CN-121980543-A - Feature perception multitasking prediction and control method for lead-zinc flotation process

CN121980543ACN 121980543 ACN121980543 ACN 121980543ACN-121980543-A

Abstract

The invention relates to the technical field of automatic control and intelligent optimization of lead-zinc flotation processes, and discloses a characteristic perception multitask prediction and control method of a lead-zinc flotation process. The method comprises the steps of constructing a double-branch structure based on one-dimensional convolution of different convolution kernels and multi-head attention, constructing a backbone network based on a self-attention mechanism and a multi-layer perceptron, constructing a flotation control model based on the double-branch structure and the backbone network, acquiring a historical time sequence and preprocessing to acquire time sequence derivative characteristics, inputting the time sequence derivative characteristics into the flotation control model, capturing dynamic characteristics under different time scales by the double-branch structure, eliminating scale conflicts of the dynamic characteristics by the backbone network based on the self-attention mechanism to acquire hidden characteristics, performing multi-task collaborative prediction based on the hidden characteristics, outputting a flotation working condition prediction result and first control output, and regulating and controlling the medicament input amount in a flotation process in real time based on the first control output. The problem that the existing artificial intelligent control model can not realize accurate control of the flotation dosing amount is solved.

Inventors

HUANG YONG
ZOU LICHAO
ZHU SIHAN
LIU YANG

Assignees

长沙矿冶研究院有限责任公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (12)

1. The characteristic perception multitasking prediction and control method for the lead-zinc flotation process is characterized by comprising the following steps: Step 1, constructing a backbone network for cross-scale feature fusion and multi-task prediction based on a self-attention mechanism and a multi-layer perceptron, and constructing a lead-zinc flotation control model adapting to industrial site time sequence features based on a double-branch structure of different time scale dynamic features and the backbone network; step 2, acquiring a historical time sequence of a control object and a control means in the lead-zinc flotation process, preprocessing the historical time sequence, and acquiring time sequence derivative characteristics representing the flotation data on the working condition change difference and the internal running state; And 3, inputting the time sequence derived features into the lead-zinc flotation control model, adaptively capturing dynamic features under different time scales based on the time sequence derived features by the double-branch structure, performing multi-task collaborative prediction by the backbone network, synchronously outputting a flotation working condition prediction result and first control output, and regulating and controlling the medicament input amount in the lead-zinc flotation process in real time based on the first control output.
2. The method for predicting and controlling the feature perception multitasking of the lead-zinc flotation process according to claim 1, wherein the double-branch structure is a double-branch structure which is constructed based on one-dimensional convolution of different convolution kernel sizes and combined with multi-head attention and is used for acquiring dynamic features of different time scales.
3. The method for predicting and controlling the feature perception multitasking of a lead-zinc flotation process according to claim 2, wherein the double-branch structure comprises a short-term branch facing a rapid working condition response and a long-term branch facing a long-term trend judgment, the short-term branch and the long-term branch are both constructed according to a network structure relation of one-dimensional convolution and multi-head attention, and a convolution kernel of the one-dimensional convolution in the short-term branch is smaller than a convolution kernel of the one-dimensional convolution in the long-term branch.
4. The method for predicting and controlling the characteristics of lead-zinc flotation process according to claim 3, wherein the short-term branch and the long-term branch are characterized in that the characteristics of one-dimensional convolution output are subjected to fitting flotation process mechanism characteristic splitting, the state characteristics of the control object and the operation characteristics of the control means in the lead-zinc flotation process are obtained, the state characteristics of the control object are used as a query matrix of the multi-head attention, the operation characteristics of the control means are used as a key matrix, and the complete characteristics of the one-dimensional convolution output are used as a value matrix.
5. The method for predicting and controlling the characteristic perception multitasking of a lead-zinc flotation process according to any one of claims 1 to 4, wherein the network structure of the backbone network comprises an additive layer, a self-attention layer, a multi-layer perceptron layer and a multitasking linear layer in sequence; The backbone network performs multi-task collaborative prediction specifically comprises the steps of mutually fusing dynamic features of different time scales based on an additional layer to obtain fusion features, obtaining hidden features through restraining scale differences and redundant information by a self-attention layer, performing strong nonlinear mapping based on a multi-layer perceptron layer to obtain joint features, and obtaining a prediction result and first control output based on the joint features and a multi-task linear layer.
6. The method for feature-aware multitasking and controlling a lead-zinc flotation process according to any one of claims 1-4, characterized in that in step 2 the historical time series for control objects and control means is built based on control object data and control means data at each time step; The control object data comprise lead-zinc concentrate grade, ore pulp concentration and metal recovery rate; the control means data comprise the amount of foaming agent, the amount of inhibitor and the amount of collector.
7. The method for predicting and controlling the characteristic perception multitasking of the lead-zinc flotation process according to claim 6, wherein the preprocessing comprises normalization and derivation processes of eliminating dimension difference and suppressing working condition noise; the normalization is expressed by the following formula: ; Wherein, the Representing normalized processed Feature vectors of the time step; Representation of A historical time series of time steps; representing a mean of the historical time series samples; standard deviation representing historical event sequence samples; the derivatization process is represented by the following formula: ; ; Wherein, the Representing derived features; Representing the mean features calculated over a predetermined time window.
8. The feature-aware multitasking and control method of a lead-zinc flotation process of any one of claims 1-4, characterized in that the loss function of said lead-zinc flotation control model is constructed based on a flotation state prediction objective function, a dosing control prediction objective function, and a prediction control co-constraint objective function; The flotation state prediction objective function is constructed based on the error combination mean square error regression loss of the prediction result and the real observation value; the dosing control prediction objective function is constructed based on the first control output and L2 regularization loss; and constructing the predictive control cooperative constraint objective function based on the target flotation state and the predicted result in combination with the target deviation penalty loss.
9. The method for predicting and controlling the characteristic-aware multitasking of a lead-zinc flotation process according to any one of claims 1 to 4, wherein the real-time controlling the dosage of the chemical in the lead-zinc flotation process based on the first control output specifically comprises: After the lead-zinc flotation control model outputs a prediction result, a deviation vector is obtained based on the production target running state and the prediction result, a second control output is obtained based on the deviation vector and the deviation sensing correction function and the first control output, and then the medicament input amount in the lead-zinc flotation process is regulated and controlled based on the second control output.
10. The method for feature-aware multitasking and controlling a lead-zinc flotation process according to claim 9, wherein said production target operating conditions include target concentrate grade and target metal recovery; the second control output is obtained based on the deviation vector and the deviation sensing correction function and the first control output, and the second control output is expressed by the following formula: ; Wherein, the Representing a second control output; Representing a first control output; Representing a dynamic characteristic control gain matrix of the adaptive flotation system; Representing the deviation vector.
11. The method for predicting and controlling the feature perception multitasking of the lead-zinc flotation process according to claim 9, wherein the controlling the dosage of the chemical in the lead-zinc flotation process based on the second control output comprises the following steps: after the second control output is obtained, carrying out constraint processing on the second control output to obtain a third control output, and regulating and controlling the dosage of the medicament in the lead-zinc flotation process based on the third control output; the constraint process is expressed by the following formula: ; Wherein, the Representing a third control output; And (3) with Representing the minimum and maximum ranges actually permitted for each type of medicament, respectively.
12. The method for predicting and controlling the characteristic perception multitasking of the lead-zinc flotation process according to claim 1, wherein the learning rate in the training process of the lead-zinc flotation control model adopts a staged adjustment strategy, and an optimizer adopts Adam or SGD and combines weight attenuation, momentum parameters and an early stopping mechanism.

Description

Feature perception multitasking prediction and control method for lead-zinc flotation process Technical Field The invention relates to the technical field of automatic control and intelligent optimization of zinc flotation technology, in particular to high-precision prediction of key indexes of a flotation tank and optimal control of reagent addition based on a multi-scale feature perception model. Background In the lead-zinc flotation production process, accurate control of ore pulp grade, concentration, recovery rate and reagent addition amount is a core link for ensuring flotation efficiency and product quality. If the key parameters cannot be timely monitored and reasonably regulated, not only the flotation recovery rate is reduced and the grade fluctuation is increased, but also the reagent waste, the unstable operation of the flotation tank and even the safety and the economic benefit of the subsequent beneficiation process are possibly caused. Therefore, how to realize the efficient prediction of the flotation state and the control of reagent addition is always a key problem to be solved in the field of mineral separation automation and intelligence. With the development of artificial intelligence and deep learning technologies, convolutional Neural Networks (CNNs), long-term memory networks (LSTM) and Transformer architectures are increasingly being introduced into flotation state prediction and process optimization research. However, the existing methods still suffer from several drawbacks in flotation process prediction and control applications, including in particular: The multiscale feature perception capability is limited in that traditional CNN feature extraction generally relies on a fixed convolution kernel, and it is difficult to capture both short-term fluctuations and long-term trends in the flotation state. LSTM can capture history dependence when processing long sequences, but is not sensitive enough to transient fluctuation response, and transducer has high reasoning delay in industrial real-time control, so that the real-time property of control is limited. The multi-task joint optimization capability is insufficient, the existing method only focuses on the prediction of a single index (such as grade or recovery rate), and the grade, the recovery rate and the medicament dosage are difficult to optimize simultaneously. The lack of a unified multitasking mechanism allows the first control output to deviate from the actual process requirements. Control response hysteresis although some studies have attempted to combine predicted output with control strategies, real-time control capability is still limited in industrial sites. The existing method is difficult to quickly and stably adjust the dosing of the medicament according to the flotation state, and particularly when the properties of ores are complex or the working conditions are frequently changed, the control effect is unreliable. The adaptability to abnormal working conditions is insufficient, the ore pulp state is complex and changeable, and short-term fluctuation, sudden abnormality or measurement noise exists. In the prior art, prediction errors and control deviations are easy to occur when the complex conditions are processed, and the flotation stability and the reagent feeding efficiency are affected. In summary, the prior art has obvious defects in aspects of multi-scale feature extraction, history and instantaneous dynamic fusion, multi-task combined prediction, real-time control, abnormal working condition adaptability and the like. Even with the CNN, LSTM or transducer methods, it is difficult to achieve high precision, fast response and stable multi-objective control in industrial flotation lines. Therefore, a new solution is needed to solve the above problems. Disclosure of Invention The invention provides a characteristic perception multitask prediction and control method for a lead-zinc flotation process, which aims to solve the problem that the existing artificial intelligent control model can not realize accurate control of flotation dosing. In order to achieve the above object, the present invention is realized by the following technical scheme: The invention provides a characteristic perception multitasking prediction and control method for a lead-zinc flotation process, which comprises the following steps: Step 1, constructing a backbone network for cross-scale feature fusion and multi-task prediction based on a self-attention mechanism and a multi-layer perceptron, and constructing a lead-zinc flotation control model adapting to industrial site time sequence features based on a double-branch structure of different time scale dynamic features and the backbone network; step 2, acquiring a historical time sequence of a control object and a control means in the lead-zinc flotation process, preprocessing the historical time sequence, and acquiring time sequence derivative characteristics representing the flotation da