Dynamic calibration and multi-agent cooperative control method for intelligent mine water dosing system

Abstract

The invention relates to the field of sewage treatment, in particular to a dynamic calibration and multi-agent cooperative control method for an intelligent mine water dosing system, which comprises the following steps: S1, collecting multi-source data; s2, extracting a characteristic alumen ustum image; s3, standard liquid is injected into the turbidity meter measuring pool to obtain the slope for first calibration, and second calibration is carried out through a back propagation neural network soft measurement model in combination with the feeding flow and the characteristic alumen ustum image; s4, the flow, the concentration, the turbidity and the image are input into a Gaussian mixture model to judge working conditions, and the total basic dosing amount is obtained through feedforward of a time sequence convolutional network-back propagation neural network mixture model; s5, according to the turbidity and the target deviation and the ideal floc characteristics and the actual deviation, the total compensation dosage is calculated through a fuzzy proportion-integral-differential algorithm, and the total basic dosage is adjusted; and S6, constructing a 9-grid reference ratio look-up table to obtain a reference addition ratio, outputting a correction coefficient according to a hard correction rule (cost priority, extreme dose protection and oscillation suppression), and finally calculating the addition amount of the main flocculant and the coagulant aid.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment, and in particular to a dynamic calibration and multi-chemical synergistic control method for an intelligent dosing system for mine water. Background Technology

[0002] Existing intelligent chemical dosing technology for coal slurry water uses a single floc feedback model to predict dosage, but it suffers from three major bottlenecks: First, core sensors such as online turbidity meters are easily contaminated and scaled in complex water conditions, leading to data drift or failure. The lack of online calibration methods causes the system to make incorrect decisions due to "perception distortion." Second, the single-model structure is insufficiently adaptive to abrupt changes in influent load and nonlinear water quality, exhibiting significant response lag and making long-term stable operation difficult. Third, the lack of a dynamic coordination mechanism for multi-chemical dosing aimed at optimizing total cost, coupled with independent control of the main flocculant and coagulant aid, results in wasted chemical consumption. These problems lead to high chemical costs per ton of water, persistently high rates of effluent exceeding standards, and poor system robustness. Therefore, a closed-loop control method integrating dual calibration, dual-model adaptation, and economic rule coordination is urgently needed.

[0003] Chinese Patent Publication No. CN114545985B discloses a dosing system and method based on floc characteristic monitoring and process feedback, relating to the field of chemical phosphorus removal and flocculation in wastewater treatment. It includes: a water quality monitoring module for monitoring the influent and effluent water quality; a floc particle size characteristic monitoring module for continuously monitoring the floc particle size characteristics during flocculation; a numerical model analysis module connected to the water quality monitoring module and the floc particle size characteristic monitoring module, which performs numerical model analysis of dosing requirements based on influent water quality parameters (feedforward), effluent water quality parameters (feedback), and dynamic changes in floc particle size characteristics (process feedback); a calculation module connected to the numerical model analysis module for self-adjustment optimization of dosing; and a dosing control module connected to the calculation module for real-time control of dosing dosage. This invention dynamically monitors the effect of the chemical phosphorus removal / flocculation process in real time, effectively providing feedback on the treatment effect and enabling predictive control. However, this scheme still suffers from problems such as inaccurate drug dosing, high drug consumption, and poor system robustness due to the lack of dual calibration of sensors, insufficient adaptive capability of single models, and lack of economic synergy mechanism for multiple drugs. Summary of the Invention

[0004] To address this, the present invention provides a dynamic calibration and multi-agent collaborative control method for an intelligent mine water dosing system, which overcomes the problems of inaccurate dosing, high chemical consumption, and poor system robustness caused by the lack of dual sensor calibration, insufficient single-model adaptive capability, and lack of economical collaborative mechanism for multiple chemicals in the prior art.

[0005] To achieve the above objectives, the present invention provides a dynamic calibration and multi-reagent synergistic control method for an intelligent mine water dosing system, comprising: Step S1: Collect data from multiple sources; Step S2: Extract features from the alum flower images in the multi-source data to obtain feature alum flower images; Step S3: Perform a first calibration on the online turbidity meter using a hardware calibration method. Also, obtain the turbidity confidence level based on the feed flow rate and the characteristic floc image from the multi-source data. Perform a second calibration on the overflow turbidity based on the turbidity confidence level to obtain the calibrated overflow turbidity. Step S4: The real-time operating conditions are judged based on multi-source data, characteristic floc images, and calibrated overflow turbidity. The total basic dosage is also obtained based on multi-source data and real-time operating conditions. Step S5: Calculate the total compensation dosage based on the turbidity of the overflow water after calibration, and adjust the total basic dosage based on the total compensation dosage. Step S6 involves constructing a benchmark ratio lookup table, obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark ratio lookup table and real-time operating conditions, outputting correction coefficients according to hard correction rules, calculating the dosage of main flocculant and coagulant aid based on the correction coefficients, total basic dosage, benchmark main flocculant dosage ratio, and benchmark coagulant aid dosage ratio, and outputting the dosage of main flocculant and coagulant aid as a synergistic control strategy.

[0006] Further, step S2 extracts features from the alum flower images in the multi-source data to obtain feature alum flower images, including: Step S21: Denoise the alum flower image in the multi-source data to obtain a denoised alum flower image; Step S22: Enhance the denoised alum flower image to obtain the enhanced alum flower image; Step S23: Quantitative features of the enhanced floc image are extracted using computer vision algorithms to obtain a characteristic floc image. The quantitative features include: average particle size of the flocs, uniformity of particle size distribution, and clarity of floc outline.

[0007] Furthermore, the first calibration of the online turbidimeter using a hardware calibration method in step S3 includes: Step S311: Inject a standard solution of a preset concentration into the turbidity meter's measuring cell, and obtain the turbidity rise slope Zc measured by the online turbidity meter within five minutes; Step S312: Compare the rising slope Zc with the preset rising slope Z0, judge the response of the online turbidimeter based on the comparison result, and perform the first calibration of the online turbidimeter based on the judgment result, wherein: When Zc≥Z0, the online turbidity meter is judged to have a fast response, and no first calibration is performed on the online turbidity meter; When Zc < Z0, the online turbidimeter is determined to be slow, and the online turbidimeter is subjected to a first calibration. The first calibration includes: triggering an online turbidimeter fault alarm, and having the online turbidimeter be repaired by staff.

[0008] Further, step S3, which involves obtaining the turbidity confidence level based on the feed flow rate and the characteristic floc image from the multi-source data, and performing a second calibration of the overflow turbidity based on the turbidity confidence level, includes: The feed flow rate and characteristic floc images are input into the preset turbidity soft measurement model. The turbidity confidence level T_virtual output by the preset turbidity soft measurement model is obtained. The turbidity T_control of the calibrated overflow water is calculated based on the overflow water turbidity T_sensor, the turbidity confidence level T_virtual, the first calibration weight coefficient w1, and the second calibration weight coefficient w2. T_control = w1 × T_sensor + w2 × T_virtual is set.

[0009] Furthermore, the pre-set turbidity soft measurement model is an online fusion of a lightweight backpropagation neural network; The pre-set turbidity soft measurement model includes a turbidity input layer, a turbidity hiding layer, and a turbidity output layer; The turbidity input layer includes nodes for feed flow rate, feed concentration, average floc size, mud height, and manual test value. The turbidity hiding layer is a single-layer structure with 10 hidden nodes, and the activation function of the turbidity hiding layer is the Tanh activation function. The output layer has 1 node and directly maps turbidity confidence.

[0010] Furthermore, step S4, which involves determining the real-time operating condition based on multi-source data, characteristic floc images, and calibrated overflow turbidity, includes: Input the feed flow rate, feed concentration, turbidity of the overflow water after calibration, average particle size of flocs, uniformity of particle size distribution, and clarity of floc outline into the pre-trained working condition judgment model to obtain the real-time working condition output by the pre-trained working condition judgment model. The real-time operating conditions include high load conditions, normal load conditions, and low load conditions.

[0011] Furthermore, step S4 also involves obtaining the total basic dosage based on multi-source data and real-time operating conditions, including: Input the feed flow rate, feed concentration, and real-time operating conditions into the pre-trained feedforward predictor to obtain the total basic dosage output by the feedforward predictor.

[0012] Further, step S5 calculates the total compensation dosage based on the turbidity of the calibrated overflow water, and adjusts the total base dosage based on the total compensation dosage, including: The target effluent turbidity, ideal average floc size, and ideal particle size distribution uniformity are obtained. The turbidity deviation is calculated based on the target effluent turbidity and the calibrated overflow turbidity, and is set as turbidity deviation = target effluent turbidity - calibrated overflow turbidity. A two-dimensional deviation is also calculated based on the ideal average floc size, ideal particle size distribution uniformity, and particle size distribution uniformity, and is set as two-dimensional deviation = [ideal particle size distribution uniformity - average floc size, ideal particle size distribution uniformity - particle size distribution uniformity]. The turbidity deviation and two-dimensional deviation are input into a pre-set multivariable feedback controller to obtain the pre-set total compensation dosage output by the multivariable feedback controller. The total basic dosage is adjusted based on the total compensation dosage to obtain the adjusted total basic dosage, which is set as total basic dosage + total compensation dosage. The value of the total basic dosage is then replaced with the value of the adjusted total basic dosage.

[0013] Further, step S6 involves constructing a benchmark ratio lookup table and obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark ratio lookup table and real-time operating conditions. Specifically, this includes: Based on the dosage ratios corresponding to high-load, normal-load, low-load, and historical conditions, a benchmark ratio lookup table is constructed to obtain the total basic dosage and the benchmark ratio lookup table with three dosage levels multiplied by three conditions in 9 cells. The real-time conditions are then input into the benchmark ratio lookup table to obtain the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio output by the benchmark ratio lookup table.

[0014] Further, step S6 outputs the correction coefficients based on the hard correction rules and multi-source data, calculates the dosage of the main flocculant and the dosage of the coagulant aid based on the correction coefficients, the total basic dosage, the benchmark main flocculant dosage ratio, and the benchmark coagulant aid dosage ratio, and outputs the dosage of the main flocculant and the dosage of the coagulant aid as a synergistic control strategy, specifically including: The hard correction rules include cost-priority rules, extreme dosage protection rules, and oscillation suppression rules; The cost priority rule includes: The price ratio A3 is calculated based on the prices of the primary flocculant (A1) and the coagulant aid (A2) from multi-source data. A3 is set to A3 = A1 / A2. The correction coefficient J is then output based on the price ratio A2, where: When A3 > 3, the correction coefficient J = -0.08 will be output. When 1.5 < A3 ≤ 3, the correction coefficient J = 0 will be output. The extreme dosage protection rules include: The total basic dosage B is compared with the preset dosage B0. Based on the comparison results, the correction coefficient J is output, and B0 is set to 70ppm, where: When B≤B0, the correction coefficient J=0 will be output; When B > B0, the correction coefficient J = -9.02 will be output. The oscillation suppression rules include: The correction coefficient J is output based on the change C of the main flocculant dosage in the previous period from the multi-source data, where: When |C|>0.15, the correction coefficient J=-0.5×sign(C) will be output; When 0.15 ≤ |C|, the correction coefficient J = 0 will be output. The dosage of main flocculant P3 and the dosage of coagulant P4 are calculated based on the correction coefficient J, the total basic dosage Q, the benchmark main flocculant dosage ratio L1 and the benchmark coagulant aid dosage ratio L2. P3 = Q × (L1 + J) and P4 = Q × [1 - (L1 + J)].

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This method, through step S1, simultaneously collects flow rate, concentration, image, turbidity, and drug price, providing a unified time base for multi-source data calibration, operating condition judgment, and economic rules, thus solving the calibration lag and coordination failure problems caused by information silos. Furthermore, through step S2, the method extracts particle size, uniformity, and contour sharpness from the alum flower image, providing reliable input for soft measurement models and operating condition identification, solving the problem of insufficient adaptive capability caused by the lack of process state feedback in single models. The method also, through step S3, uses hardware slope calibration and… The method employs a dual-channel calibration of image and flow rate soft measurement to reduce the long-term reliability of overflow turbidity from ±15% to ±3%, overcoming the problem of inaccurate dosing caused by the lack of dual sensor calibration. Furthermore, in step S4, the method uses calibration multidimensional features to look up the operating conditions and total basic dosage, and feedforward prediction shortens the load step response time by 60%, solving the hysteresis and poor generalization problems caused by insufficient adaptive capability of single models. In step S5, the method uses calibration turbidity deviation to drive fuzzy PID to generate the total compensation dosage, and positive and negative corrections keep the steady-state turbidity deviation within ±2 NTU in the long term, solving the problem of excessive effluent and wasted chemicals caused by the inability of single models to eliminate steady-state errors. Finally, in step S6, the method uses a benchmark lookup combined with chemical price ratio, extreme dosage, and hard correction rules for oscillation amplitude to output the main-coagulant dosage from the zero model, reducing the chemical cost per ton of water by 18.7% compared to existing technologies, thus compensating for the lack of an economic synergy mechanism for multiple chemicals. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating the dynamic calibration and multi-agent synergistic control method of the intelligent mine water dosing system in this embodiment. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, it is a flowchart illustrating the dynamic calibration and multi-agent synergistic control method of the intelligent mine water dosing system in this embodiment, including: Step S1: Collect data from multiple sources; Step S2: Extract features from the alum flower images in the multi-source data to obtain feature alum flower images; Step S3: Perform a first calibration on the online turbidity meter using a hardware calibration method. Also, obtain the turbidity confidence level based on the feed flow rate and the characteristic floc image from the multi-source data. Perform a second calibration on the overflow turbidity based on the turbidity confidence level to obtain the calibrated overflow turbidity. Step S4: The real-time operating conditions are judged based on multi-source data, characteristic floc images, and calibrated overflow turbidity. The total basic dosage is also obtained based on multi-source data and real-time operating conditions. Step S5: Calculate the total compensation dosage based on the turbidity of the overflow water after calibration, and adjust the total basic dosage based on the total compensation dosage. Step S6 involves constructing a benchmark ratio lookup table, obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark ratio lookup table and real-time operating conditions, outputting correction coefficients according to hard correction rules, calculating the dosage of main flocculant and coagulant aid based on the correction coefficients, total basic dosage, benchmark main flocculant dosage ratio, and benchmark coagulant aid dosage ratio, and outputting the dosage of main flocculant and coagulant aid as a synergistic control strategy.

[0020] Specifically, this method is applied to an intelligent dosing system for mine water. Through a six-step closed-loop process, it achieves dual sensor calibration, adaptive dual-model integration, and coordinated economic rules, precisely reducing chemical consumption and improving system robustness. Step S1 synchronously collects flow rate, concentration, image data, turbidity, and chemical price, providing a unified time base for multi-source data calibration, operational condition judgment, and economic rules, thus solving the calibration lag and coordination failure problems caused by information silos. Step S2 further extracts particle size, uniformity, and contour clarity from floc images, providing reliable input for soft measurement models and operational condition identification, addressing the adaptive energy limitations caused by the lack of process state feedback in single models. To address the issue of insufficient power, the method further employs step S3 to calibrate turbidity using both hardware slope calibration and image-flow soft measurement dual-channel calibration, reducing the long-term reliability of overflow turbidity from ±15% to ±3%, thus overcoming the problem of inaccurate dosing caused by the lack of dual sensor calibration. Step S4 uses calibration multidimensional features to look up the operating conditions and total basic dosage, and feedforward prediction shortens the load step response time by 60%, solving the lag and poor generalization problems caused by insufficient single-model adaptive capability. Step S5 uses calibration turbidity deviation to drive fuzzy PID to generate the total compensation dosage, and positive and negative corrections keep the steady-state turbidity deviation within ±2 NTU in the long term, solving the problem of excessive effluent and wasted chemicals caused by the inability of a single model to eliminate steady-state errors. Step S6 combines benchmark lookup with chemical price ratio, extreme dosage, and hard correction rules for oscillation amplitude, outputting the main-coagulant dosage from the zero model, reducing the chemical cost per ton of water by 18.7% compared to existing technologies, thus compensating for the lack of an economic synergy mechanism for multiple chemicals.

[0021] Specifically, the multi-source data includes feed flow rate, feed concentration, floc image, main flocculant price, coagulant aid price, and the change in main flocculant dosage in the previous cycle. The feed flow rate refers to the instantaneous volumetric flow rate measured in real time by an electromagnetic flowmeter at the input end of the mine water intelligent dosing system, with units of m³. 3 / h, used to characterize the volumetric load of mine water entering the flocculation reaction zone per unit time; step S1 collects the feed flow rate using the exponential weighted moving average filtering method of the electromagnetic flowmeter; the feed concentration refers to the suspended solids mass concentration measured in real time by the ultrasonic concentration meter at the input end of the intelligent mine water dosing system, in g / L, used to characterize the solid phase load intensity of coal slime particles per unit volume of feed; step S1 collects the feed concentration using the median filtering method of the ultrasonic concentration meter; the floc image refers to the image of the floc formation underwater in the flocculation reaction zone. Real-time video frames captured by an industrial camera, containing floc morphology, size, and distribution information, are used to quantify the flocculation effect. Step S1 acquires floc images using a synchronous acquisition method with an underwater industrial camera and an LED ring light source. The price of the coagulant aid refers to the real-time purchase price of the coagulant aid (PAC) within the current settlement period, expressed in yuan / kg, and is used for cost priority correction rules. Step S1 collects the price of the coagulant aid using a timed retrieval method via the ERP interface. The change in the dosage of the main flocculant in the previous period refers to the difference between the dosage of the main flocculant in the previous control period (10 min) and the dosage in the period before that, expressed in ppm, and is used for oscillation suppression rules. Step S1 collects the change in the dosage of the main flocculant in the previous period using a local cache differential method.

[0022] Specifically, step S2 extracts features from the alum flower images in the multi-source data to obtain feature alum flower images, including: Step S21: Denoise the alum flower image in the multi-source data to obtain a denoised alum flower image; Step S22: Enhance the denoised alum flower image to obtain the enhanced alum flower image; Step S23: Quantitative features of the enhanced floc image are extracted using computer vision algorithms to obtain a characteristic floc image. The quantitative features include: average particle size of the flocs, uniformity of particle size distribution, and clarity of floc outline.

[0023] Specifically, this embodiment uses a cascaded algorithm of bilateral filtering and non-local means (NLM) denoising to denoise the alum floc images from the multi-source data, resulting in a denoised alum floc image. Bilateral filtering is an edge-preserving smoothing denoising method that removes high-frequency noise while retaining floc edge information. Its weighting function considers both spatial distance and pixel grayscale difference to prevent edge blurring. Non-local means denoising utilizes image redundancy information by searching for similar pixel blocks in the entire image and performing a weighted average, effectively suppressing recurring small suspended particle noise in underwater alum floc images. This embodiment further enhances the denoised alum floc image using contrast-limited adaptive histogram equalization (CLAHE) and a multi-scale Retinex algorithm, resulting in an enhanced alum floc image. Contrast-limited adaptive histogram equalization (CLAHE) is a local contrast enhancement technique that divides the image into 8×8... Tiles are subjected to histogram equalization, and the contrast amplification factor is limited (clipLimit=2.0) to avoid excessive noise amplification, significantly improving the local contrast between the flocs and the background water. The multi-scale Retinex algorithm is an illumination normalization algorithm that estimates and removes illumination components at different scales through Gaussian filtering, solving the problem of bright spots or dark areas in images caused by uneven underwater illumination, and ensuring that the floc texture remains consistent across different illumination regions. The computer vision algorithm refers to a cascaded feature extraction process that includes "morphological segmentation - edge extraction - statistical quantization". The average particle size of the flocs... The diameter refers to the size of all independent floc particles identified by the Canny edge detection and watershed segmentation algorithm in the enhanced floc image, expressed as the diameter of the equivalent circle of pixel area. The arithmetic mean of the diameters of all particles is used. The particle size distribution uniformity refers to the ratio of the standard deviation to the mean of the diameter of independent floc particles in the enhanced floc image (coefficient of variation). The closer the value is to 0, the more uniform the particle size. The floc contour sharpness refers to the average gradient amplitude of the extracted floc edge pixels in the enhanced floc image, quantified by the response intensity of the Sobel operator. The larger the value, the sharper and more distinct the floc edge.

[0024] Specifically, step S3, which involves performing a first calibration of the online turbidimeter using a hardware calibration method, includes: Step S311: Inject a standard solution of a preset concentration into the turbidity meter's measuring cell, and obtain the turbidity rise slope Zc measured by the online turbidity meter within five minutes; Step S312: Compare the rising slope Zc with the preset rising slope Z0, judge the response of the online turbidimeter based on the comparison result, and perform the first calibration of the online turbidimeter based on the judgment result, wherein: When Zc≥Z0, the online turbidity meter is judged to have a fast response, and no first calibration is performed on the online turbidity meter; When Zc < Z0, the online turbidimeter is determined to be slow, and the online turbidimeter is subjected to a first calibration. The first calibration includes: triggering an online turbidimeter fault alarm, and having the online turbidimeter be repaired by staff.

[0025] Specifically, the turbidimeter measurement cell refers to the closed sample flow cell built into the optical sensing unit of the online turbidimeter. The preset concentration standard solution refers to a turbidity standard solution prepared with formazin polymer and having national first-class standard material traceability. Its concentration gradient covers three levels: zero-point standard solution (0 NTU), mid-point standard solution (100 NTU), and full-scale standard solution (500 NTU). In this embodiment, the turbidity collected by the online turbidimeter within five minutes is used to obtain the rising slope of the turbidity measured by the online turbidimeter within five minutes. The measured turbidity refers to the turbidity of the turbidimeter measurement cell measured by the online turbidimeter. This embodiment does not limit the specific implementation method for triggering the online turbidimeter fault alarm. Those skilled in the art can set it according to the actual situation, such as triggering the online turbidimeter fault alarm by issuing a rapid buzzer alarm sound.

[0026] Specifically, step S3 involves obtaining the turbidity confidence level based on the feed flow rate and the characteristic floc image from the multi-source data, and performing a second calibration of the overflow turbidity based on the turbidity confidence level, including: Input the feed flow rate and characteristic floc image into the preset turbidity soft measurement model, obtain the turbidity confidence T_virtual output by the preset turbidity soft measurement model, and calculate the turbidity T_control of the calibrated overflow water based on the overflow water turbidity T_sensor, turbidity confidence T_virtual, first calibration weight coefficient w1 and second calibration weight coefficient w2, and set T_control=w1×T_sensor+w2×T_virtual; The pre-set turbidity soft measurement model is an online fusion of a lightweight backpropagation neural network; The pre-set turbidity soft measurement model includes a turbidity input layer, a turbidity hiding layer, and a turbidity output layer; The turbidity input layer includes nodes for feed flow rate, feed concentration, average floc size, mud height, and manual test value. The turbidity hiding layer is a single-layer structure with 10 hidden nodes, and the activation function of the turbidity hiding layer is the Tanh activation function. The output layer has 1 node and directly maps turbidity confidence.

[0027] Specifically, the first calibration weighting coefficient refers to the weighting coefficient corresponding to the overflow turbidity in the calculation process of the overflow turbidity after calibration, and the second calibration weighting coefficient refers to the weighting coefficient corresponding to the turbidity confidence level in the calculation process of the overflow turbidity after calibration. When the calculation of the overflow turbidity after calibration is in the default state, the online turbidity meter is trusted as the main factor, and w1=0.8 and w2=0.2 are set. When the online turbidity meter fault alarm is triggered and the online turbidity meter is repaired by the staff, the preset turbidity soft measurement model is completely relied upon, and w1=0 and w2=1 are set. The default state refers to the online turbidity meter operating normally. The feed flow rate node refers to the instantaneous volumetric flow rate characteristic unit in the soft measurement model input layer, which is in cubic meters per hour (m3 / h). Its value is obtained by the output of the electromagnetic flowmeter and the exponential weighted moving average filter, and is used to characterize the volumetric load of mine water entering the flocculation reaction zone per unit time. The feed concentration node refers to the suspended solids mass concentration characteristic unit in the soft sensing model input layer, expressed in grams per liter (g / L). Its value is obtained from the ultrasonic concentration meter output and median filtering, reflecting the solid load intensity of the coal slime particles in the feed. The floc average particle size node refers to the equivalent circle diameter statistical characteristic unit in the soft sensing model input layer, expressed in micrometers (μm). Its value is the arithmetic mean of the particle sizes of all independent flocs in the current frame, used to quantify the degree of particle aggregation during the flocculation reaction process. The mud layer height node refers to the interface position characteristic unit in the soft sensing model input layer, representing the vertical distance from the mud-water interface to the bottom of the thickener, expressed in meters (m). Its value is output by the ultrasonic mud-water interface meter, used to indirectly reflect the solid-liquid separation efficiency and underflow concentration change trend in the settling zone. The artificial laboratory test value node refers to the value in the soft sensing model input layer representing the laboratory's gravimetric method (GB / T). 11901) The effluent turbidity reference value characteristic unit, measured every 4 hours in NTU, is used as the true value of the slow variable to perform online drift correction for virtual turbidity, thereby suppressing the accumulation of long-term model bias. The turbidity hidden layer is a single-layer structure, which can minimize the number of matrix operations in the forward propagation. The measured inference time is stable at 8-10 ms / operation, which is significantly better than multi-layer networks (>15 ms), thus reserving sufficient computational margin for subsequent calculations. The 10 hidden nodes are not empirically assigned, but are precise solutions determined through dual optimization of offline grid search and online verification: comparing the range of 8-15 hidden nodes on the validation set, when the number of hidden nodes increases from 8 to 10, the MSE increases from 32 NTU. 2 Significantly reduced to 23 NTU 2 (meeting <25 NTU) 2 (Accuracy threshold); while when the number of hidden nodes increases to 12, the MSE only slightly decreases to 22 NTU. 2However, the 20% increase in the number of model parameters caused the inference latency to approach 12ms. Therefore, 10 hidden nodes is the minimum complexity configuration under the premise of achieving the accuracy target, realizing Pareto optimality between real-time performance and accuracy.

[0028] Specifically, step S4, which determines the real-time operating condition based on multi-source data, characteristic floc images, and calibrated overflow turbidity, includes: Input the feed flow rate, feed concentration, turbidity of the overflow water after calibration, average particle size of flocs, uniformity of particle size distribution, and clarity of floc outline into the pre-trained working condition judgment model to obtain the real-time working condition output by the pre-trained working condition judgment model. The real-time operating conditions include high load conditions, normal load conditions, and low load conditions.

[0029] Specifically, the pre-trained operating condition judgment model refers to a Gaussian mixture model that takes the feed flow rate, feed concentration, calibrated overflow turbidity, average floc size, particle size distribution uniformity, and floc outline clarity as inputs and real-time operating conditions as outputs. This embodiment does not limit the specific setting of the Gaussian mixture model. Those skilled in the art can set it according to the actual situation, as long as it meets the output of real-time operating conditions. For example, the number of cluster centers K of the Gaussian mixture model can be set to 3, corresponding to three real-time operating conditions, namely high load, normal load, and low load, with the covariance type being diagonal covariance and the prior probability π=[0.25,0.55,0.20].

[0030] Specifically, step S4 further involves obtaining the total basic dosage based on multi-source data and real-time operating conditions, including: Input the feed flow rate, feed concentration, and real-time operating conditions into the pre-trained feedforward predictor to obtain the total basic dosage output by the feedforward predictor.

[0031] Specifically, the pre-trained feedforward predictor is a hybrid model of temporal convolutional network and backpropagation neural network, which is composed of a cascaded temporal convolutional network module and a backpropagation neural network module (BP module). The two are coupled through a feature concatenation layer. The temporal convolutional network module is composed of stacked one-dimensional causal convolutional layers with the following parameters: kernel size of 3, dilation factor list of [1, 2, 4, 8, 16], number of convolutional filters of 64, activation function of ReLU activation function, residual connection of 1×1 convolutional residual mapping in each layer, output feature map size of 1×64, and output as TCN feature vector. The input layer of the backpropagation neural network module includes feed flow rate node, feed concentration node, real-time condition node, TCN feature vector node and current time node. The hidden layer of the backpropagation neural network module includes a single hidden layer and ten nodes, using Tanh activation function, and Xavier uniform distribution to initialize the weights of the hidden layer. The output layer of the backpropagation neural network module has only one node for outputting the total basic dosage.

[0032] Specifically, step S5 calculates the total compensation dosage based on the turbidity of the overflow water after calibration, and adjusts the total base dosage based on the total compensation dosage, including: The target effluent turbidity, ideal average floc size, and ideal particle size distribution uniformity are obtained. The turbidity deviation is calculated based on the target effluent turbidity and the calibrated overflow turbidity, and is set as turbidity deviation = target effluent turbidity - calibrated overflow turbidity. A two-dimensional deviation is also calculated based on the ideal average floc size, ideal particle size distribution uniformity, and particle size distribution uniformity, and is set as two-dimensional deviation = [ideal particle size distribution uniformity - average floc size, ideal particle size distribution uniformity - particle size distribution uniformity]. The turbidity deviation and two-dimensional deviation are input into a pre-set multivariable feedback controller to obtain the pre-set total compensation dosage output by the multivariable feedback controller. The total basic dosage is adjusted based on the total compensation dosage to obtain the adjusted total basic dosage, which is set as total basic dosage + total compensation dosage. The value of the total basic dosage is then replaced with the value of the adjusted total basic dosage.

[0033] Specifically, this embodiment obtains the target effluent turbidity, ideal average floc size, and ideal particle size distribution uniformity through water quality standards. The water quality standards refer to the water quality requirements for mine water set by the staff. The pre-set multivariable feedback controller is a multivariable feedback controller based on a fuzzy proportional-integral-differential algorithm. The fuzzy proportional-integral-differential algorithm is an adaptive control algorithm that maps a two-dimensional deviation vector (average floc size deviation, particle size distribution uniformity deviation) to a three-dimensional PID gain correction. It achieves online gain self-tuning through membership functions, rule bases, and defuzzification operations, as detailed below: Membership function: The turbidity deviation e_T and the two-dimensional deviation vectors (floc average particle size deviation e_F_D and particle size distribution uniformity deviation e_F_U) are respectively subjected to triangular membership functions, the universe of discourse is normalized to [-6, 6], and the linguistic variables are unified as {NB, NM, NS, ZO, PS, PM, PB}. Rule base: Establish 7×7×7=343 “IF-THEN” rules, with typical rules in the form of: IF Turbidity deviation is NB and Floc average particle size deviation is NM and Particle size distribution uniformity deviation is NS THEN ΔKp is PM, ΔKi is PS, ΔKd is NS; Defuzzing: Mamdani minimum operation activation, maximal-synthesis aggregation, and centroid method deblurring are adopted to output ΔKp, ΔKi, and ΔKd, which are superimposed on the initial values ​​Kp0, Ki0, and Kd0 to form real-time gains Kp(k), Ki(k), and Kd(k). Incremental output: Using the two-dimensional bias weighted sum e(k) = w_T·e_T(k) + w_D·e_F_D(k) + w_U·e_F_U(k) as the unified error, and substituting it into the incremental PID formula: ΔDose_comp(k) = Kp(k)[e(k) - e(k-1)] + Ki(k)e(k) + Kd(k)[e(k) - 2e(k-1) + e(k-2)], the total compensation dosage Dose_comp(k) is obtained, where w_T = 0.60, w_D = 0.25, and w_U = 0.15.

[0034] Specifically, step S6 involves constructing a benchmark mix ratio lookup table and obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark mix ratio lookup table and real-time operating conditions. This includes: Based on the dosage ratios corresponding to high-load, normal-load, low-load, and historical conditions, a benchmark ratio lookup table is constructed to obtain the total basic dosage and the benchmark ratio lookup table with three dosage levels multiplied by three conditions in 9 cells. The real-time conditions are then input into the benchmark ratio lookup table to obtain the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio output by the benchmark ratio lookup table.

[0035] Specifically, this embodiment does not limit the specific construction method of the benchmark ratio lookup table. Those skilled in the art can set it according to the actual situation. For example, when the total basic dosage falls in the range of 20-50 ppm and the real-time operating condition is the normal load, the benchmark main flocculant dosage ratio is found to be 0.40. The rest can be deduced from the table. The dosage ratio corresponding to the historical operating condition refers to the dosage ratio corresponding to the real-time operating condition collected in history.

[0036] Specifically, step S6 outputs the correction coefficients based on the hard correction rules and multi-source data, calculates the dosage of the main flocculant and the dosage of the coagulant aid based on the correction coefficients, the total basic dosage, the benchmark main flocculant dosage ratio, and the benchmark coagulant aid dosage ratio, and outputs the dosage of the main flocculant and the dosage of the coagulant aid as a synergistic control strategy, specifically including: The hard correction rules include cost-priority rules, extreme dosage protection rules, and oscillation suppression rules; The cost priority rule includes: The price ratio A3 is calculated based on the prices of the primary flocculant (A1) and the coagulant aid (A2) from multi-source data. A3 is set to A3 = A1 / A2. The correction coefficient J is then output based on the price ratio A2, where: When A3 > 3, the correction coefficient J = -0.08 will be output. When 1.5 < A3 ≤ 3, the correction coefficient J = 0 will be output. The extreme dosage protection rules include: The total basic dosage B is compared with the preset dosage B0. Based on the comparison results, the correction coefficient J is output, and B0 is set to 70ppm, where: When B≤B0, the correction coefficient J=0 will be output; When B > B0, the correction coefficient J = -9.02 will be output. The oscillation suppression rules include: The correction coefficient J is output based on the change C of the main flocculant dosage in the previous period from the multi-source data, where: When |C|>0.15, the correction coefficient J=-0.5×sign(C) will be output; When 0.15 ≤ |C|, the correction coefficient J = 0 will be output. The dosage of main flocculant P3 and the dosage of coagulant P4 are calculated based on the correction coefficient J, the total basic dosage Q, the benchmark main flocculant dosage ratio L1 and the benchmark coagulant aid dosage ratio L2. P3 = Q × (L1 + J) and P4 = Q × [1 - (L1 + J)].

[0037] Specifically, the preset dosage refers to a preset value used to judge the output of the correction coefficient.

[0038] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for dynamic calibration and multi-reagent synergistic control of an intelligent mine water dosing system, characterized in that, include: Step S1: Collect data from multiple sources; Step S2: Extract features from the alum flower images in the multi-source data to obtain feature alum flower images; Step S3: Perform a first calibration on the online turbidity meter using a hardware calibration method. Also, obtain the turbidity confidence level based on the feed flow rate and the characteristic floc image from the multi-source data. Perform a second calibration on the overflow turbidity based on the turbidity confidence level to obtain the calibrated overflow turbidity. Step S4: The real-time operating conditions are judged based on multi-source data, characteristic floc images, and calibrated overflow turbidity. The total basic dosage is also obtained based on multi-source data and real-time operating conditions. Step S5: Calculate the total compensation dosage based on the turbidity of the overflow water after calibration, and adjust the total basic dosage based on the total compensation dosage. Step S6 involves constructing a benchmark ratio lookup table, obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark ratio lookup table and real-time operating conditions, outputting correction coefficients according to hard correction rules, calculating the dosage of main flocculant and coagulant aid based on the correction coefficients, total basic dosage, benchmark main flocculant dosage ratio, and benchmark coagulant aid dosage ratio, and outputting the dosage of main flocculant and coagulant aid as a synergistic control strategy.

2. The dynamic calibration and multi-agent synergistic control method for the intelligent mine water dosing system according to claim 1, characterized in that, Step S2 extracts features from the alum flower images in the multi-source data to obtain feature alum flower images, including: Step S21: Denoise the alum flower image in the multi-source data to obtain a denoised alum flower image; Step S22: Enhance the denoised alum flower image to obtain the enhanced alum flower image; Step S23: Quantitative features of the enhanced floc image are extracted using computer vision algorithms to obtain a characteristic floc image. The quantitative features include: average particle size of the flocs, uniformity of particle size distribution, and clarity of floc outline.

3. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 1, characterized in that, The first calibration of the online turbidimeter using the hardware calibration method in step S3 includes: Step S311: Inject a standard solution of a preset concentration into the turbidity meter's measuring cell, and obtain the turbidity rise slope Zc measured by the online turbidity meter within five minutes; Step S312: Compare the rising slope Zc with the preset rising slope Z0, judge the response of the online turbidimeter based on the comparison result, and perform the first calibration of the online turbidimeter based on the judgment result, wherein: When Zc≥Z0, the online turbidity meter is judged to have a fast response, and no first calibration is performed on the online turbidity meter; When Zc < Z0, the online turbidimeter is determined to be slow, and the online turbidimeter is subjected to a first calibration. The first calibration includes: triggering an online turbidimeter fault alarm, and having the online turbidimeter be repaired by staff.

4. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 3, characterized in that, Step S3, which involves obtaining the turbidity confidence level based on the feed flow rate from the multi-source data and the characteristic floc image, and performing a second calibration of the overflow turbidity based on the turbidity confidence level, includes: The feed flow rate and characteristic floc images are input into the preset turbidity soft measurement model. The turbidity confidence level T_virtual output by the preset turbidity soft measurement model is obtained. The turbidity T_control of the calibrated overflow water is calculated based on the overflow water turbidity T_sensor, the turbidity confidence level T_virtual, the first calibration weight coefficient w1, and the second calibration weight coefficient w2. T_control = w1 × T_sensor + w2 × T_virtual is set.

5. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 4, characterized in that, The pre-set turbidity soft measurement model is an online fusion of a lightweight backpropagation neural network; The pre-set turbidity soft measurement model includes a turbidity input layer, a turbidity hiding layer, and a turbidity output layer; The turbidity input layer includes nodes for feed flow rate, feed concentration, average floc size, mud height, and manual test value. The turbidity hiding layer is a single-layer structure with 10 hidden nodes, and the activation function of the turbidity hiding layer is the Tanh activation function. The output layer has 1 node and directly maps turbidity confidence.

6. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 5, characterized in that, Step S4, which involves judging the real-time operating conditions based on multi-source data, characteristic floc images, and calibrated overflow turbidity, includes: Input the feed flow rate, feed concentration, turbidity of the overflow water after calibration, average particle size of flocs, uniformity of particle size distribution, and clarity of floc outline into the pre-trained working condition judgment model to obtain the real-time working condition output by the pre-trained working condition judgment model. The real-time operating conditions include high load conditions, normal load conditions, and low load conditions.

7. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 6, characterized in that, Step S4 further involves obtaining the total basic dosage based on multi-source data and real-time operating conditions, including: Input the feed flow rate, feed concentration, and real-time operating conditions into the pre-trained feedforward predictor to obtain the total basic dosage output by the feedforward predictor.

8. The dynamic calibration and multi-reagent synergistic control method for the intelligent mine water dosing system according to claim 7, characterized in that, Step S5 calculates the total compensation dosage based on the turbidity of the calibrated overflow water, and adjusts the total base dosage based on the total compensation dosage, including: The target effluent turbidity, ideal average floc size, and ideal particle size distribution uniformity are obtained. The turbidity deviation is calculated based on the target effluent turbidity and the calibrated overflow turbidity, and is set as turbidity deviation = target effluent turbidity - calibrated overflow turbidity. A two-dimensional deviation is also calculated based on the ideal average floc size, ideal particle size distribution uniformity, and particle size distribution uniformity, and is set as two-dimensional deviation = [ideal particle size distribution uniformity - average floc size, ideal particle size distribution uniformity - particle size distribution uniformity]. The turbidity deviation and two-dimensional deviation are input into a pre-set multivariable feedback controller to obtain the pre-set total compensation dosage output by the multivariable feedback controller. The total basic dosage is adjusted based on the total compensation dosage to obtain the adjusted total basic dosage, which is set as total basic dosage + total compensation dosage. The value of the total basic dosage is then replaced with the value of the adjusted total basic dosage.

9. The dynamic calibration and multi-agent synergistic control method for the intelligent mine water dosing system according to claim 8, characterized in that, Step S6 involves constructing a benchmark mix ratio lookup table and obtaining the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio based on the benchmark mix ratio lookup table and real-time operating conditions. Specifically, this includes: Based on the dosage ratios corresponding to high-load, normal-load, low-load, and historical conditions, a benchmark ratio lookup table is constructed to obtain the total basic dosage and the benchmark ratio lookup table with three dosage levels multiplied by three conditions in 9 cells. The real-time conditions are then input into the benchmark ratio lookup table to obtain the benchmark main flocculant dosage ratio and benchmark coagulant aid dosage ratio output by the benchmark ratio lookup table.

10. The dynamic calibration and multi-agent synergistic control method for the intelligent mine water dosing system according to claim 9, characterized in that, Step S6 outputs the correction coefficients based on the hard correction rules and multi-source data. It calculates the dosages of the main flocculant and coagulant based on the correction coefficients, the total base dosage, the baseline main flocculant dosage ratio, and the baseline coagulant aid dosage ratio. The main flocculant and coagulant aid dosages are then output as a synergistic control strategy, specifically including: The hard correction rules include cost-priority rules, extreme dosage protection rules, and oscillation suppression rules; The cost priority rule includes: The price ratio A3 is calculated based on the prices of the primary flocculant (A1) and the coagulant aid (A2) from multi-source data. A3 is set to A3 = A1 / A2. The correction coefficient J is then output based on the price ratio A2, where: When A3 > 3, the correction coefficient J = -0.08 will be output. When 1.5 < A3 ≤ 3, the correction coefficient J = 0 will be output. The extreme dosage protection rules include: The total basic dosage B is compared with the preset dosage B0. Based on the comparison results, the correction coefficient J is output, and B0 is set to 70ppm, where: When B≤B0, the correction coefficient J=0 will be output; When B > B0, the correction coefficient J = -9.02 will be output. The oscillation suppression rules include: The correction coefficient J is output based on the change C of the main flocculant dosage in the previous period from the multi-source data, where: When |C|>0.15, the correction coefficient J=-0.5×sign(C) will be output; When 0.15 ≤ |C|, the correction coefficient J = 0 will be output. The dosage of main flocculant P3 and the dosage of coagulant P4 are calculated based on the correction coefficient J, the total basic dosage Q, the benchmark main flocculant dosage ratio L1 and the benchmark coagulant aid dosage ratio L2. P3 = Q × (L1 + J) and P4 = Q × [1 - (L1 + J)].

Images