Intelligent dry separation dynamic optimization method and system for coal preparation plant

Abstract

The present application relates to the field of intelligent control of coal preparation plant, and discloses a kind of coal preparation plant intelligent dry separation dynamic optimization method and system.The method is by obtaining raw coal property data and environmental parameters, generates standardized data matrix through multi-source data fusion processing;Based on deep learning network generates parameter suggestion set and carries out feasibility check;Intelligent optimization algorithm is used to search parameter space, and generates optimized parameter combination;Finally, through instruction conversion and encapsulation, the dynamic adjustment of process parameters is realized.The present application realizes the intelligent decision and dynamic optimization of coal preparation process, has the advantages of high data processing accuracy, high parameter optimization efficiency, fast system response speed, etc., can significantly improve the clean coal recovery rate, reduce energy consumption, and ensure the safe and stable operation of production.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of coal preparation plants, and in particular to a dynamic optimization method and system for intelligent dry coal preparation plants. Background Technology

[0002] In the field of intelligent control of coal preparation plants, existing solutions for the same object / scenario as a dynamic optimization method and system for intelligent dry coal preparation plants typically rely on traditional physical sorting techniques and fixed parameter settings. These solutions suffer from limitations such as difficulty in real-time adjustment of coal preparation parameters, inability to effectively handle multi-source data fusion and prediction, and a lack of dynamic optimization mechanisms. Existing methods largely depend on static models and empirical judgments, which can easily lead to low sorting efficiency and resource waste under complex coal quality variations and environmental conditions, making it difficult to achieve stable dynamic optimization in coal preparation plants. Regarding the joint processing of how to achieve dynamic adjustment during the dry coal preparation process in coal preparation plants based on raw data streams and complete datasets through multi-source data fusion, LSTM parameter prediction, and genetic algorithm collaborative optimization, existing technologies generally suffer from common shortcomings in data fusion, parameter prediction, and optimization control. This makes it difficult to form a consistent data acquisition-processing-prediction-optimization control process in dynamic optimization application scenarios in coal preparation plants, resulting in insufficient dynamic adjustment of coal preparation parameters. Summary of the Invention

[0003] This invention provides a method and system for dynamic optimization of intelligent dry separation in coal preparation plants, which solves the problem of how to achieve dynamic adjustment in the dry separation process of coal preparation plants based on raw data streams and complete datasets through multi-source data fusion, LSTM parameter prediction and genetic algorithm collaborative optimization.

[0004] To address the aforementioned technical problems, this invention provides a dynamic optimization method for intelligent dry coal preparation plants, comprising:

[0005] The particle size distribution and density characteristics of raw coal, along with environmental parameters, are acquired. Noise filtering, missing value imputation, and time series alignment are then performed to generate a standardized data matrix. The environmental parameters include temperature, humidity, and vibration intensity data. The noise filtering process includes: using a combination of wavelet transform and adaptive filtering algorithms to identify and remove high-frequency noise and abnormal spikes in the signal; dynamically adjusting the filtering parameters according to a preset threshold; and marking and recording abnormal data points in the log system.

[0006] Obtain a standardized data matrix, perform sliding window segmentation, normalization, and time dimension expansion to generate a feature vector set and a three-dimensional feature tensor;

[0007] Based on the three-dimensional feature tensor, perform forward propagation operation and regularization correction of LSTM network to generate parameter proposal set and perform feasibility verification to generate effective parameter space;

[0008] To obtain an effective parameter space, the population is initialized, parallel fitness calculation is performed, elite selection and crossover mutation are carried out, and an optimized parameter combination is generated. The process of generating the optimized parameter combination includes: adopting an adaptive mutation rate adjustment mechanism to automatically adjust the mutation amplitude according to the number of iterations and population diversity; periodically introducing random new individuals to prevent getting trapped in local optima; after completing the genetic operation, updating the population structure to form a new generation of candidate parameter set containing elite individuals and newly generated individuals.

[0009] Obtain optimized parameter combinations, perform device instruction conversion, safety boundary verification, and vibration- and temperature-resistant encapsulation, and generate execution instruction packages;

[0010] Receive execution instruction packets, monitor the dry separation process, extract energy efficiency indicators for evaluation and recovery rate calculation, and generate optimization suggestion parameters.

[0011] Furthermore, the process of generating a standardized data matrix includes:

[0012] Key features were extracted from multi-source data. A method combining principal component analysis and information entropy assessment was used to automatically screen out key indicators that reflect the trend of raw coal particle size change, density distribution characteristics and environmental fluctuations, and to remove redundant information.

[0013] For missing values ​​in key feature data, an algorithm combining temporal interpolation and multiple interpolation is used for filling. First, the missing pattern is determined, and then, based on the complete data in the adjacent time window and the correlation of environmental parameters, a reasonable estimate of the missing values ​​is achieved.

[0014] The completed data underwent multiple rounds of iterative verification to ensure that its statistical characteristics were highly consistent with the distribution of the original data, thus generating a complete dataset.

[0015] Furthermore, generating a standardized data matrix also includes:

[0016] Time series alignment processing is performed on multi-source data. An algorithm based on dynamic time warping and linear interpolation is used to adjust the time axis of each data stream to achieve accurate synchronization of multi-source data, while taking into account device clock drift and network transmission latency.

[0017] The aligned data undergoes standardization processing using a Z-score-based standardization method to transform the data for each feature dimension into a standard normal distribution interval with a mean of zero and a variance of one, thereby eliminating dimensional differences.

[0018] During the standardization process, the mean and variance parameters are dynamically updated to adapt to the time-varying characteristics of raw coal properties and generate a standardized data matrix.

[0019] Furthermore, the process of obtaining the effective parameter space includes:

[0020] A multidimensional data set of parameter combinations is extracted from the effective parameter space and used as the initial solution set for genetic algorithm optimization, so that it meets the physical limits of the equipment, the process flow and safety standards.

[0021] A strategy combining random sampling and stratified sampling is adopted to initialize the population of the genetic algorithm. The parameter space is uniformly covered by the Latin hypercube sampling method to avoid local clustering during population initialization.

[0022] By combining historically optimal parameter combinations, a hybrid initialization mechanism is adopted, setting some individuals as empirically optimal solutions to accelerate the convergence speed;

[0023] The parameters of each individual are encoded using real numbers to ensure the continuity and accuracy of the genetic operation. The encoding structure includes key dry sorting parameters such as wind speed, vibration frequency, sorting angle, and feeding speed.

[0024] Furthermore, the process of generating the optimal parameter combination includes:

[0025] Each individual parameter combination is extracted from the candidate parameter set, and the fitness function is calculated in parallel to evaluate the performance of each individual in the dry separation process.

[0026] Based on a predefined multi-objective fitness function, and taking into account the clean coal recovery rate, energy consumption index and equipment operation safety, a fitness evaluation model is constructed.

[0027] The fitness calculation task is assigned to edge computing nodes or cloud computing platforms that support parallel computing, and a multi-threaded and distributed computing framework is adopted to achieve efficient evaluation of the candidate parameter set.

[0028] Each computing node independently executes the simulation model, inputs individual parameters from the candidate parameter set, simulates the material sorting behavior and equipment response during the dry sorting process, and outputs the corresponding performance indicators.

[0029] After completing the fitness assessment of all individuals, the results of each calculation node are summarized to form a scoring matrix, in which each element corresponds to the overall fitness value of an individual.

[0030] Furthermore, generating optimized parameter combinations also includes:

[0031] Individuals in the candidate parameter set are ranked according to their fitness values, and an elite retention strategy is adopted to directly copy the top-ranked individuals to the next generation.

[0032] Based on the fitness ratio selection method, parent individuals are randomly selected from the remaining individuals for crossover.

[0033] The crossover operation employs a simulated binary crossover method with real-number encoding, combined with dynamic crossover probability, to generate diverse offspring individuals that meet the constraints.

[0034] Mutation operations are performed on the offspring individuals after crossover, using a Gaussian mutation strategy to dynamically adjust the mutation amplitude to balance exploration and utilization.

[0035] Furthermore, the process of crossover mutation treatment includes:

[0036] After mutation, constraint satisfaction checks are performed on all newly generated individuals, and parameter combinations that do not meet the safety boundaries are eliminated or corrected.

[0037] Furthermore, the generation of the standardized data matrix includes:

[0038] The noise filtering algorithm is executed, and the function expression is:

[0039]

[0040] Among them, T wave To improve the wavelet threshold function; x s σ represents the raw signal of the s-th sensor channel; s α represents the standard deviation of the corresponding channel noise. 1 =3.5 and β 1 =2.2 is the adaptive shape parameter;

[0041] Furthermore, an adaptive filtering algorithm is employed:

[0042] y t =γ t ·W t ·x t +(1-γ t )·y t-1

[0043] Among them, y t The filtered output signal at time t; x t W represents the original input signal at time t; t γ is a time-varying filter matrix; t Forgetting factor; y t-1 This represents the filtered output signal at time t-1;

[0044] Furthermore, principal component analysis (PCA) is used to extract key features:

[0045]

[0046] in, Let i be the weight of the i-th feature in the p-th principal component; Let H be the eigenvalue of the p-th principal component; n represents the total number of feature dimensions; maxH p H represents the maximum value of the feature information entropy in the p-th principal component;p (x i ) represents the information entropy of the i-th feature; ∈ represents the smoothing coefficient;

[0047] Furthermore, missing value imputation employs spatiotemporal interpolation:

[0048]

[0049] in, The result of imputing missing values ​​at time t; The summation symbol indicates that the summation is performed over the time index k from the lower bound k = t - Δ to the upper bound k = t + Δ; Δ is the size of the time window; ω k As a time decay weight; Observations; The ambient humidity gradient is η; η = 0.2 is the coupling coefficient.

[0050] Alignment is performed using a dynamic time warping algorithm:

[0051]

[0052] Among them, D sync D is a measure of the synchronization error between two time series. sync (X i ,X j (x) represents two time series X i and X j Synchronization error measurement between; X i (m),X j (n) represents the data from the i-th and j-th sensors at time points m and n; A is the set of alignment paths; To find the minimum value on set A; μ is the square of the L2 norm; i and σ i ...

[0053] Furthermore, standardization is carried out:

[0054]

[0055] Among them, Z k For standardized data; x k This is the original data; and δ is the sliding window statistic; δ is the stability factor.

[0056] Furthermore, the process of generating the optimal parameter combination includes:

[0057] Initialize the population by sampling and define multidimensional parameter sampling:

[0058]

[0059] in, Let π be the d-th dimension parameter value of the i-th individual; d It is a randomly arranged sequence of integers; The numbers are uniformly distributed random numbers; The upper and lower bounds of the parameters are defined; N is the population size.

[0060] Furthermore, parallel fitness calculations are performed, and a multi-objective fitness function is defined:

[0061]

[0062] Among them, F i R represents the multi-objective fitness value of the i-th individual. i E represents the recovery rate of clean coal. i Energy consumption indicators; Let be the deviation of the j-th safety constraint; Let α be the safety constraint limit for the j-th term; 2 β 2 γ 2 R represents the weighting coefficients; min ,R max E min E max These are the minimum and maximum values ​​of clean coal recovery rate and energy consumption indicators in historical data;

[0063] Furthermore, a genetic operation is performed on the scoring matrix to define simulated binary crossover:

[0064]

[0065] in, This represents the d-th dimension parameter value of the k-th offspring individual after crossover; The parameter values ​​corresponding to the two individuals in the parent generation; β q is the control parameter for the crossover operation; u is a random variable uniformly distributed in [0,1].

[0066] Furthermore, we define adaptive Gaussian mutation:

[0067]

[0068] in, This represents the new value of the d-th dimension parameter for the k-th individual after mutation; σ represents the original value of the d-th dimension parameter of the k-th individual before mutation; d The magnitude of variation in parameter dimensions; A standard normally distributed random number generator; g is the current iteration number; Gmax This represents the maximum number of iterations.

[0069] Furthermore, an intelligent dry separation dynamic optimization system for coal preparation plants, applied to any of the methods described above, includes:

[0070] The installation unit is used to acquire the installation parameters of the dry separation equipment and complete the installation requirements.

[0071] The data acquisition unit is used to collect raw coal property data and environmental parameters, and to perform noise filtering within a preset monitoring range to obtain the raw data stream.

[0072] The data processing unit is used to extract key features from the raw data stream, imputate missing values, generate a complete dataset, and perform time series alignment on the complete dataset to generate a standardized data matrix.

[0073] The judgment and arbitration unit is used to judge and arbitrate events based on a standardized data matrix and judgment thresholds, and output control commands.

[0074] The execution unit is used to output control commands to the execution components of the dry separation equipment to complete dynamic adjustment and control.

[0075] The recording and updating unit is used to record the effective time information of control commands and update the parameter adjustment strategy table.

[0076] The key innovations of this invention include:

[0077] (1) By acquiring raw coal property data and environmental parameters and performing noise filtering, a raw data stream is generated to achieve dynamic optimization of multi-source data fusion.

[0078] (2) The LSTM parameter prediction model is used to align the complete dataset to time series and generate a standardized data matrix, thereby improving the accuracy and timeliness of data prediction.

[0079] (3) Use genetic algorithms to initialize the population and select elites to generate optimal parameter combinations, so as to realize the dynamic adjustment and optimization of coal preparation parameters.

[0080] The following are its main beneficial effects:

[0081] (1) The generation of raw data stream can effectively handle the problem of multi-source data fusion, ensure the consistency of data acquisition and processing in the dry separation process of coal preparation plant, and is applicable to varying coal quality and environmental conditions.

[0082] (2) The generation of standardized data matrices improves the prediction accuracy of time series data, making parameter prediction based on LSTM model more reliable and suitable for complex dynamic optimization scenarios in coal preparation plants.

[0083] (3) The generation of optimized parameter combinations effectively solves the limitations of traditional fixed parameter settings through the dynamic adjustment mechanism of genetic algorithm, and is suitable for the needs of real-time optimization control in the dry separation process of coal preparation plant. Attached Figure Description

[0084] Figure 1 A flowchart illustrating a dynamic optimization method for intelligent dry coal preparation in a coal preparation plant, provided as an embodiment of this application;

[0085] Figure 2 This is a structural block diagram of an intelligent dry separation dynamic optimization system for a coal preparation plant, provided as an embodiment of this application. Detailed Implementation

[0086] Example 1: Refer to Figure 1 This is a flowchart illustrating a dynamic optimization method for intelligent dry coal preparation in a coal preparation plant, provided by an embodiment of the present invention. The process may include at least steps S100-S600:

[0087] S100: Obtain the particle size distribution, density characteristics, and environmental parameters of raw coal; perform noise filtering, missing value imputation, and time series alignment processing to generate a standardized data matrix.

[0088] S200: Obtain a standardized data matrix, perform sliding window segmentation, normalization, and time dimension expansion to generate a feature vector set and a three-dimensional feature tensor.

[0089] S300: Based on the three-dimensional feature tensor, perform forward propagation operation and regularization correction of LSTM network to generate parameter proposal set and perform feasibility verification to generate effective parameter space;

[0090] S400: Obtain the effective parameter space, perform population initialization, parallel fitness calculation, elite selection and crossover mutation processing, and generate optimized parameter combinations;

[0091] S500: Obtain optimized parameter combinations, perform device instruction conversion, safety boundary verification, and vibration- and temperature-resistant encapsulation, and generate execution instruction packages;

[0092] S600 receives and executes instruction packets, monitors the dry separation process, extracts energy efficiency indicators for evaluation and recovery rate calculation, and generates optimized parameter suggestions.

[0093] Step S100 includes at least steps S110-S130:

[0094] S110. Obtain raw coal property data and environmental parameters, perform noise filtering processing, and obtain the raw data stream;

[0095] Specifically, the particle size distribution, density characteristics, and environmental parameters of raw coal, including but not limited to temperature, humidity, and vibration intensity, are collected in real time from a multi-source sensor system as input. The multi-source sensor system transmits data to the data acquisition interface module via industrial Ethernet or wireless transmission protocols to ensure data integrity and timeliness. Further, noise filtering is performed on the acquired raw data stream using a combination of wavelet transform (WT) and adaptive filtering algorithms to identify and remove high-frequency noise and abnormal spikes in the signal, ensuring data stability and accuracy. During noise filtering, the system dynamically adjusts the filtering parameters based on preset thresholds to ensure adaptability to different sensor data. Abnormal data points are marked and recorded in the log system for subsequent anomaly analysis and tracking. After noise filtering, the generated raw data stream serves as the output field of this step, accurately reflecting the temporal information of the preprocessed raw coal properties and environmental parameters. This output field, "Raw Data Stream," is consumed as input in subsequent step S120 for key feature extraction and missing value imputation.

[0096] S120. Extract key features from the original data stream, imput missing values, and generate a complete dataset.

[0097] This step takes the "raw data stream" output by S110 as input and first extracts key features from the multi-source data. Specifically, a method combining Principal Component Analysis (PCA) and information entropy assessment is used to automatically select key indicators reflecting the trend of raw coal particle size variation, density distribution characteristics, and environmental fluctuations. During the extraction process, the system performs multi-dimensional analysis on sensor data at different time points, eliminating redundant information to ensure the representativeness and stability of the features. Subsequently, for missing values ​​in the key feature data, an algorithm combining time-series interpolation and multiple imputation is used for imputation. Specifically, the system first determines the missing pattern (random or non-random missing) and, based on complete data within the adjacent time window and the correlation of environmental parameters, achieves a reasonable estimate of the missing values. The imputation process strictly follows data consistency constraints to avoid introducing abnormal biases. The imputed data undergoes multiple rounds of iterative verification to ensure that its statistical characteristics are highly consistent with the distribution of the original data. Finally, the generated complete dataset serves as the output field of this step, containing multi-source fused data after feature extraction and missing value imputation. This "complete dataset" is invoked in subsequent step S130 for time series alignment and standardization.

[0098] S130. Perform time series alignment on the complete dataset to generate a standardized data matrix;

[0099] This step takes the "complete dataset" output by S120 as input and first performs time series alignment processing on the multi-source data. Specifically, to address the issues of inconsistent sampling frequencies and timestamp deviations among different sensor data, the system employs an algorithm combining Dynamic Time Warping (DTW) and linear interpolation to adjust the time axis of each data stream, achieving precise synchronization of multi-source data. This alignment process considers device clock drift and network transmission latency to ensure data time point consistency. Subsequently, the aligned data undergoes standardization processing, specifically using a Z-score-based standardization method to transform the data of each feature dimension into a standard normal distribution interval with a mean of zero and a variance of one, eliminating dimensional differences and improving the convergence speed and stability of subsequent model training. During the standardization process, the system dynamically updates the mean and variance parameters to adapt to the time-varying characteristics of raw coal properties. After processing, the generated standardized data matrix is ​​stored in the form of a multi-dimensional array, structurally representing the fused data features of the multi-source sensors. This output field, "Standardized Data Matrix," serves as the output of this step and is passed to the subsequent step S210 for the partitioning of the time series sample set and feature modeling. Meanwhile, the standardized data matrix provides a unified and high-quality data foundation for the subsequent S200, S300, S400, S500 and S600 modules, ensuring the coordinated operation of the entire intelligent dry separation dynamic optimization method and system.

[0100] In another embodiment:

[0101] The input source consists of granularity distribution, density characteristics, and environmental parameter data collected by a multi-source sensor system. After being transmitted to the data acquisition interface module via industrial Ethernet, a noise filtering algorithm based on improved wavelet transform is executed. Formula ① defines the improved wavelet threshold function:

[0102]

[0103] in:

[0104] T wave Improved wavelet threshold function;

[0105] x s This represents the raw signal of the s-th sensor channel;

[0106] σ s The standard deviation of the corresponding channel noise is calculated using a sliding window.

[0107] α 1 =3.5 and β 1 =2.2 is the adaptive shape parameter, obtained by training with historical data.

[0108] This function combines nonlinear attenuation of high-frequency noise with preservation of signal characteristics.

[0109] After processing using formula ①, an improved adaptive filtering algorithm is employed:

[0110] y t =γ t ·W t ·x t +(1-γ t )·y t-1

[0111] in:

[0112] y t The filtered output signal represents time t;

[0113] x t The original input signal at time t;

[0114] W t It is a time-varying filtering matrix, dynamically adjusted by ambient temperature (data source C) and vibration intensity (data source D);

[0115] γ t ∈[0,1] is the forgetting factor, calculated based on the signal-to-noise ratio;

[0116] y t-1 : Represents the filtered output signal at the previous time (t-1).

[0117] The output field "raw data stream" is consumed by the "raw data stream" of S120.

[0118] Step S120 takes the original data stream as input and uses an improved principal component analysis algorithm to extract key features. Formula ② defines the feature weight calculation:

[0119]

[0120] in:

[0121] The weight of the i-th feature in the p-th principal component;

[0122] The eigenvalue of the p-th principal component;

[0123] n represents the total number of feature dimensions;

[0124] maxH p This represents the maximum value of the feature information entropy in the p-th principal component;

[0125] H p (x i) is the information entropy of the i-th feature, calculated from the density distribution histogram;

[0126] ∈=1×10 -5 This is the smoothing coefficient.

[0127] This formula dynamically weights the principal component variance contribution rate with information entropy. Missing value imputation uses an improved spatiotemporal interpolation formula ③:

[0128]

[0129] in:

[0130] The result of imputing missing values ​​at time t;

[0131] The summation symbol indicates that the summation is performed over the time index k from the lower limit k = t - Δ to the upper limit k = t + Δ.

[0132] Δ is the size of the time window;

[0133] ω k As a time decay weight;

[0134] Observations;

[0135] For the ambient humidity gradient;

[0136] η = 0.2 is the coupling coefficient.

[0137] The output field "Complete Dataset" is consumed by "Complete Dataset" in S130.

[0138] After receiving the complete dataset, S130 uses an improved dynamic time warping algorithm for alignment. Formula ④ defines the synchronization error metric:

[0139]

[0140] in:

[0141] D sync : A measure of synchronization error between two time series;

[0142] D sync (X i ,X j ): Two time series X i and X j The synchronization error between them is measured;

[0143] X i (m),X j (n): Data from the i-th and j-th sensors at time points m and n;

[0144] A is the set of alignment paths;

[0145] Find the minimum value in set A;

[0146] The square of the L2 norm;

[0147] μ i and σ i Let be the mean and standard deviation of the data from the i-th sensor, respectively;

[0148] λ = 0.5 is the time offset penalty coefficient.

[0149] Furthermore, the standardization process employs an improved Z-score formula⑤:

[0150]

[0151] in:

[0152] Z k Standardized data;

[0153] x k : Raw data;

[0154] and For sliding window statistics;

[0155] δ = 0.01 is the stability factor.

[0156] The output field "Standardized Data Matrix" is consumed by the "Standardized Data Matrix" of S210.

[0157] Formula referencing relationships and data flow:

[0158] Formula ① (improved wavelet threshold function) processes the raw sensor signal and outputs the "raw data stream" as the input of S120.

[0159] Formula ② (feature weight calculation) extracts key features based on the "raw data stream" and combines it with Formula ③ (missing value imputation) to generate a "complete dataset" for use by S130.

[0160] Formula ④ (synchronization error metric) and Formula ⑤ (standardization) complete time series alignment and standardization based on the "complete dataset", outputting a "standardized data matrix" to provide input for subsequent modeling.

[0161] Specifically, Formula ① processes multi-source sensor signals to generate a "raw data stream," which serves as the input for Formula ②. Formula ② combines information entropy to calculate feature weights, assisting in the extraction of key features. Formula ③, in conjunction with Formula ③, imputes missing values, outputting a "complete dataset." Formula ④ calculates the time series synchronization error based on the "complete dataset," guiding sequence alignment. Formula ⑤ standardizes the aligned data, generating a "standardized data matrix." This matrix provides high-quality data input for subsequent model training or control algorithms.

[0162] Technical achievements in this section: By improving the wavelet threshold function and spatiotemporal interpolation algorithm, noise suppression and feature enhancement of multi-source data are achieved; combined with the synchronization error metric of dynamic time warping, a standardized data matrix is ​​constructed, providing high-quality input for time series modeling.

[0163] Step S200 includes at least steps S210-S230:

[0164] S210. Obtain the standardized data matrix, perform sliding window segmentation, and obtain the time series sample set;

[0165] This step takes the "standardized data matrix" output by S130 as input and first performs sliding window segmentation on the standardized data matrix. Specifically, the system gradually extracts multidimensional data subsets within continuous time periods according to preset time window length and sliding step size parameters to form a time-series sample set. The sliding window parameters can be dynamically adjusted according to the fluctuation characteristics of raw coal properties to adapt to the time-varying nature of data under different operating conditions. During the sliding window segmentation process, the system performs special processing on boundary data, using zero-filling or repeated-filling strategies to ensure the consistency of sample length. Furthermore, for potentially abnormal window samples, the system removes or marks them based on statistical thresholds and anomaly detection algorithms to prevent abnormal data from affecting subsequent model training and prediction. After the sliding window segmentation is completed, the generated time-series sample set is stored in the form of a multidimensional array, fully reflecting the continuous change characteristics of multi-source fused data in the time dimension. This output field "time-series sample set" is passed to the subsequent step S220 as the output of this step for density distribution feature extraction and normalization processing. Meanwhile, the time-series sample set provides basic data support for the construction of feature vectors in the subsequent S200 module and the prediction of LSTM parameters in the S300 module, ensuring the consistency and accuracy of the time-series feature modeling of the entire intelligent dry selection dynamic optimization method and system.

[0166] S220. Extract density distribution features from the time series sample set, perform normalization processing, and generate a feature vector set;

[0167] This step takes the "time-series sample set" output by S210 as input and first extracts density distribution features from the data of each time window in the time-series sample set. Specifically, based on the raw coal particle size and density measurement data, the system uses the Kernel Density Estimation (KDE) method to perform non-parametric estimation of the particle size and density data distribution within the sample, obtaining a continuous and smooth probability density function. The kernel function type and bandwidth parameters are dynamically optimized through cross-validation to achieve accurate fitting of the density distribution under different raw coal characteristics. Further, the system combines multi-source sensor data to extract statistical features describing the peak value, distribution width, skewness, and kurtosis of the particle size distribution, comprehensively characterizing the temporal changes in the physical properties of raw coal. During the extraction process, the system identifies and corrects abnormal peak values ​​to prevent noise interference from causing feature distortion. Subsequently, for the density distribution features, the system performs normalization processing, using the Min-Max Normalization method to map the feature values ​​to a predefined standard interval, eliminating differences in the dimensions of different features and improving the consistency of subsequent model inputs. The normalization parameters are dynamically updated based on historical data to adapt to the long-term changing trends of raw coal properties. After normalization, the generated feature vector set is organized in matrix form, containing multi-dimensional density distribution features, fully reflecting the temporal evolution of raw coal properties. This output field, "feature vector set," is passed to the subsequent step S230 for time dimension expansion and three-dimensional feature tensor generation. Simultaneously, the feature vector set provides accurate and stable feature representation for the LSTM network input of module S300, supporting the accuracy and robustness of parameter prediction.

[0168] S230. Extend the feature vector set in the time dimension to generate a three-dimensional feature tensor;

[0169] This step takes the "feature vector set" output by S220 as input and first performs time dimension expansion processing on the feature vector set to generate a three-dimensional feature tensor. Specifically, according to the requirements of temporal feature modeling, the system stacks the two-dimensional feature matrices according to the time step to form a three-dimensional array structure containing three dimensions: time step, feature dimension, and sample batch. This three-dimensional feature tensor structurally expresses the temporal dynamic features of multi-source fused data, meeting the modeling requirements of subsequent Long Short-Term Memory (LSTM) networks for temporal dependencies. Furthermore, during the generation of the three-dimensional feature tensor, the system performs correlation analysis between features within the time window, adjusts the data arrangement order within the tensor, and optimizes the continuity and timeliness of temporal information. To ensure data integrity, the system interpolates and fills in missing time steps using a time-weighted linear interpolation method to ensure the continuity and smoothness of the tensor data. In addition, the system performs data integrity verification on the generated three-dimensional feature tensor, using a checksum and data version control mechanism to prevent damage and errors during data transmission. The generated 3D feature tensor is stored in an efficient data format, supporting fast reading and parallel computation. This output field, "3D Feature Tensor," is passed to the subsequent step S310 for forward propagation and parameter prediction in the LSTM network. The 3D feature tensor also provides a unified and high-quality temporal feature input for modules S300, S400, and the subsequent S500, ensuring the accuracy and continuity of the temporal feature modeling stage of the entire intelligent dynamic optimization method and system.

[0170] Step S300 includes at least steps S310-S330:

[0171] S310. Obtain the three-dimensional feature tensor, perform forward propagation operation of LSTM network to obtain the initial prediction value;

[0172] This step takes the "three-dimensional feature tensor" output by S230 as input. Specifically, the three-dimensional feature tensor is used as input data for a Long Short-Term Memory (LSTM) network model for forward propagation to achieve initial prediction of key parameters in the dry coal preparation process. The three-dimensional feature tensor includes three dimensions: time step, feature dimension, and sample batch, comprehensively reflecting the temporal dynamic characteristics of raw coal properties. First, the system loads the current model parameter weights and bias configurations according to the pre-trained LSTM model structure to ensure that the model state is consistent with the historical training results. Subsequently, the input three-dimensional feature tensor is fed into the LSTM unit step by step. The LSTM unit filters and memorizes temporal information through gating mechanisms (including forget gate, input gate, and output gate), dynamically capturing the dependencies of raw coal properties over time. Specifically, the forget gate determines which historical information to retain or forget based on the current input and the hidden state of the previous time step; the input gate controls the writing of the current input information; and the output gate determines the hidden state of the output, forming a dynamic expression of temporal features. During the forward propagation process, a bidirectional transmission mechanism is employed between network layers to enhance the ability to capture temporal context information. Furthermore, for outliers and missing time steps in the input data, the system automatically adjusts its internal state using state update rules based on the gating mechanism, avoiding the negative impact of outlier data on the prediction results. To ensure computational efficiency, the forward propagation process is executed in parallel on an edge computing platform supporting GPU acceleration, shortening the model response time. After completing the forward propagation for all time steps, the system maps the output hidden state to the target parameter space through a fully connected layer, generating the corresponding initial prediction value. This initial prediction value serves as the output field of this step and is passed to the subsequent step S320 for key parameter extraction and regularization correction. Simultaneously, the initial prediction value provides the basis for prediction in the subsequent parameter proposal set generation in module S300 and the genetic algorithm optimization in module S400, ensuring the prediction accuracy and timeliness of the intelligent dynamic optimization method and system.

[0173] S320. Extract key parameters from the initial predicted values, perform regularization corrections, and generate a parameter suggestion set;

[0174] This step takes the "initial predicted value" output by S310 as input. Specifically, it extracts a set of key parameters related to the dry separation parameters from the initial predicted value and performs regularization correction to generate a parameter suggestion set that conforms to the system's operating constraints. First, the system selects parameters that significantly affect the performance of the dry separation equipment, such as wind speed, vibration frequency, separation angle, and feeding speed, based on preset parameter mapping rules, and constructs a subset of key parameters. Subsequently, for these key parameters, a regularization method based on the L2 norm is used for correction to suppress overfitting and smooth parameter fluctuations, thereby improving the stability of the prediction results. During the regularization process, the system dynamically adjusts the regularization coefficient based on historical operating data to ensure that the parameter suggestion set conforms to actual working conditions while avoiding the occurrence of extreme values. Furthermore, the system combines equipment safety boundaries and process constraints to perform range trimming on the regularized parameters to prevent exceeding the equipment's carrying capacity or the process's allowable range. This trimming process dynamically updates the upper and lower limits of parameters based on the technical specifications provided by the equipment manufacturer and on-site feedback data to adapt to different raw coal properties and environmental conditions. To ensure the rationality of the parameter recommendations, the system performs multiple rounds of iterative verification. In each iteration, the corrected parameters are input into the simulation model to evaluate their impact on dry separation efficiency and clean coal recovery rate. The regularization strategy is adjusted based on the feedback. Abnormal parameters are automatically marked and recorded in the log system for subsequent manual intervention and model optimization. Finally, the generated parameter recommendation set contains regularized corrected dry separation parameter configurations that comply with safety and process constraints. This parameter recommendation set serves as the output field of this step and is passed to subsequent step S330 for feasibility verification and the construction of the effective parameter space. Simultaneously, the parameter recommendation set provides an initial solution space for the genetic algorithm optimization in module S400, supporting the collaborative optimization process between the intelligent dry separation dynamic optimization method and the system.

[0175] S330. Perform feasibility verification on the parameter suggestion set and generate an effective parameter space;

[0176] This step takes the "parameter suggestion set" output by S320 as input. Specifically, it verifies the feasibility of the parameter suggestion set to generate an effective parameter space that meets the system's operational requirements. First, based on a multi-dimensional constraint model, the system checks the constraint satisfaction of each parameter combination in the parameter suggestion set. The constraints cover equipment physical limits, process requirements, safety standards, and environmental adaptability indicators. The constraint detection uses a method combining constraint programming (CP) and heuristic search to quickly filter out parameter combinations that satisfy all constraints. Further, the system combines historical operating data and real-time monitoring information, using fuzzy logic reasoning to score the applicability of parameter combinations and eliminate unstable or potentially high-risk parameter configurations. For edge cases, the system triggers a manual review mechanism, recording the relevant parameter combinations and their risk assessment results for expert judgment. Understandably, the system performs multi-dimensional coverage analysis on the effective parameter space to ensure that parameter combinations have good adaptability under various key operating conditions. To prevent the parameter space from being too sparse or too dense, the system dynamically adjusts the sampling density based on the parameter distribution characteristics, employing the Latin Hypercube Sampling (LHS) method to generate a uniformly distributed subset of candidate parameters, thus improving the search efficiency of subsequent genetic algorithm optimization. The system also performs version control and data integrity verification on the effective parameter space to ensure the consistency and traceability of the parameter data. After the above processing, the generated effective parameter space is stored in a structured data format and used as the output field of this step, passed to subsequent step S410 for population initialization in the genetic algorithm. This effective parameter space provides a scientifically reasonable search range for the genetic algorithm optimization in module S400, supporting the intelligent dynamic optimization method and the system's global optimal solution exploration.

[0177] Step S400 includes at least steps S410-S430:

[0178] S410. Obtain the effective parameter space, perform population initialization, and generate a candidate parameter set;

[0179] This step takes the "effective parameter space" output by S330 as input. Specifically, it first extracts a multi-dimensional data set of parameter combinations from the effective parameter space as the initial solution set for genetic algorithm optimization. The effective parameter space contains a subset of parameters that have undergone rigorous constraint detection and sampling optimization to ensure that they meet the physical limits of the equipment, the process flow, and safety standards. Based on the preset population size parameters and the distribution characteristics of the effective parameter space, the system initializes the population of the genetic algorithm using a strategy combining random sampling and stratified sampling. Specifically, the system uses Latin Hypercube Sampling (LHS) to uniformly cover the parameter space, avoiding local clustering during population initialization and improving search diversity. Furthermore, during initialization, a hybrid initialization mechanism is used, incorporating historically optimal parameter combinations, to set some individuals as empirically optimal solutions to accelerate convergence. During initialization, the system uses real-number encoding for the parameters of each individual to ensure the continuity and accuracy of the genetic operation. The individual encoding includes key sorting parameters such as wind speed, vibration frequency, sorting angle, and feeding speed, and the encoding structure is compatible with subsequent genetic operators. To ensure the quality of the initial population, the system performs individual validity checks, removing individuals that violate equipment safety boundaries or process constraints. This information is recorded in the log system for subsequent analysis and optimization. After initialization, the generated candidate parameter set is passed as the output field of this step to the subsequent step S420 for parallel fitness calculation. This candidate parameter set provides a diverse search basis for the genetic algorithm iterations of modules S420 and S430, supporting intelligent dynamic optimization methods and the system's global optimization process.

[0180] S420. Extract individuals from the candidate parameter set, perform parallel fitness calculations, and generate a scoring matrix;

[0181] This step takes the "candidate parameter set" output by S410 as input. Specifically, it extracts each individual parameter combination from the candidate parameter set and performs parallel calculation of the fitness function to evaluate the performance of each individual in the dry separation process. First, the system constructs a fitness evaluation model based on a predefined multi-objective fitness function, comprehensively considering clean coal recovery rate, energy consumption indicators, and equipment operation safety. The fitness function combines experimental data-driven models and simulation models to accurately reflect the impact of parameter combinations on the dry separation effect. Further, the system allocates fitness calculation tasks to edge computing nodes or cloud computing platforms that support parallel computing, and adopts a multi-threaded and distributed computing framework to achieve efficient evaluation of the candidate parameter set. Specifically, each computing node independently executes the simulation model, inputs individual parameters from the candidate parameter set, simulates material sorting behavior and equipment response in the dry separation process, and outputs corresponding performance indicators. During the calculation process, the system dynamically monitors the computing load and node status, automatically adjusts the task allocation strategy, and ensures optimal utilization of computing resources. In case of calculation anomalies or timeouts, the system triggers a fault tolerance mechanism, reschedules the task, or uses backup nodes to ensure the integrity and accuracy of the fitness calculation. After completing the fitness assessment of all individuals, the system summarizes the results from each computation node to form a scoring matrix, where each element corresponds to the overall fitness value of an individual. The scoring matrix is ​​normalized to eliminate differences in the dimensions of different indicators, improving the selection efficiency of subsequent genetic operations. The scoring data is recorded in a database, supporting historical performance analysis and model iterative optimization. This scoring matrix serves as the output field of this step and is passed to subsequent step S430 for elite selection and genetic operations. The scoring matrix provides an objective evaluation basis for the generation of optimization parameter combinations in module S430, ensuring the scientific and real-time nature of the optimization decisions of the intelligent dynamic selection optimization method and system.

[0182] S430. Perform elite selection and cross-mutation on the scoring matrix to generate an optimal parameter combination;

[0183] This step takes the "scoring matrix" output by S420 as input. Specifically, it performs elite selection, crossover, and mutation operations on the scoring matrix to generate a new generation of optimized parameter combinations, realizing the iterative evolution of the genetic algorithm. First, the system sorts the individuals in the candidate parameter set according to their fitness values ​​and adopts an elitism strategy, directly copying the top-ranked individuals to the next generation to ensure the inheritance of high-quality solutions. Then, the system randomly selects parent individuals from the remaining individuals using the fitness proportionate selection method and performs a crossover operation. Specifically, the crossover operation uses a simulated binary crossover (SBX) method with real-number encoding, combined with dynamic crossover probabilities, to generate diverse and constrained offspring individuals. During the crossover process, the system performs boundary checks on the generated offspring parameter combinations to prevent parameters from exceeding equipment and process limitations. Further, the system performs mutation operations on the crossover offspring individuals, employing a Gaussian mutation strategy to dynamically adjust the mutation amplitude to balance exploration and utilization. The mutation process incorporates an adaptive mutation rate adjustment mechanism, automatically adjusting based on the number of iterations and population diversity to avoid premature convergence. After mutation, the system performs constraint satisfaction checks on all newly generated individuals, eliminating or correcting parameter combinations that do not meet safety boundaries to ensure the safety and practicality of the optimization process. To enhance the algorithm's global search capability, the system periodically introduces random new individuals to prevent getting trapped in local optima. After completing the genetic operation, the system updates the population structure, forming a new generation of candidate parameter sets containing elite individuals and newly generated individuals. This new generation of parameter sets is recorded in the version control system, supporting algorithm backtracking and performance analysis. Finally, the generated optimized parameter combination is used as the output field of this step and passed to the subsequent step S510 for dynamic parameter deployment. The optimized parameter combination provides high-quality control command input to the S500 module, ensuring the accuracy and response speed of parameter adjustment for the intelligent dynamic optimization method and system.

[0184] In another embodiment:

[0185] In S410, the population is initialized using an improved Latin hypercube sampling method, with the effective parameter space as input. Equation ⑥ defines multidimensional parameter sampling:

[0186]

[0187] in:

[0188] The d-th dimension parameter value for the i-th individual is derived from the wind speed (data source G) and vibration frequency (data source H) in the effective parameter space.

[0189] πd It is a randomly arranged sequence of integers;

[0190] Uniformly distributed random numbers;

[0191] upper and lower bounds of parameters;

[0192] N: Population size.

[0193] The output field "candidate parameter set" is consumed by the "candidate parameter set" of S420.

[0194] Furthermore, parallel fitness calculations are performed on the candidate parameter set in S420. Equation ⑦ defines the multi-objective fitness function:

[0195]

[0196] in:

[0197] F i : The multi-objective fitness value of the i-th individual;

[0198] R i For clean coal recovery rate;

[0199] E i Energy consumption indicators;

[0200] Let be the deviation of the j-th safety constraint;

[0201] The j-th safety constraint limit;

[0202] α 2 =0.6, β 2 =0.3, γ 2 =0.1 is the weighting coefficient;

[0203] R min ,R max E min E max The minimum and maximum values ​​of clean coal recovery rate and energy consumption indicators in historical data.

[0204] The output field "Rating Matrix" is consumed by the "Rating Matrix" of S430.

[0205] Furthermore, an improved genetic operation is performed on the scoring matrix in S430. Equation ⑧ defines the simulated binary crossover:

[0206]

[0207] in:

[0208] The parameter value of the d-th dimension of the k-th offspring after crossover;

[0209] The parameter values ​​corresponding to the two parent individuals;

[0210] β q : These are the control parameters for cross-operation, calculated as follows η c =2.5 is the distribution index;

[0211] u is a random variable uniformly distributed in [0,1].

[0212] Formula 9 defines adaptive Gaussian mutation:

[0213]

[0214] in:

[0215] The new value of the d-th dimension parameter of the k-th individual after mutation;

[0216] The original value of the d-th dimension parameter of the k-th individual before mutation;

[0217] σ d The magnitude of variation in parameter dimensions;

[0218] Standard normal distribution random number generator;

[0219] g is the current iteration number;

[0220] G max =100 is the maximum number of iterations.

[0221] The output field "Optimized parameter combination" is consumed by the "Optimized parameter combination" of S510.

[0222] Explanation of formula dependencies and transitive relationships:

[0223] Formula ⑥ (Latin hypercube sampling) generates a "candidate parameter set" in S410 for fitness calculation in S420.

[0224] Formula ⑦ (multi-objective fitness function) calculates fitness based on the "candidate parameter set" and outputs a "score matrix" for S430 genetic operations.

[0225] Formulas ⑧ and ⑨ (genetic operators) select and mutate the "score matrix" to generate "optimized parameter combinations" and complete the parameter optimization closed loop.

[0226] Specifically, formula ⑥ samples the effective parameter space to generate a "candidate parameter set"; formula ⑦ performs multi-objective fitness evaluation on the candidate parameters to form a "score matrix"; formulas ⑧ and ⑨ use a genetic algorithm to perform crossover and mutation operations on the score matrix to generate an "optimized parameter combination", thus completing the parameter optimization process and achieving system performance improvement and security constraint satisfaction.

[0227] Technical benefits in this section: By improving Latin hypercube sampling and multi-objective fitness functions, global exploration of the parameter space is achieved; combined with adaptive genetic operators, the safety and process feasibility of optimizing parameter combinations are ensured.

[0228] Step 500 includes at least steps S510-S530:

[0229] S510: Obtain the optimized parameter combination, perform equipment instruction conversion, and generate a control instruction set;

[0230] This step takes the "optimized parameter combination" output by S430 as input. Specifically, it obtains the optimized parameter combination from the genetic algorithm optimization module. This optimized parameter combination includes multi-dimensional sorting parameter configurations generated through elite selection, crossover, and mutation operations. First, the system receives the optimized parameter combination and performs equipment instruction conversion processing on key parameters in each parameter combination, such as wind speed, vibration frequency, sorting angle, and feeding speed. Specifically, the system maps the optimized parameter combination to a control instruction format recognizable by the equipment according to the equipment control protocol and instruction set specifications. During the mapping process, the system calls a preset parameter mapping table and conversion function to ensure that the parameter values ​​are consistent with the equipment input range and unit standards. Furthermore, the system performs format verification on the mapped control instructions, including instruction code legality, parameter boundary compliance, and check code integrity, to prevent equipment malfunctions due to incorrect instruction formats. For multi-device collaborative control scenarios, the system performs format verification based on the equipment topology. The system generates a distributed control instruction set based on communication protocols, supporting parallel issuance and synchronous execution. After the control instruction set is generated, the system encapsulates the instruction data into a unified instruction packet format, supporting transmission via industrial Ethernet, fieldbus, or wireless network. To ensure the reliability of instruction transmission, the system adds a sequence number, timestamp, and checksum to the instruction packet, along with a confirmation and retransmission mechanism, to ensure complete delivery of the instruction. The entire device instruction conversion process is executed on edge computing nodes, adjusting the instruction generation strategy based on real-time monitoring feedback. Abnormal instructions or conversion failure events are recorded in detail in the log system for subsequent fault diagnosis and optimization. After completing the above processing, the generated control instruction set is passed as the output field of this step to the subsequent step S520 for safety boundary verification and post-verification instruction generation. The control instruction set also provides the control input basis for the dry separation process monitoring and energy efficiency assessment of the S600 module, ensuring the accuracy and real-time performance of the intelligent dry separation dynamic optimization method and system instruction execution.

[0231] S520: Extract key parameters from the control instruction set, perform safety boundary verification, and generate post-verification instructions;

[0232] This step takes the "control instruction set" output by the S510 as input. Specifically, it extracts key parameters from the control instruction set, performs safety boundary verification, and generates verified instructions that comply with equipment operation safety specifications. First, the system parses the control instruction set, identifies the various sorting parameters and their corresponding values, and loads the corresponding safety boundary parameter set based on the technical specifications and operation manual provided by the equipment manufacturer. The safety boundary parameter set covers key indicators such as upper and lower limits of wind speed, extreme values ​​of vibration frequency, sorting angle range, and feeding speed limits. The system performs boundary detection on each parameter in the control instructions, using a rule-based judgment algorithm to determine whether the parameter exceeds the safety range. For detected out-of-bounds parameters, the system automatically executes a boundary correction strategy to adjust the parameter value to the closest safety boundary value, ensuring equipment operation safety. During the correction process, the system records the parameter differences before and after correction and the reasons for correction, writing them to the safety verification log to support subsequent auditing and analysis. Furthermore, the system combines real-time equipment status monitoring data to perform a dynamic safety assessment, determining the applicability of the current parameter combination in the actual operating environment. The dynamic safety assessment employs a fuzzy logic-based reasoning model, comprehensively considering factors such as equipment temperature, vibration intensity, and ambient humidity to achieve real-time determination of parameter safety. For parameter combinations that fail the dynamic safety assessment, the system triggers an alarm mechanism, suspends the command issuance process, and feeds back the abnormal information to the operation terminal, supporting manual intervention. In the post-verification command generation stage, the system reconstructs control commands based on the corrected parameters, ensuring the integrity and consistency of the command format and content. The post-verification commands undergo multiple verifications, including parameter consistency checks, command format verification, and security policy compliance confirmation, ensuring that the commands meet all safety specifications. After completing the safety boundary verification and post-verification command generation, the system uses the generated post-verification commands as the output field of this step, passing them to the subsequent step S530 for vibration-resistant and temperature-resistant packaging and execution command package generation. The post-verification commands provide a safe and compliant control basis for the execution status monitoring of the S600 module, ensuring the safe and stable operation of the intelligent dry-selection dynamic optimization method and system.

[0233] S530: The verified instructions are encapsulated in a vibration-resistant and temperature-resistant package to generate an execution instruction package;

[0234] This step takes the "verified instruction" output by the S520 as input. Specifically, it performs vibration-resistant and temperature-resistant encapsulation processing on the verified instruction to generate an execution instruction package suitable for the field execution environment. First, the system loads vibration-resistant encapsulation and temperature control scheme parameters based on the characteristics of the field hardware environment. These scheme parameters cover vibration-resistant material type, encapsulation structure design, thermal conductivity, and temperature regulation strategy. For the verified instruction, the system embeds the instruction data into a dedicated encapsulation protocol, employing a multi-layer data protection mechanism, including error detection codes (such as CRC check codes), data encryption, and redundant coding, to improve the anti-interference capability and security of the instruction package. Furthermore, based on the equipment's vibration frequency characteristics and amplitude range, combined with field vibration monitoring data, the system designs and applies a dynamic vibration-resistant filtering algorithm to reduce the instruction transmission error rate caused by mechanical vibration. Regarding temperature control, the system monitors the equipment's operating environment temperature in real time, and dynamically adjusts the thermal conductivity parameters of the encapsulation material and the internal heat dissipation mechanism based on temperature sensor data to ensure the stability of the instruction package in high or low temperature environments. The vibration-resistant and temperature-resistant encapsulation process adopts a modular design, supporting rapid disassembly and maintenance while ensuring sealing performance and electromagnetic compatibility. During the encapsulation process, the system implements version control and integrity verification for the instruction packets, employing digital signature technology to prevent data tampering and ensure the authenticity and integrity of the instruction packets. After encapsulation, the generated execution instruction packets are stored in a standardized data format, supporting rapid parsing and execution by field devices. The execution instruction packets are sent to the dry separation equipment control unit via industrial fieldbus or wireless network, supporting real-time control and feedback. The system monitors the transmission status of the instruction packets and, combined with an acknowledgment and response mechanism, achieves reliable instruction issuance and execution tracking. Abnormal transmission or execution failure events are automatically recorded and trigger a retransmission mechanism to ensure the continuity and accuracy of instruction execution. This execution instruction packet, as an output field of this step, is passed to the subsequent step S610 for dry separation process monitoring and operation status log generation. The execution instruction packet provides the execution basis for the closed-loop optimization feedback of the S600 module, ensuring the stability and efficiency of the intelligent dry separation dynamic optimization method and the field deployment and execution of the system.

[0235] Step S600 includes at least steps S610-S630:

[0236] S610: Obtain the execution instruction package, monitor the dry selection process, and generate an operation status log;

[0237] This step takes the "execution command packet" output by the S530 as input. Specifically, it obtains the execution command packet after vibration-resistant and temperature-resistant encapsulation. The execution command packet contains multi-dimensional control commands for the dry separation equipment in the coal preparation plant and its encapsulation information. First, the system receives the execution command packet through an industrial fieldbus or wireless communication module. The communication module supports real-time data transmission and a two-way confirmation mechanism to ensure the integrity and timeliness of the command packet. Specifically, the system parses the received execution command packet to extract the encapsulated control command data and its related metadata, including the serial number, timestamp, and checksum. Subsequently, the system matches the version information in the command packet with the current firmware version of the equipment to confirm the compatibility of the command packet and prevent execution anomalies caused by version inconsistencies. After confirming compatibility, the system decrypts the command data and verifies the error detection code to ensure that the data has not been tampered with and there are no transmission errors. After unpacking, the control commands are transmitted to the control unit of the dry separation equipment. Specifically, the control unit adjusts key parameters of the dry separation equipment, such as wind speed, vibration frequency, sorting angle, and feeding speed, according to the command content. During execution, the system monitors the equipment status in real time, collecting information such as vibration intensity, temperature, and operating load through built-in sensors to form a continuous operating status data stream. This operating status data undergoes preprocessing at edge computing nodes, including anomaly detection and data compression, ensuring the accuracy and transmission efficiency of the monitoring data. Further, the system generates an operating status log according to a preset format, covering equipment response time, parameter execution deviations, and abnormal event records. The operating status log is periodically uploaded to the central monitoring platform, supporting remote diagnostics and maintenance decisions. Abnormal status triggering mechanisms include equipment over-limit alarms, communication interruption detection, and command execution failure retransmission; related events are recorded in detail and operators are notified. After the above processing, the generated operating status log serves as the output field of this step and is passed to subsequent step S620 for energy efficiency index extraction and recovery rate calculation. The operating status log also provides basic data support for the closed-loop optimization feedback of the S600 module, ensuring the real-time monitoring and dynamic adjustment capabilities of the intelligent dry separation dynamic optimization method and system.

[0238] S620. Extract energy efficiency indicators from the operation status log, calculate the recovery rate, and generate an evaluation report;

[0239] This step takes the "Operating Status Log" output by S610 as input. Specifically, it extracts energy efficiency indicators from the operating status log, calculates the recovery rate, and generates an evaluation report. First, the system performs structured analysis on the equipment parameter execution data and environmental monitoring data contained in the operating status log to extract key energy efficiency indicators, including but not limited to clean coal recovery rate, unit energy consumption, equipment load rate, and failure rate. The energy efficiency indicator extraction employs a method combining a rule engine and machine learning. The rule engine defines the indicator calculation formula based on industry standards and equipment technical specifications, while the machine learning model dynamically adjusts weights and thresholds based on historical operating data to improve the accuracy and adaptability of indicator calculation. Further, the system utilizes the collected raw coal particle size and density characteristic data, combined with operating parameters, and employs a combination of statistical analysis and simulation models to calculate the clean coal recovery rate. This calculation process considers material sorting efficiency, equipment operating status, and environmental factors to ensure the comprehensiveness of the recovery rate evaluation. For the calculation of unit energy consumption, the system obtains real-time power data from energy consumption monitoring sensors, combines it with equipment operating time and load conditions, calculates the total energy consumption using an integral method, and obtains the unit energy consumption indicator through normalization. The calculated results of the energy efficiency indicators are dynamically updated, supporting real-time feedback and historical trend analysis. Subsequently, based on the energy efficiency indicators, combined with equipment health status and production plans, the system generates a comprehensive evaluation report. The evaluation report includes the current values ​​of the energy efficiency indicators, historical trends, abnormal fluctuation analysis, and explanations of influencing factors. During report generation, the system employs a template engine and data visualization technology, combining quantitative indicators with charts to improve readability and decision support capabilities. To ensure the accuracy of the evaluation report, the system performs multi-level checks on the input data, including integrity checks, outlier removal, and data consistency verification. Abnormal data is automatically marked and recorded in the log system. After the evaluation report is completed, the system stores it in a central database and pushes it to subsequent step S630 via a message queue mechanism for difference analysis and optimization suggestion generation. The evaluation report also provides decision-making basis for the closed-loop optimization feedback of the S600 module, supporting continuous improvement and performance enhancement of the intelligent dry-selection dynamic optimization method and system.

[0240] S630. Perform a difference analysis on the evaluation report and generate optimization suggestion parameters;

[0241] This step takes the "assessment report" output by S620 as input. Specifically, it performs a difference analysis on the assessment report and generates optimized parameter suggestions. First, the system compares the energy efficiency indicators in the current assessment report with historical benchmark data, calculating the differences and trend deviations. The difference analysis uses time series analysis and multivariate statistics, combined with anomaly detection algorithms to identify the root causes of performance fluctuations. Further, based on the difference analysis results, combined with equipment operating status and environmental parameters, the system calls a preset optimization rule base and a machine learning optimization model to generate adjustment suggestions for dry separation parameters. The optimization rule base covers parameter control methods such as wind speed adjustment strategies, vibration frequency optimization schemes, sorting angle fine-tuning, and feeding speed matching. The rules are based on industry experience and equipment manufacturer technical specifications. The machine learning optimization model is trained on historical operating data and has adaptive capabilities, automatically recommending optimal parameter combinations based on changes in raw coal properties. When generating optimized parameter suggestions, the system considers equipment safety boundaries and process constraints to ensure the feasibility and safety of the suggested parameters. Furthermore, the system performs multiple rounds of simulation verification on the optimized suggested parameters, simulating the operational effect of the adjusted dry separation process, evaluating the improvement in clean coal recovery rate and energy consumption changes, and ensuring the practicality and effectiveness of the suggestions. During the simulation, the system dynamically adjusts the parameter weights to optimize the stability and response speed of the suggested parameters. For suggested parameters with potential risks, the system automatically marks them and triggers a manual review process to ensure the safety and controllability of the optimization process. After generating the optimized suggested parameters, the system outputs them in structured data form and transmits them to the "Optimized Suggested Parameters" in step S110, realizing closed-loop feedback. The optimized suggested parameters provide an adjustment basis for the multi-source data fusion of module S100, supporting the continuous self-adaptation and performance improvement of the intelligent dry separation dynamic optimization method and system.

[0242] Example 2: Figure 2 A structural block diagram of an intelligent dry separation dynamic optimization system for a coal preparation plant according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0243] Installation unit 01 is used to acquire installation parameters for the dry separation equipment and complete installation constraints. Specifically, it receives an installation parameter file provided by the equipment manufacturer, and under the configured installation constraints, completes parameter parsing, equipment calibration, and confirmation of the installation location to form an installation configuration file. The installation configuration file records the installation parameters and maintains a consistent association with the equipment status. The installation parameters are transmitted to the data acquisition unit for use as initialization parameters, while the installation log is retained for subsequent traceability.

[0244] Data acquisition unit 02 is used to collect raw coal property data and environmental parameters, and perform noise filtering within a preset monitoring range to obtain a raw data stream. Specifically, it receives installation parameters output from the installation unit and raw coal property data collected in real time by the sensor, performs data preprocessing, noise filtering, and effective range verification on the raw coal property data, and forms a raw data stream based on the monitoring range; the raw data stream is then transmitted as input data to the data processing unit for retrieval, and the corresponding timestamp information is registered in the data buffer for the data processing unit to read.

[0245] Data processing unit 03 is used to extract key features from the raw data stream, imputate missing values, generate a complete dataset, and align the complete dataset to a time series to generate a standardized data matrix. Specifically, it receives the raw data stream from the data acquisition unit, performs feature extraction and missing value imputation based on preset feature extraction rules, and generates a complete dataset when the data integrity condition is met; it aligns the complete dataset to a time series to generate a standardized data matrix, which is then used as input data by the judgment and arbitration unit, while recording the correspondence between the matrix and the processing strategy in the repository.

[0246] The judgment and arbitration unit 04 is used to perform event judgment and arbitration based on a standardized data matrix and a judgment threshold, and output control commands. Specifically, it completes event judgment and arbitration based on the standardized data matrix from the data processing unit, generates control commands, outputs the control commands to the execution unit, and sends the arbitration results back to the recording and updating unit for registration.

[0247] Execution unit 05 is used to output the control command to the execution component of the dry separation equipment to complete dynamic adjustment control. Specifically, it receives the control command from the judgment and arbitration unit, performs command parsing and dynamic control signal generation, generates an execution signal, outputs the execution signal to the execution component of the dry separation equipment, and sends the control effective status back to the recording and updating unit for registration.

[0248] Recording and updating unit 06 is used to record the effective time information of control commands and update the parameter adjustment strategy table. Specifically, it receives the control effective status from the execution unit and the arbitration result from the decision and arbitration unit, performs effective time recording, strategy table updating and log archiving, and obtains the updated parameter adjustment strategy table; the updated parameter adjustment strategy table is provided to the data acquisition unit as the basis for parameter adjustment, and maintains an index relationship consistent with the time series.

Claims

1. A dynamic optimization method for intelligent dry coal preparation in a coal preparation plant, characterized in that, include: The particle size distribution, density characteristics, and environmental parameters of raw coal are obtained, and noise filtering, missing value imputation, and time series alignment are performed to generate a standardized data matrix. The environmental parameters include temperature, humidity, and vibration intensity data; The noise filtering process includes: using a combination of wavelet transform and adaptive filtering algorithms to identify and remove high-frequency noise and abnormal spikes in the signal; dynamically adjusting the filtering parameters according to a preset threshold; and marking and recording abnormal data points in the log system. Obtain a standardized data matrix, perform sliding window segmentation, normalization, and time dimension expansion to generate a feature vector set and a three-dimensional feature tensor; Based on the three-dimensional feature tensor, perform forward propagation operation and regularization correction of LSTM network to generate parameter proposal set and perform feasibility verification to generate effective parameter space; To obtain an effective parameter space, the population is initialized, parallel fitness calculation is performed, elite selection and crossover mutation are carried out, and an optimized parameter combination is generated. The process of generating the optimized parameter combination includes: adopting an adaptive mutation rate adjustment mechanism to automatically adjust the mutation amplitude according to the number of iterations and population diversity; periodically introducing random new individuals to prevent getting trapped in local optima; after completing the genetic operation, updating the population structure to form a new generation of candidate parameter set containing elite individuals and newly generated individuals. The generated optimized parameter combination specifically includes: Each individual parameter combination is extracted from the candidate parameter set, and the fitness function is calculated in parallel to evaluate the performance of each individual in the dry separation process. Based on a predefined multi-objective fitness function, and taking into account the clean coal recovery rate, energy consumption index and equipment operation safety, a fitness evaluation model is constructed. The fitness calculation task is assigned to edge computing nodes or cloud computing platforms that support parallel computing, and a multi-threaded and distributed computing framework is adopted to achieve efficient evaluation of the candidate parameter set. Each computing node independently executes the simulation model, inputs individual parameters from the candidate parameter set, simulates the material sorting behavior and equipment response during the dry sorting process, and outputs the corresponding performance indicators. After completing the fitness assessment of all individuals, the results of each calculation node are summarized to form a scoring matrix, in which each element corresponds to the overall fitness value of an individual. Obtain optimized parameter combinations, perform device instruction conversion, safety boundary verification, and vibration- and temperature-resistant encapsulation, and generate execution instruction packages; Receive execution instruction packets, monitor the dry separation process, extract energy efficiency indicators for evaluation and recovery rate calculation, and generate optimization suggestion parameters.

2. The intelligent dry coal preparation plant dynamic optimization method according to claim 1, characterized in that, The process of generating a standardized data matrix includes: Key features were extracted from multi-source data. A method combining principal component analysis and information entropy assessment was used to automatically screen out key indicators that reflect the trend of raw coal particle size change, density distribution characteristics and environmental fluctuations, and to remove redundant information. For missing values ​​in key feature data, an algorithm combining temporal interpolation and multiple interpolation is used for filling. First, the missing pattern is determined, and then, based on the complete data in the adjacent time window and the correlation of environmental parameters, a reasonable estimate of the missing values ​​is achieved. The completed data underwent multiple rounds of iterative verification to ensure that its statistical characteristics were highly consistent with the distribution of the original data, thus generating a complete dataset.

3. The intelligent dry separation dynamic optimization method for coal preparation plants according to claim 1, characterized in that, Generating a standardized data matrix also includes: Time series alignment processing is performed on multi-source data. An algorithm based on dynamic time warping and linear interpolation is used to adjust the time axis of each data stream to achieve accurate synchronization of multi-source data, while taking into account device clock drift and network transmission delay. The aligned data undergoes standardization processing using the Z-score standardization method to transform the data of each feature dimension into a standard normal distribution interval with a mean of zero and a variance of one, thereby eliminating dimensional differences. During the standardization process, the mean and variance parameters are dynamically updated to adapt to the time-varying characteristics of raw coal properties and generate a standardized data matrix.

4. The intelligent dry separation dynamic optimization method for coal preparation plants according to claim 1, characterized in that, The process of obtaining the effective parameter space includes: A multidimensional data set of parameter combinations is extracted from the effective parameter space and used as the initial solution set for genetic algorithm optimization, so that it meets the physical limits of the equipment, the process flow and safety standards. A strategy combining random sampling and stratified sampling is adopted to initialize the population of the genetic algorithm. The parameter space is uniformly covered by the Latin hypercube sampling method to avoid local clustering during population initialization. By combining historically optimal parameter combinations, a hybrid initialization mechanism is adopted, setting some individuals as empirically optimal solutions to accelerate the convergence speed; Each individual parameter is encoded using real numbers, and the individual codes include wind speed, vibration frequency, sorting angle, and feeding speed.

5. The intelligent dry coal preparation plant dynamic optimization method according to claim 1, characterized in that, Generating optimized parameter combinations also includes: Individuals in the candidate parameter set are ranked according to their fitness values, and an elite retention strategy is adopted to directly copy the top-ranked individuals to the next generation. Based on the fitness ratio selection method, parent individuals are randomly selected from the remaining individuals for crossover. The crossover operation employs a simulated binary crossover method with real-number encoding, combined with dynamic crossover probability, to generate diverse offspring individuals that meet the constraints. Mutation operations are performed on the offspring individuals after crossover, using a Gaussian mutation strategy to dynamically adjust the mutation amplitude to balance exploration and utilization.

6. The intelligent dry coal preparation plant dynamic optimization method according to claim 1, characterized in that, The process of crossover mutation includes: After mutation, constraint satisfaction checks are performed on all newly generated individuals, and parameter combinations that do not meet the safety boundaries are eliminated or corrected.

7. The intelligent dry coal preparation plant dynamic optimization method according to claim 1, characterized in that, The methods for generating standardized data matrices include: The noise filtering algorithm is executed, and the function expression is: in, To improve the wavelet threshold function; Indicates the first The raw signals from each sensor channel; This represents the standard deviation of the corresponding channel noise. and For adaptive shape parameters; Furthermore, an adaptive filtering algorithm is employed: in, The filtered output signal represents time t; The original input signal at time t; It is a time-varying filter matrix; Forgetting factor; This represents the filtered output signal at time t−1; Furthermore, principal component analysis (PCA) is used to extract key features: in, For the first 3D features in the 1st dimension The weights in each principal component; For the first Eigenvalues ​​of each principal component; Indicates the total number of feature dimensions; This represents the maximum value of the feature information entropy in the p-th principal component; For the first Information entropy of 3D features; For smoothing coefficients; Furthermore, missing value imputation employs spatiotemporal interpolation: in, For a moment The result of imputing missing values; The summation symbol indicates the summation over time. From the lower limit Up to the limit The range is accumulated; Δ is the size of the time window; As a time decay weight; Observed values; For the ambient humidity gradient; The coupling coefficient; Alignment is performed using a dynamic time warping algorithm: in, A measure of synchronization error between two time series; Two time series and The synchronization error between them is measured; For the first Each sensor at a time point Data; For the set of aligned paths; ( ) represents the minimum value on set A; The square of the L2 norm; and The first Mean and standard deviation of data from each sensor; This is the time offset penalty coefficient; Furthermore, standardization is carried out: in, The data is standardized. This is the original data; and For sliding window statistics; This is a stability factor.

8. The intelligent dry coal preparation plant dynamic optimization method according to claim 1, characterized in that, The process of generating optimized parameter combinations includes: Initialize the population by sampling and define multidimensional parameter sampling: in, For the first The first individual Dimensional parameter values; It is a randomly arranged sequence of integers; Uniformly distributed random numbers; These are the upper and lower bounds of the parameters; Population size; Furthermore, parallel fitness calculations are performed, and a multi-objective fitness function is defined: in, For the first The multi-objective fitness value of an individual; For clean coal recovery rate; Energy consumption indicators; For the first Deviation of a safety constraint; For the first Safety constraint limits; , , These are the weighting coefficients; These are the minimum and maximum values ​​of clean coal recovery rate and energy consumption indicators in historical data; Furthermore, a genetic operation is performed on the scoring matrix to define simulated binary crossover: in, For the cross-th The first offspring individual Dimensional parameter values; These are the parameter values ​​corresponding to the two individuals in the parent generation; These are the control parameters for cross-operation; Let be a random variable uniformly distributed in [0,1]. Furthermore, we define adaptive Gaussian mutation: in, For the mutated first The individual New values ​​for the dimension parameter; For the first time before the mutation The individual The original value of the dimension parameter; The magnitude of variation in parameter dimensions; A standard normally distributed random number generator; This represents the current iteration number; This represents the maximum number of iterations.

9. A dynamic optimization system for intelligent dry coal preparation in a coal preparation plant, applied to the method described in any one of claims 1-8, characterized in that, include: The installation unit is used to acquire the installation parameters of the dry separation equipment and complete the installation requirements. The data acquisition unit is used to collect raw coal property data and environmental parameters, and to perform noise filtering within a preset monitoring range to obtain the raw data stream. The data processing unit is used to extract key features from the raw data stream, imputate missing values, generate a complete dataset, and perform time series alignment on the complete dataset to generate a standardized data matrix. The judgment and arbitration unit is used to judge and arbitrate events based on a standardized data matrix and judgment thresholds, and output control commands. The execution unit is used to output control commands to the execution components of the dry separation equipment to complete dynamic adjustment and control. The recording and updating unit is used to record the effective time information of control commands and update the parameter adjustment strategy table.

Images