality record, retrieve the corresponding working condition category label from the historical labeled data sample, verify the working condition category label with the batch working condition category item by item, replace the inconsistent batch working condition category with the working condition category label, and form a batch working condition verification record. The batch quality records and batch operating condition verification records are merged according to the raw coal batches to form batch verification data; Based on whether the batch operating condition category has been replaced and whether the parameter deviation exists, the batch verification data that only has the batch operating condition category replaced is classified as category correction data, the batch verification data that only has the parameter deviation is classified as parameter correction data, and the batch verification data that has both the batch operating condition category replaced and the parameter deviation is classified as joint correction data. The category correction data, parameter correction data and joint correction data are combined to form the branched batch verification data. Using the parameter deviation position in the parameter correction data and joint correction data as the filtering condition, batch verification data with the same parameter deviation position are selected from the historical batch verification data. The selected batch verification data and the branch batch verification data are arranged according to the formation time to generate a correction sample sequence. The classification hyperplane of the support vector machine condition classifier is updated by combining the raw coal batch feature dataset corresponding to the classification correction data and joint correction data in the correction sample sequence with the verified batch condition category, thus obtaining the updated support vector machine condition classifier. The network weights of the temporal neural network model are updated by correcting the phase feature sequences corresponding to the parameter correction data and joint correction data in the correction sample sequence and the actual phase parameters. The updated support vector machine condition classifier and the updated temporal neural network model are used as the intelligent control results of jigging process parameters.

7. The intelligent control method for jigging process parameters based on multi-algorithm fusion as described in claim 6, characterized in that: The continuous transition parameter value is an intermediate parameter value set in chronological order between the end parameter value of the previous pulsation phase and the beginning parameter value of the next pulsation phase, based on monotonically piecewise cubic Hermite interpolation. The end parameter value of the previous pulsation phase is the last valid jigging process parameter value corresponding to the end time of the previous pulsation phase. The starting parameter value of the next pulsation phase is the first effective jigging process parameter value corresponding to the beginning moment of the next pulsation phase.

8. An intelligent control system for jigging process parameters based on multi-algorithm fusion, based on the intelligent control method for jigging process parameters based on multi-algorithm fusion as described in any one of claims 1 to 7, characterized in that: include, The acquisition module is used to acquire multi-source data of the jigging process, perform feature engineering on the multi-source data, divide the raw coal into batches, and generate a raw coal batch feature dataset. The deduction module is used to identify the working conditions of raw coal batch feature dataset in real time through a support vector machine working condition classifier, output the working condition category of the batch, select the continuous features corresponding to the raw coal batch according to the working condition category, and divide the continuous features according to the time sequence of the pulsation phase, using the opening and closing time of the air valve as the boundary of the pulsation phase, forming a phase-sequence feature sequence. The phase parameters of each pulsation phase are generated by performing time-series deduction on the phase-sequence feature sequence through a time-series neural network model. The execution module is used to splice the phase parameters of each pulsating phase to form a composite pulsating waveform, convert the composite pulsating waveform into a jigging process parameter adjustment scheme and send it to the jigging machine controller for execution, and generate batch execution result data. The calibration module is used to read the product ash content, clean coal yield and actual phase parameters from the batch execution result data, compare the actual phase parameters with the phase parameters one by one, and verify the batch operating condition category according to the product ash content and clean coal yield to form batch calibration data. The batch calibration data is used to correct the support vector machine operating condition classifier and the time-series neural network model to form the intelligent control result of jigging process parameters.

Citation Information

Patent Citations

  • Efficient and high-quality jigging control method of thermal coal

    CN101875025A

  • Clean coal ash content prediction optimization method based on dense medium washing

    CN118469120A