CN-122021719-A - Tailing grain grading optimization intelligent decision-making method based on machine learning

CN122021719ACN 122021719 ACN122021719 ACN 122021719ACN-122021719-A

Abstract

The invention relates to the technical field of artificial intelligence and machine learning, and discloses a road base tailing grain grading optimization decision-making method based on machine learning. The method comprises the steps of obtaining feedback data of a multidimensional real-time working condition and a thickener state, performing time alignment and feature standardization, inputting a dynamic graph neural network embedded with sedimentation dynamics constraint, outputting optimal grain grading target distribution, generating an countermeasure network reverse mapping into a process regulation instruction through conditions, and realizing online model updating. According to the invention, through high-precision fusion perception of multi-source heterogeneous data, fine granularity state representation of the whole tailing sedimentation process is realized, sedimentation coupling relations among particles with different particle diameters are explicitly modeled by utilizing a graph structure, and physical interpretability and generalization capability of grading prediction are improved.

Inventors

LI WEN
XU HAILIN
YANG FULI
JIA JUN
WANG YANAN
HAN XIPING
YAN JIAKUN
LIU ZHIDONG
LIU NING
CHEN XIAOLING
LI JING
LV WANG
ZHANG PENGFEI
MEI XU
MENG YONG
Qin Xinliang
JING JIAN
YAN RUNBO
GUO HAIYAN
FANG SHUYAN
ZHANG MENG
WANG JIE
PAN XIAOLONG
YU CHANGJIANG
ZHANG YI
WANG ZHANJUN
QIN RUI
HAN ZE
MA WENBO
MA DONGMEI
GUO JIANXUN
ZHAO RUIBIN
GUO YANYAN
ZHAO SHENG
ZHANG YONGQIANG

Assignees

内蒙古交科路桥建设有限公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. The road base tailing grain grading optimization decision-making method based on machine learning is characterized by comprising the following steps of: acquiring multidimensional real-time working condition data of tailing slurry; Synchronously collecting feedback data of the operation state of the thickener, wherein the feedback data of the operation state of the thickener comprise overflow water turbidity, underflow emission concentration, mud layer interface height, sedimentation velocity gradient and harrow current value; performing time alignment and feature standardization processing on the multidimensional real-time working condition data and thickener running state feedback data to generate a high-dimensional feature vector sequence under a unified time reference; Inputting the high-dimensional feature vector sequence into a pre-trained dynamic graph neural network model, wherein the dynamic graph neural network model takes a particle swarm sedimentation dynamics equation as physical constraint, constructing a particle size section of a node representing tailing particle section, and introducing a gating circulation unit to perform memory coding on historical working condition evolution trend, so as to output optimal particle grading target distribution under the current working condition; Based on the optimal grain distribution target distribution, a reverse process parameter mapping module is called, the reverse process parameter mapping module adopts a condition to generate an countermeasure network architecture, a generator of the reverse process parameter mapping module receives the target distribution and the current working condition characteristic as input and outputs a group of executable process control instructions, wherein the process control instructions comprise the adjustment quantity of the overflow port diameter of the grinding classification cyclone, the molecular weight selection mark of a flocculating agent, the feeding rate of the flocculating agent and the opening increment of the underflow discharge valve of the thickener; And issuing the process control instruction to a field execution mechanism, collecting new running state feedback data in the next sampling period, and generating the discriminator weight of the countermeasure network by using the graph structure parameters and the conditions for updating the dynamic graph neural network model on line to realize closed loop self-adaptive optimization.
2. The intelligent decision-making method for optimizing tailing granule gradation based on machine learning according to claim 1, wherein the step of obtaining multidimensional real-time working condition data of the tailing slurry specifically comprises the following steps: Collecting particle size distribution data of the ground mineral products by an online laser particle size analyzer; measuring the instantaneous flow and the volume concentration in a feeding pipeline of the thickener by combining a mass flowmeter and a densimeter; Monitoring the turbidity of the water body at the overflow weir of the thickener in real time through a high-precision turbidity sensor; Continuously measuring the interface height of the mud layer in the thickener by using a radioisotope interface detector; the current transformer arranged on the harrow machine driving motor is used for collecting harrow machine current signals, and the steady state current average value is extracted after low-pass filtering and is used as an agent index of the mud layer compaction degree.
3. The intelligent decision-making method for tailing granule grading optimization based on machine learning according to claim 2, wherein the time alignment and feature standardization processing for the multidimensional real-time working condition data and thickener operation state feedback data specifically comprises: Resampling all other sensor data by adopting a cubic spline interpolation method by taking a time stamp of a feeding flow signal of a thickener as a reference master clock, and unifying the data to 1 data point per second; Respectively calculating the average value and standard deviation of the sliding window length of 300 seconds for each characteristic dimension, and adopting a Z-score method to dynamically standardize so as to eliminate the dimension difference and the long-term drift influence of the sensor; And performing cumulative distribution function conversion on the granularity distribution data, reconstructing the granularity distribution data into a continuous monotonically increasing function in a discrete histogram form, expanding and retaining the first 20-order coefficient as a frequency domain characteristic through Fourier series, compressing the data dimension and retaining the distribution form key information.
4. The intelligent decision-making method for optimizing the grain size distribution of tailings based on machine learning according to claim 3, wherein outputting the optimal grain size distribution target distribution under the current working condition comprises the following steps: defining a graph node set, wherein each node corresponds to a particle size interval, the particle size interval is divided into 30 interval sections according to logarithmic equidistant, and a full size range from 0.8 micrometers to 600 micrometers is covered; Initializing a node characteristic vector which comprises mass fraction of a particle size section, stokes sedimentation speed and absolute value of surface potential; Constructing an adjacency matrix, wherein the connection weight between adjacent grain diameter nodes is determined by the reciprocal of the sedimentation velocity difference of two nodes, and the weight of non-adjacent nodes is set to be 0; in the forward propagation process of the network, a physical consistency check module is introduced after each layer of graph convolution operation, and the forced output grading distribution meets the mass conservation constraint, namely the sum of mass fractions of all particle size sections is equal to 1; And adding a sedimentation flux conservation term in the loss function, wherein the sedimentation flux conservation term calculates the residual square between the sedimentation interface speed predicted by the model and the actually measured mud layer interface height change rate.
5. The intelligent decision-making method for tailings grain composition optimization based on machine learning of claim 4 wherein outputting a set of executable process control instructions comprises: splicing the optimal grain distribution target distribution and the high-dimensional characteristic vector of the current working condition to form a 105-dimensional input vector; Inputting the 105-dimensional input vector into a generator of a four-layer fully-connected neural network structure, wherein the number of neurons of an hidden layer is 256, 128 and 64 in sequence, and the dimension of an output layer is 4, and the parameters respectively correspond to the diameter adjustment quantity of an overflow port of the grinding classifying cyclone, the molecular weight selection identification of a flocculating agent, the feeding rate of the flocculating agent and the opening increment of an underflow discharge valve of a thickener; The Wasserstein distance is used as a condition to generate an optimization target of the countermeasure network, and a gradient penalty term is applied to ensure training stability.
6. The intelligent decision-making method for optimizing tailing granule gradation based on machine learning according to claim 5, wherein the process control instruction is issued to an on-site execution mechanism, and new operation state feedback data is collected in a next sampling period, for updating graph structure parameters of a dynamic graph neural network model and a discriminator weight of a condition generation countermeasure network on line, comprising: receiving the process control instruction through a programmable logic controller, driving an electric control valve to control the opening of an underflow discharge valve, switching a flocculant storage tank supply pipeline, and sending a cyclone overflow port adjustment request to a DCS system between grinding cars; after the end of the next sampling period, re-executing the collection of the multidimensional real-time working condition data and the running state feedback data; When the number of accumulated new samples reaches 1000, the sedimentation coupling strength among all the grain diameter nodes is recalculated based on the new data, the adjacent matrix is updated, and the stacking weight parameters of the drawing volume are smoothly updated by adopting an exponential moving average strategy, wherein the smoothing factor is set to be 0.95.
7. The intelligent decision-making method for tailing granule gradation optimization based on machine learning according to claim 6, wherein when the turbidity of overflow water is greater than a preset threshold value of 50NTU in 5 continuous control periods, triggering a model structure recalibration mechanism: increasing the division density of the particle size interval from 30 sections to 40 sections; All node feature vectors are reinitialized based on the most recently acquired data.
8. The intelligent decision-making method for tailing granule grading optimization based on machine learning according to claim 7, wherein the dimension of the high-dimensional feature vector sequence is 75 dimensions, the frequency domain features of the granularity distribution are 40 dimensions, and the rest of working conditions and state parameters are 35 dimensions.
9. The intelligent decision-making method for tailing granule grading optimization based on machine learning according to claim 8, wherein each layer of graph rolling operation of the dynamic graph neural network model is connected to a gating circulation unit for memory coding of historical working condition evolution trend.
10. The intelligent decision making method for tailing granule gradation optimization based on machine learning according to claim 9, wherein the discriminator of the condition generation countermeasure network adopts a three-layer full connection structure, and the input is a splicing vector of a real process parameter and a target gradation distribution or the output of a generator is a splicing vector of a target gradation distribution, and the output is an authenticity probability value.

Description

Tailing grain grading optimization intelligent decision-making method based on machine learning Technical Field The invention belongs to the technical field of artificial intelligence and machine learning, and particularly relates to a road base tailing grain grading optimization decision-making method based on machine learning. Background The cement stabilized base is used as a core bearing layer of the road structure, and the service durability of the cement stabilized base directly determines the service life and the operation and maintenance cost of the road engineering. In the actual operation process, the base layer is extremely easy to generate micro cracks under the influence of multiple factors such as seasonal temperature and humidity circulation alternation, repeated action of running load and the like, and the initial stage of the micro cracks has no obvious structural damage, but becomes a main channel for water infiltration. After moisture invades, fine aggregates in the base layer are washed, chain diseases such as mud pumping, plate bottom void, crack expansion and the like are caused, and finally, pavement subsidence, cracking and even structural damage are caused. The prior art adopts passive protection means aiming at the problems, such as adding a soil paving grille, spraying a waterproof coating or post grouting repair, and the like, and the methods have high maintenance cost and complicated construction, can not block disease evolution after crack generation from the root, lack active repair capability and are difficult to meet the requirement of a high-grade road on long-term stability of a base layer. Under the background, tailings are used as main solid waste generated in the metal mine beneficiation process, and the recycling utilization of the tailings provides a new way for improving the performance of base materials and reducing engineering cost. The tailings can be used as part or all of aggregate to replace natural aggregate for cement stabilization base, so that solid waste can be consumed in a large scale, environmental pressure can be reduced, grading composition of materials can be improved, and mechanical properties, crack resistance and durability of the base are further affected. However, reasonable regulation and control of tailing grain composition is a key precondition for realizing efficient and stable engineering application The tailing grain grading optimization generally relates to the collaborative design of a plurality of physical parameters such as grain size distribution, fineness modulus, non-uniformity coefficient and the like so as to meet the requirements of strength, compactness and permeability in specific engineering scenes. Currently, decision methods in the field mainly depend on engineer experience or static models based on historical data, such as a fixed proportioning method, a table look-up method, a linear interpolation method and the like. The prior art has the common problems of strong subjectivity, weak generalization capability, response lag and the like. The nonlinear coupling relation among multiple variables is difficult to quantify by artificial experience, key interaction effects are easy to ignore, and a static model cannot sense dynamic working conditions such as raw material fluctuation, equipment state change or external environment disturbance, so that an optimization result deviates from actual requirements. Especially under the scenes of complicated multi-source tailing mixed discharge, seasonal temperature and humidity change, emergency treatment and the like, the traditional method cannot adjust the grading strategy in real time, lacks the capability of predicting the evolution trend of engineering performance, and severely restricts the safe and efficient utilization of tailing resources. Therefore, there is a need for an intelligent decision method that can fuse multi-source heterogeneous data, adaptively learn the grading-performance mapping law, and dynamically generate an optimization scheme. Disclosure of Invention The invention provides a machine learning-based road base tailing grain composition optimization decision method, and aims to solve the problems of low grading efficiency, water quality deterioration and resource waste caused by difficulty in adapting to dynamic change of mineral separation working conditions in a static mathematical model due to the fact that tailing grain composition regulation and control depend on artificial experience judgment in the prior art. According to the invention, a multisource heterogeneous data fusion perception system is constructed, a particle grading prediction and regulation model with self-adaption dynamic working conditions is established, and accurate collaborative optimization of tailing sedimentation characteristics, overflow clarity and underflow concentration is realized. The invention provides a road base tailing grain composition optimization decision-making method b