Dynamic sorting regulation and control method based on deep learning coal gasification slag ash content prediction

Abstract

The invention provides a separation dynamic regulation and control method based on deep learning coal gasification slag ash content prediction, and belongs to the technical field of mineral processing intelligent control. The method comprises the steps that original video data and actual ash content labels are collected based on the coal gasification slag reselection process; preprocessing the original video data, and obtaining an image data set by randomly extracting video frame pictures; building a deep learning ash content prediction model, training by utilizing an image data set and combining an actual ash content label, obtaining a real-time prediction result through verification, testing and evaluation, and obtaining sorting equipment subjected to preliminary adjustment; and real-time monitoring data are collected, and according to a real-time prediction result, the running state of the sorting equipment is automatically adjusted by calculating the deviation between theoretical sorting efficiency and actual sorting efficiency, so that sorting dynamic regulation and control are completed. The problems that the running state and the sorting effect of an existing sorting system depend on feeding characteristics, the real-time prediction and regulation capacity is lacked, and the quality stability is poor are solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent control of mineral processing, and particularly relates to a separation dynamic regulation and control method based on deep learning coal gasification slag ash content prediction. BACKGROUND

[0002] The key to the resource utilization of coal gasification slag lies in the effective separation of its main components, residual carbon and ash. Due to the high porosity and specific surface area of residual carbon, and the dense inorganic mineral particles of ash, there are obvious differences in their physical properties, which make them have the material basis for separation. However, the existing traditional separation technologies such as flotation and gravity separation still face many challenges in practical application. Although flotation can be used to recover residual carbon, the high dosage of reagent, high cost and low recovery rate are caused by the significant adsorption of coal gasification slag residual carbon on the collector; although gravity separation can efficiently handle coarse particles, it is difficult to effectively separate fine particles with a particle size of less than 0.2 mm, and such particle size usually accounts for more than 80% in actual samples, which becomes a bottleneck restricting the wide application of gravity separation. It is found that there is a significant difference in the distribution of carbon and ash in coal gasification slag at different particle sizes, which provides a new idea for efficient classification and separation. The hydrocyclone is suitable for rapid classification of fine particles due to its high efficiency, low cost and small footprint; and the spiral separator is suitable for high-precision separation of coarse particles. Therefore, using the combined process of "hydrocyclone + spiral separator" to divide and separate the particle size of coal gasification slag can effectively improve the separation efficiency of carbon and ash. However, the particle size composition of coal gasification slag is highly volatile, and the raw material properties are complex, which makes the running state and separation effect of the gravity separation equipment highly dependent on the characteristics of the feed, especially the ash content. The current combined separation system lacks real-time prediction and regulation and control capability of ash index, resulting in poor product quality stability and restricting the high-value utilization of subsequent resource products.

[0003] With the development of artificial intelligence and deep learning technology, deep learning models based on video monitoring have been widely applied in the field of mineral processing. Based on deep learning models such as three-dimensional convolutional network (3D CNN) or time series convolutional network (TSN), as a kind of deep learning model that can effectively capture the time sequence features in video data, it shows good performance in dynamically changing environment. By analyzing continuous video frames, TSN can extract key dynamic features in the separation process of coal gasification slag, and then realize accurate prediction of ash content. SUMMARY

[0004] In view of the above shortcomings in the prior art, the separation dynamic regulation and control method based on deep learning coal gasification slag ash content prediction provided by the present application solves the problem that the running state and separation effect of the existing combined separation system are highly dependent on the characteristics of the feed, lack real-time prediction and regulation and control capability of ash index, and the product quality stability is poor.

[0005] To achieve the above object, the technical scheme adopted by the present application is as follows: a sorting dynamic regulation and control method based on deep learning coal gasification slag ash content prediction, comprising the following steps: S1, based on the coal gasification slag reselection process, collecting original video data and actual ash content label; S2, pre-processing the original video data, and obtaining image data set by randomly extracting video frame pictures; S3, building a deep learning ash content prediction model, and using the image data set, combining the actual ash content label, training the deep learning ash content prediction model, and through verification, testing and evaluation, obtaining real-time prediction results, and obtaining the sorting equipment after preliminary adjustment; S4, according to the sorting equipment after preliminary adjustment, collecting real-time monitoring data, according to the real-time prediction results, calculating the deviation between the theoretical sorting efficiency and the actual sorting efficiency, automatically adjusting the running state of the sorting equipment, and completing the sorting dynamic regulation and control.

[0006] The beneficial effects of the present application are: the present application realizes accurate detection of product ash content by collecting ore pulp video image data and building a deep learning ash content prediction model, dynamically regulates and controls the operating parameters of the hydrocyclone and the spiral separator according to the prediction results, optimizes the separation efficiency of residual carbon and ash in the coal gasification slag, improves the sorting precision and resource utilization rate, significantly improves the sorting efficiency of fine coal gasification slag and the stability of product quality, reduces resource waste, and provides an intelligent solution for high-value utilization of coal gasification slag.

[0007] Further, the S1 comprises the following steps: S101, based on the coal gasification slag reselection process, responding to the stable flow of ore pulp, using the preset video monitoring equipment to collect at the preset time and preset acquisition frame rate, obtaining the original video data; S102, collecting the ore pulp in the original video data collection time period, and performing filtration and drying treatment on the collected ore pulp, obtaining the treated ore pulp, and performing ash content test and detection on the treated ore pulp, and obtaining the actual ash content label through the determination.

[0008] The beneficial effects of the further scheme are that: the application collects original video data when the ore pulp is stably flowing, and collects, ash analyzes and detects the ore pulp in the original video data collection time period to obtain actual ash labels, ensures that the obtained video data and ash analysis data are representative, and further improves the training quality and prediction accuracy of the ash prediction model, and provides a high-quality label data set for subsequent deep learning ash prediction model training, and enhances the learning ability of the model; by collecting and detecting the ash of the ore pulp at different time periods, more changes can be covered, so that the model is more sensitive to the changes in the ash of the ore pulp under different conditions in actual application, thereby improving the robustness and adaptability of the ash prediction; by collecting actual ash data and combining with video data for analysis, more accurate ash detection can be realized, the real-time performance and accuracy of ash control in the separation process are improved, and the separation process is further optimized.

[0009] Further, the S2 comprises the following steps: S201, denoising and enhancing the original video data, and dividing the original video data into time periods according to a preset frame rate to obtain preprocessed short video data; S202, dividing the preprocessed short video data into a training set, a validation set and a test set by hierarchical sampling; S203, obtaining an image data set by randomly extracting video frames in the training set.

[0010] The beneficial effects of the further scheme are that: the application divides the original video data of the long video into time periods to obtain short video data, and randomly extracts video frames to make an image data set, which avoids overfitting, so that the subsequent deep learning ash prediction model can learn different features from more diverse image data, improve the generalization ability of the model, and be more suitable for actual scenes; by randomly extracting video frames, different environmental changes (such as light, background, etc.) can be simulated, so that the training data is more abundant, thereby improving the recognition and processing ability of the model for image data under different conditions. It can also effectively reduce the amount of calculation required in the training process, save time and computing resources, and improve the training efficiency. Further, the generated image data set covers various dynamic situations in the video, so that the model can be exposed to more dynamic scenes during training, which helps to improve the accuracy of image classification or target detection.

[0011] Further, the S3 comprises the following steps: S301, a preset timing segmentation network, a two-dimensional convolutional neural network and a three-dimensional convolutional neural network are adopted to build a deep learning ash prediction model; S302, based on the deep learning ash prediction model, the time sequence feature extraction is carried out on the image data set, and the deep learning ash prediction model is supervised and trained in combination with the actual ash label, and the prediction result is obtained; S303, the deep learning ash prediction model is verified and tested by using the verification set and the test set, and is evaluated by using the preset evaluation method, and the classification effect evaluation result is obtained; S304, in response to the classification effect evaluation result meeting the preset evaluation requirement, the prediction result output by the trained deep learning ash prediction model is taken as the real-time prediction result; S305, based on the sorting equipment including a water medium cyclone and a spiral separator, the operating parameters including the flow rate of the automatically adjusted water medium cyclone, the cyclone classification pressure, the separation angle of the spiral separator and the feed flow rate are dynamically adjusted according to the real-time prediction result, and the sorting equipment after preliminary adjustment is obtained.

[0012] The beneficial effects of the above further scheme are: the deep learning ash prediction model is constructed by the time sequence segmentation network, the two-dimensional convolutional neural network and the three-dimensional convolutional neural network, the accurate detection of product ash is realized, the prediction result is obtained through classification effect evaluation, and the preliminary dynamic adjustment is completed through dynamic adjustment of operating parameters, and the adjustment in the ash prediction aspect is realized.

[0013] Further, the S4 comprises the following steps: S401, according to the sorting equipment after preliminary adjustment, the real-time monitoring data including the feed size distribution, the overflow yield and the ash data are collected, the ideal sorting process is simulated by using the classification efficiency curve, the theoretical sorting model is constructed, and the theoretical sorting efficiency and the underflow theoretical ash are calculated; S402, the underflow theoretical yield is calculated by using the size distribution and the theoretical sorting efficiency, and the comparative theoretical sorting efficiency is calculated according to the underflow theoretical ash in combination with the feed ash; S403, the solid concentration data are obtained by monitoring the overflow and underflow solid concentration, and the actual underflow yield is calculated in combination with the mass balance; S404, according to the actual underflow yield, the actual sorting efficiency is obtained by calculating the actual residual carbon recovery rate and the actual ash recovery rate in combination with the ash data; S405, according to the comparative theoretical sorting efficiency and the actual sorting efficiency, the deviation of the comparative theoretical sorting efficiency and the actual sorting efficiency is calculated; S406. In response to the deviation between the theoretical and actual sorting efficiency exceeding a preset threshold, the classification pressure of the hydrocyclone, the sorting angle of the spiral separator, or the feed flow rate are automatically adjusted by minimizing the deviation, so as to keep the ash content in the sorting process within a predetermined range, thereby obtaining a dynamically adjusted sorting device and completing the dynamic control of sorting.

[0014] Furthermore, the calculation expression for the theoretical sorting efficiency is as follows: ; in, This represents the theoretical sorting efficiency. Indicates the first i Feed mass fraction at each particle size level Indicates the first i The theoretical recovery rate of each particle size class entering the underflow under ideal sorting conditions.

[0015] Furthermore, the calculation expression for the comparative theoretical sorting efficiency is as follows: ; ; ; in, This represents the theoretical sorting efficiency. This represents the theoretical residual carbon recovery rate. This represents the theoretical ash recovery rate. Indicates the theoretical yield of the bottom flow. This indicates the ash content of the bottom flow theory. This indicates the ash content of the feed material.

[0016] Furthermore, the calculation expression for the actual sorting efficiency is as follows: ; ; ; in, Indicates the actual sorting efficiency. This indicates the actual residual carbon recovery rate. This indicates the actual ash recovery rate. Indicates the actual miscarriage rate. This indicates the ash content of the bottom flow.

[0017] The beneficial effects of the above-mentioned further solutions are as follows: By calculating the theoretical sorting efficiency and the actual sorting efficiency, and combining the difference for dynamic control, the present invention further improves the effectiveness of control based on ash content prediction. Attached Figure Description

[0018] Figure 1 Flow chart of the method of the present application.

[0019] Figure 2 Three-dimensional convolutional neural network diagram in the present embodiment.

[0020] Figure 3 Time series convolutional network structure diagram in the present embodiment.

[0021] Figure 4 Three-dimensional convolutional neural network model training curve in the present embodiment.

[0022] Figure 5 Time series convolutional network model training curve in the present embodiment. DETAILED DESCRIPTION

[0023] The specific embodiments of the present application are described below to facilitate understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all applications utilizing the concept of the present application are within the scope of protection.

[0024] Before explaining the present embodiment, the following terms are explained: Tromp curve: grading efficiency curve; I3D neural network: three-dimensional convolutional neural network model independently built in the present embodiment; TSN: time series convolutional network; Inception: the name of a deep learning model architecture.

[0025] EMBODIMENT The present application proposes a dynamic control method for coal gasification slag ash content prediction based on deep learning, which aims to predict the ash content in real time through a deep learning model, and intelligently adjust the operating parameters of the separation equipment based on the prediction results, so as to optimize the separation effect and improve the separation efficiency of coal gasification slag, while reducing energy consumption; It relates to real-time monitoring of mineral slurry video image data in the coal gasification slag separation process using a deep learning prediction model, accurately predicting the ash content of the separation product through a deep learning prediction model, and dynamically adjusting the separation system (including a water medium cyclone and a spiral separator) according to the prediction results to optimize the separation process and improve resource utilization; the present application can be widely used in the efficient recovery and utilization of coal gasification slag, ore separation and other coal resources, and has the advantages of significant energy saving and emission reduction, improved separation precision and intelligent control.

[0026] As Figure 1As shown, the present application provides a dynamic regulation method for separation of coal gasification slag ash content prediction based on deep learning, and the implementation method is as follows: S1, based on the process of coal gasification slag gravity separation, collect original video data and actual ash content label, the specific steps are as follows: S101, based on the process of coal gasification slag gravity separation, in response to the stable flow of ore pulp, use the preset video monitoring device to collect at the preset time and the preset acquisition frame rate, obtain the original video data; S102, collect the ore pulp in the original video data collection time period, and perform filtration and drying treatment on the collected ore pulp, obtain the treated ore pulp, and perform ash content test and detection on the treated ore pulp, through determination, obtain the actual ash content label.

[0027] In this embodiment, for the gravity separation process of coal-based solid waste coal gasification slag, a high-speed camera or an optical imaging device is used to collect ore pulp video data in the process of coal gasification slag gravity separation in real time, to obtain the flow of ore pulp, material movement and the appearance characteristics of different outlet products; Connect the power supply, computer, industrial camera, adjust the initial set camera parameters, open the coding software for controlling shooting, connect the discharge port to the inclined water flow slope, so that the ore pulp is uniformly and stably conveyed to the inclined water flow slope device, the ore pulp flows into the storage tank, the storage tank overflows to the buffer area, and then stably flows into the shooting area. At this time, start shooting and collecting, set the preset time of each video to 15 seconds, the preset acquisition is 40 frames per second, and the original video data is collected. When the shooting and collection are finished, close the feeder, place a sample collection tank under the inclined water flow slope, collect the ore pulp in the original video data collection time period, and send it to the laboratory for ash content test and detection. Through determination, the actual ash content label corresponding to the video segment is obtained.

[0028] In this embodiment, the video monitoring device is specifically a high-speed camera, an optical imaging device or an infrared imaging device, which can capture the flow and appearance characteristics of coal gasification slag in the separation process in real time.

[0029] S2, pre-process the original video data, and obtain the image data set by randomly extracting video frame pictures, the specific steps are as follows: S201, denoise and enhance the original video data, and divide the original video data into time segments according to the preset division frame rate, to obtain the pre-processed short video data; S202, divide the pre-processed short video data into training set, validation set and test set by hierarchical sampling; S203, randomly extract video frame pictures in the training set to obtain the image data set.

[0030] In this embodiment, the original video data is preprocessed by denoising, enhancing and frame extraction, etc. to ensure the image quality, and the key features in the ore pulp flow, material movement and separation process are extracted through the preprocessing. In order to effectively process long video data, the original video data (long video) collected is processed by using a video clipping method or a key video frame representing the whole video method. The long video with a time length of 15 seconds and a total of 600 frames is clipped into short videos with 20 frames each, and the preprocessed short video data is obtained. The preprocessed short video data is divided into training set, validation set and test set by using a hierarchical sampling method. According to the training set, the ore pulp image data set of the separation product is made by randomly extracting video frames from the ore pulp video in the training set.

[0031] S3, a deep learning ash content prediction model is built, and the image data set is used to train the deep learning ash content prediction model combined with the actual ash content label. Through verification, testing and evaluation, real-time prediction results are obtained, and a preliminarily adjusted separation equipment is obtained. The specific steps are as follows: S301, a preset timing segmentation network, a two-dimensional convolutional neural network and a three-dimensional convolutional neural network are used to build a deep learning ash content prediction model.

[0032] In this embodiment, as shown in Figure 2 and Figure 3 , the deep learning ash content prediction model is one or more of a timing segmentation network (TSN), a two-dimensional convolutional neural network (2D CNN) or a three-dimensional convolutional neural network (3D CNN). As shown in Figure 2 , the I3D model network is independently built as a deep learning ash content prediction model, which specifically includes: a first convolutional layer, a first max pooling layer, a second convolutional layer, a third convolutional layer, a second max pooling layer, a first Inception module, a second Inception module, a third max pooling layer, a fourth Inception module, a fifth Inception module, a sixth Inception module, a seventh Inception module, a fourth max pooling layer, an eighth Inception module, a ninth Inception module, an average pooling layer and a fourth convolutional layer connected in sequence. The first convolutional layer has a convolution kernel of with a step of 2. The second convolutional layer has a convolution kernel of , the third convolutional layer has a convolution kernel of , and the fourth convolutional layer has a convolution kernel of . The first max pooling layer has a kernel of The step size is 1, 2, 2; the second max-pooling layer kernel is The step size is 1, 2, 2; the third max-pooling layer kernel is The step size is 2; the fourth maximum pooling layer kernel is The step size is 2; the average pooling layer kernel is ; The receptive fields of the first max pooling layer and the second convolutional layer are set to 7, 11, 11; the receptive fields of the second max pooling layer and the first Inception module are set to 7, 11, 11; the receptive fields of the fourth max pooling layer and the eighth Inception module are set to 7, 11, 11. The first, second, fourth, fifth, sixth, seventh, eighth, and ninth Inception modules all have the same structure and specifically include: The fifth max-pooling layer, the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer are connected to the previous layer; the eighth convolutional layer is connected to the fifth max-pooling layer; the ninth convolutional layer is connected to the fifth convolutional layer; and the tenth convolutional layer is connected to the sixth convolutional layer. The seventh, eighth, ninth, and tenth convolutional layers are connected to the splicing layer and output to the next layer.

[0033] S302. Based on the deep learning gray prediction model, the prediction results are obtained by extracting temporal features from the image dataset and combining them with the actual gray labels.

[0034] In this embodiment, the model is trained based on the method of using video cropping or key video frames to represent the whole video. The deep learning gray prediction model is trained using an image dataset. The deep learning method is used to capture the temporal information and spatial correlation in the video. The image dataset is trained on a self-built 3D convolutional neural network model (I3D neural network) or a temporal convolutional network (TSN) prediction model. By extracting temporal features from the image dataset and combining them with actual gray labels, the deep learning gray prediction model is trained under supervision to obtain the prediction results.

[0035] In this embodiment, the self-built temporal convolutional network model adopts a multi-segment temporal learning method. It segments the video data into multiple time segments, extracts features from each segment, and combines these extracted features to generate a global feature vector for grayscale prediction. Specifically: This method classifies videos in an image dataset using TSN (Time Sequencing Network). By dividing the video into multiple shorter segments, and then extracting and classifying features from each segment individually, it models the entire video over a long period. The network significantly reduces computational complexity and storage requirements by sparsely sampling temporal information during training—selecting only a subset of video segments for learning. Specifically, each video segment generates a classification result, which is then predicted using a formula shown below: ; in, This indicates temporal convolution processing. This indicates the video segments that have been divided. Indicates video clip Feature extraction, This represents a combination function that combines features from multiple video segments. This represents the prediction function. In this embodiment, the Softmax function is used as the prediction function, and its specific expression is shown below: ; in, Representing various features in a video clip, This represents the total number of features in a video clip. Indicates the first i One characteristic, Indicates the first j Each feature, through a prediction formula, yields the probability that each segment belongs to a certain behavioral category, thus obtaining the prediction result.

[0036] S303. The deep learning gray prediction model is validated and tested using the validation set and the test set, and evaluated using the preset evaluation method to obtain the classification performance evaluation results. S304. In response to the classification effect evaluation result meeting the preset evaluation requirements, the prediction result output by the trained deep learning gray prediction model is used as the real-time prediction result. S305. Based on the sorting equipment including a hydrocyclone and a spiral separator, the operating parameters including the flow rate of the hydrocyclone, the hydrocyclone classification pressure, the sorting angle of the spiral separator, and the feed flow rate are dynamically adjusted according to the real-time prediction results to obtain the sorting equipment after preliminary adjustment.

[0037] In this embodiment, the deep learning gray prediction model is validated and tested using a validation set and a test set, and evaluated using a preset evaluation method to obtain the classification performance evaluation results. The preset evaluation method is to use metrics such as confusion matrix, accuracy, precision, recall, and mean absolute error to evaluate the model's classification performance. The evaluation of the model classification performance includes, but is not limited to, the following steps: First, accuracy evaluation: calculate the accuracy of the model's prediction results with respect to the actual labels to evaluate the model's classification ability; Second, recall and precision: for gray classification prediction, recall and precision can be used to evaluate model performance; Finally, real-time evaluation: evaluate the time required for the model to process a single frame to ensure that it meets real-time requirements. When the classification performance evaluation results meet the preset evaluation requirements, the prediction results output by the trained deep learning gray prediction model will be used as the real-time prediction results. Based on the prediction results, the operating parameters of the hydrocyclone and spiral separator are dynamically adjusted, including the feed flow rate, the hydrocyclone's classification pressure, and the separator's separation angle and speed, to optimize the separation effect of coal gasification slag, improve the ash recovery rate, and reduce energy consumption. The dynamic adjustment step includes automatically adjusting the flow rate of the hydrocyclone, the hydrocyclone's classification pressure, the spiral separator's separation angle, and the feed flow rate, and adjusting the above parameters according to the real-time prediction results of the ash content, thereby optimizing the separation accuracy and resource utilization rate of the coal gasification slag, completing the initial adjustment, and obtaining the initially adjusted separation equipment.

[0038] S4. Based on the preliminarily adjusted sorting equipment, collect real-time monitoring data. Based on the real-time prediction results, calculate the deviation between the theoretical and actual sorting efficiency, and automatically adjust the operating status of the sorting equipment to complete the dynamic control of sorting. The specific steps are as follows: S401. Based on the preliminarily adjusted sorting equipment, collect real-time monitoring data including feed particle size distribution, overflow yield and ash content data, and simulate the ideal sorting process by using the grading efficiency curve to construct a theoretical sorting model and calculate the theoretical sorting efficiency and underflow theoretical ash content. S402. Calculate the theoretical underflow yield using particle size distribution and theoretical separation efficiency, and calculate the comparative theoretical separation efficiency based on the theoretical underflow ash content and the feed ash content. S403. By monitoring the solid concentration in the overflow and underflow, solid concentration data is obtained, and the actual underflow yield is calculated in conjunction with the mass balance. S404. Based on the actual bottom flow rate and combined with ash content data, the actual residual carbon recovery rate and actual ash recovery rate are calculated to obtain the actual sorting efficiency. S405. Based on the comparison of theoretical sorting efficiency and actual sorting efficiency, calculate the deviation between the comparison of theoretical sorting efficiency and actual sorting efficiency; S406. In response to the deviation between the theoretical and actual sorting efficiency exceeding a preset threshold, the classification pressure of the hydrocyclone, the sorting angle of the spiral separator, or the feed flow rate are automatically adjusted by minimizing the deviation, so as to keep the ash content in the sorting process within a predetermined range, thereby obtaining a dynamically adjusted sorting device and completing the dynamic control of sorting.

[0039] In this embodiment, the theoretical sorting efficiency is calculated based on real-time monitoring of feed particle size, overflow yield, and ash content. This calculation is then compared with the actual sorting efficiency. The sorting parameters are dynamically adjusted based on the comparison results to obtain a dynamically adjusted sorting device, thus completing the dynamic control of the sorting process. Specifically: Theoretical sorting efficiency calculation: Based on real-time monitoring of feed particle size distribution, overflow yield, and ash content data, the theoretical sorting efficiency is calculated using a theoretical sorting model. The theoretical sorting model considers the influence of feed particle size distribution on sorting behavior and uses a tromp curve to simulate the ideal sorting process. The theoretical sorting efficiency is calculated within the theoretical sorting model. The formula is: ; in, This represents the theoretical sorting efficiency. Indicates the first i Feed mass fraction at each particle size level Indicates the first i The theoretical recovery rate of each particle size class entering the underflow under ideal sorting conditions is obtained from the preset Tromp curve by using the theoretical cut particle size (d50) and particle size distribution parameters. The theoretical undercurrent yield was calculated using particle size distribution and theoretical sorting efficiency. And based on theoretical sorting efficiency The theoretical ash content of the bottom flow was calculated. ; At the same time, combined with the ash content of the feed And the theoretical sorting efficiency Calculated bottom flow theoretical ash content Calculate the theoretical residual carbon recovery rate and theoretical ash recovery rate The calculation expression is as follows: ; ; in, This represents the theoretical undercurrent yield calculated based on particle size distribution and theoretical sorting efficiency; Finally, the comparative theoretical sorting efficiency is obtained for comparison with the actual efficiency. Specifically, it is expressed as: ; Actual sorting efficiency calculation: Based on real-time monitoring of overflow and underflow solid concentrations, the actual sorting efficiency is calculated, and the actual underflow yield is calculated using mass balance and solid concentration data. The actual residual carbon recovery rate was calculated by combining the ash content data. and actual ash recovery rate The calculation expression is as follows: ; ; in, Indicates bottom flow ash content, Indicates the ash content of the feed material; The formula for calculating the actual sorting efficiency is shown below: ; in, Indicates the actual sorting efficiency; The deviation between the theoretical and actual sorting efficiency is calculated, and the calculation expression is shown below: ; in, This indicates the deviation between the theoretical sorting efficiency and the actual sorting efficiency. like If the preset threshold is exceeded, the separation process is optimized by minimizing the deviation. By automatically adjusting the grading pressure of the hydrocyclone, the separation angle of the spiral separator, or the feed flow rate, the ash content in the separation process is kept within a predetermined range, resulting in a dynamically adjusted separation device, thus completing the dynamic control of the separation.

[0040] In this embodiment, the parameters of the sorting system are continuously adjusted by collecting ash data in real time to ensure that the ash content remains stable within a predetermined range, thereby improving the resource recovery rate and sorting stability.

[0041] In this embodiment, the sorted material is coal gasification slag, or industrial solid waste with particle size differences, including but not limited to fly ash, coarse coal slime, coal slurry, or mineral processing tailings.

[0042] In this embodiment, the coal gasification slag gravity separation process includes, but is not limited to, hydrocyclone grading and spiral separation of coal gasification slag. It is also applicable to other material processing scenarios with obvious particle size distribution characteristics and solid-liquid separation requirements, such as the separation process of fly ash, coarse coal slime, coal slime water, and mineral beneficiation tailings. The coal gasification slag gravity separation process includes, but is not limited to, using water hydrocyclones or spiral separators for separation, and is also applicable to combined separation processes that combine water hydrocyclones and spiral separators; any of the above-mentioned separation systems can be used in conjunction with the ash content prediction and feedback control method based on video images and deep learning models provided by this invention to achieve intelligent control of different separation processes.

[0043] In this embodiment, the obtained slurry video dataset is divided as follows: 4892 videos are divided into 3590 training samples, 476 validation samples, and 476 test samples. The specific training process uses the I3D model and SGD optimizer, with an initial learning rate of 10⁻², a learning decay rate of 10⁻⁴, a batch size of 16, and the cross-entropy loss function. The training curve based on the 3D convolutional neural network model is shown below. Figure 4 As shown; from Figure 4 As can be seen, the accuracy and loss value of the 3D convolutional neural network model tend to stabilize after 100 iterations. The accuracy on the training set can reach 98%, and the loss value drops below 0.2, showing good convergence. During training, the model gradually learns the data patterns and classifies the given task more and more accurately. The model successfully reduces the gap between the predicted value and the actual label during the optimization process. The accuracy of the trained model on the test set is 93.35%, which is 16.87% higher than the image classification accuracy. The classification prediction results based on the 3D convolutional neural network model are shown in Table 1 below. Table 1 is a classification evaluation table of the 3D neural network model. It can be seen from Table 1 that the precision, recall, and F1 score of each gray class are at a high level. The I3D model performs excellently in the classification task between different gray partitions and can achieve highly accurate results for each category.

[0044] Table 1

[0045] In this embodiment, the obtained slurry video dataset is divided as follows: 4892 videos are divided into 3590 training samples, 476 validation samples, and 476 test samples. The specific training process uses the TSN model and SGD optimizer, with an initial learning rate of 10⁻², a learning decay rate of 10⁻⁴, a batch size of 16, and the cross-entropy loss function. The video input consists of 15 frames per segment, with one segment and one randomly selected frame from each segment. The training curve based on the temporal segmentation network model is shown below. Figure 5As shown, the TSN network curve tends to stabilize after 40 iterations, with an accuracy of 97% and a loss value of around 0.3. It converges quickly during training and can learn effective feature representations in a relatively short training time. The TSN model, after training, achieves a prediction accuracy of 94.68% on the test set, indicating that the TSN model has achieved good overall performance in the task of classifying slurry videos of sorted products. The TSN model classification performance evaluation is shown in Table 2. Table 2 is the TSN model classification evaluation table.

[0046] Table 2

[0047] In this embodiment, the collected field slurry dataset is systematically validated under actual production conditions. The TSN model is trained, and after 100 iterations, the training set accuracy can reach 95%, the loss value drops to below 0.3, and the model converges quickly. The prediction model is evaluated, such as accuracy, recall and F1 score. The results are shown in Table 3, which shows that it can accurately predict the gray interval.

[0048] Table 3

[0049] In this embodiment, the present invention uses a deep learning prediction model to monitor the slurry video image data in real time during the production process of coal gasification slag gravity separation, accurately predict the ash content and dynamically adjust the separation system parameters, optimize the separation process, improve resource utilization, and is widely used in the separation and industrial application of materials with obvious particle size characteristics such as coal gasification slag.

Claims

1. A dynamic control method for coal gasification slag separation based on deep learning-based prediction of ash content, characterized in that, Includes the following steps: S1. Based on the coal gasification slag gravity separation process, collect raw video data and actual ash content labels; S2. Preprocess the raw video data and obtain an image dataset by randomly extracting video frames; S3. Build a deep learning gray prediction model, and use image datasets and actual gray labels to train the deep learning gray prediction model. Through verification, testing and evaluation, obtain real-time prediction results and obtain a sorting device that has been initially adjusted. S4. Based on the preliminarily adjusted sorting equipment, collect real-time monitoring data, and based on the real-time prediction results, automatically adjust the operating status of the sorting equipment by calculating and comparing the deviation between the theoretical sorting efficiency and the actual sorting efficiency, thereby completing the dynamic control of sorting.

2. The dynamic control method for coal gasification slag ash content prediction based on deep learning as described in claim 1, characterized in that, S1 includes the following steps: S101. Based on the coal gasification slag gravity separation process, in response to the stable flow of slurry, the raw video data is obtained by using a preset video monitoring device at a preset time and preset acquisition frame rate. S102. Collect the slurry during the original video data acquisition period, filter and dry the collected slurry to obtain the processed slurry, and perform ash content analysis and testing on the processed slurry to obtain the actual ash content label.

3. The dynamic control method for coal gasification slag ash content prediction based on deep learning as described in claim 1, characterized in that, S2 includes the following steps: S201. The original video data is denoised and enhanced, and the original video data after denoising and enhancement is divided into time periods according to the preset frame rate to obtain preprocessed short video data. S202. Using hierarchical sampling, the preprocessed short video data is divided into training set, validation set and test set; S203. An image dataset is created by randomly extracting video frames from the training set.

4. The dynamic control method for coal gasification slag ash content prediction based on deep learning according to claim 1, characterized in that, S3 includes the following steps: S301. A deep learning gray prediction model is obtained by constructing a pre-defined temporal segmentation network, a two-dimensional convolutional neural network, and a three-dimensional convolutional neural network. S302. Based on the deep learning gray prediction model, the deep learning gray prediction model is trained under supervision by extracting temporal features from the image dataset and combining them with the actual gray labels to obtain the prediction results. S303. The deep learning gray prediction model is validated and tested using the validation set and the test set, and evaluated using the preset evaluation method to obtain the classification performance evaluation results. S304. In response to the classification effect evaluation result meeting the preset evaluation requirements, the prediction result output by the trained deep learning gray prediction model is used as the real-time prediction result. S305. Based on the sorting equipment including a hydrocyclone and a spiral separator, the operating parameters including the flow rate of the hydrocyclone, the hydrocyclone classification pressure, the sorting angle of the spiral separator, and the feed flow rate are dynamically adjusted according to the real-time prediction results to obtain the sorting equipment after preliminary adjustment.

5. The dynamic control method for coal gasification slag ash content prediction based on deep learning according to claim 1, characterized in that, S4 includes the following steps: S401. Based on the preliminarily adjusted sorting equipment, collect real-time monitoring data including feed particle size distribution, overflow yield and ash content data, and simulate the ideal sorting process by using the grading efficiency curve to construct a theoretical sorting model and calculate the theoretical sorting efficiency and underflow theoretical ash content. S402. Calculate the theoretical underflow yield using particle size distribution and theoretical sorting efficiency, and calculate the comparative theoretical sorting efficiency based on the theoretical underflow ash content and the feed ash content. S403. By monitoring the solid concentration in the overflow and underflow, solid concentration data is obtained, and the actual underflow yield is calculated in conjunction with the mass balance. S404. Based on the actual bottom flow rate and combined with ash data, the actual residual carbon recovery rate and actual ash recovery rate are calculated to obtain the actual sorting efficiency. S405. Based on the comparison of theoretical sorting efficiency and actual sorting efficiency, calculate the deviation between the comparison of theoretical sorting efficiency and actual sorting efficiency; S406. In response to the deviation between the theoretical and actual sorting efficiency exceeding a preset threshold, the classification pressure of the hydrocyclone, the sorting angle of the spiral separator, or the feed flow rate are automatically adjusted by minimizing the deviation, so as to keep the ash content in the sorting process within a predetermined range, thereby obtaining a dynamically adjusted sorting device and completing the dynamic control of sorting.

6. The dynamic control method for coal gasification slag ash content prediction based on deep learning according to claim 5, characterized in that, The formula for calculating the theoretical sorting efficiency is as follows: in, This represents the theoretical sorting efficiency. Indicates the first i Feed mass fraction at each particle size level Indicates the first i The theoretical recovery rate of each particle size class entering the underflow under ideal sorting conditions.

7. The dynamic control method for coal gasification slag ash content prediction based on deep learning according to claim 6, characterized in that, The calculation expression for the comparative theoretical sorting efficiency is as follows: in, This represents the theoretical sorting efficiency. Indicates the theoretical residual carbon recovery rate. This represents the theoretical ash recovery rate. Indicates the theoretical yield of the bottom flow. This indicates the ash content of the bottom flow theory. This indicates the ash content of the feed material.

8. The dynamic control method for coal gasification slag ash content prediction based on deep learning according to claim 7, characterized in that, The formula for calculating the actual sorting efficiency is as follows: in, Indicates the actual sorting efficiency. This indicates the actual residual carbon recovery rate. This indicates the actual ash recovery rate. Indicates the actual miscarriage rate. This indicates the ash content of the bottom flow.

Images