Intelligent decision method and system for jigging separation
By using an intelligent decision-making system and training a large model with sensor data and manually labeled data, the problem of unstable bed state in traditional mineral sorting has been solved, and efficient and stable mineral sorting results have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
In traditional mineral sorting processes, operators rely on manual experience, which leads to unstable bed condition control, often fluctuating between too tight and too loose, thus reducing jigging sorting efficiency.
An intelligent decision-making system is adopted, which acquires historical sensor data and manually labeled ore particle size distribution data of the mineral jigging process, combines the physical rules of jigging, uses DeepSeek-R1-1.5B as the base model, and fine-tunes it through low-rank adaptive algorithm and reward model to build a large model, so as to realize intelligent analysis and decision-making of jigging conditions.
It significantly reduces reliance on manual labor and regulatory lag, enabling real-time dynamic optimization of bed conditions, improving stratification accuracy, reducing concentrate loss, and enhancing the efficiency and stability of mineral jigging.
Smart Images
Figure CN122114165A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mineral sorting technology, and in particular to an intelligent decision-making method and system for jigging sorting. Background Technology
[0002] Jigging is a core step in the clean utilization of coal. Its essence is to use periodic water flow to cause the ore to be stratified according to density on the screen: high-density minerals sink to the bottom, medium-density minerals are in the middle, and low-density gangue floats to the top.
[0003] In traditional mineral sorting, operators initially select and adjust the coal feed rate based on the characteristics of the incoming ore and past experience. Simultaneously, they manually formulate initial control strategies for air pressure and water flow, taking into account the ore density distribution test results. Afterward, operators must strictly adhere to product discharge regulations, visually observing the product discharge status to determine the bed looseness and thickness. They then manually adjust the pulsation parameters, feed rate, and discharge speed. For example, if an operator observes excessive concentrate discharge from the discharge port, indicating an overly loose bed and concentrate loss, they will immediately reduce the water replenishment by feel, while simultaneously lowering the air pressure and pulsation intensity to compact the bed, and quickly and manually closing the concentrate discharge gate.
[0004] Traditional mineral sorting initial control strategies rely heavily on human experience, resulting in unstable bed condition control, often fluctuating between too tight (inhibiting stratification) and too loose (concentrate loss), which directly reduces jigging efficiency. Summary of the Invention
[0005] Therefore, it is necessary to provide an intelligent decision-making system for jigging to address the aforementioned technical problems.
[0006] This invention provides an intelligent decision-making method for jigging sorting, comprising: Acquire historical sensor data of the mineral jigging process, as well as manually labeled ore particle size distribution data and bed looseness data; By fusing the jigging physical rules with historical sensor data, manually labeled ore particle size distribution data, and bed looseness data, historical structured prompt words are obtained. The large model is trained using historical structured cue words, including: using DeepSeek-R1-1.5B as the base model, freezing the pre-trained weight matrix of the base model using a low-rank adaptation algorithm and adding a low-rank matrix to the linear layer weight matrix of the base model to obtain a preliminary fine-tuned model; performing forward propagation calculations on the historical structured cue words using the preliminary fine-tuned model to obtain jigging data prediction results; using a reward model to score the quality of the jigging data prediction results, and using the score results as reward signals, updating the parameters of the preliminary fine-tuned model in reverse using a proximal policy optimization algorithm to obtain the trained large model; By combining real-time sensor data from the mineral jigging process with the physical rules of jigging, real-time structured prompts are obtained. These real-time structured prompts are then input into a trained large model to obtain jigging data analysis results. Based on these results, decision-making suggestions for the jigging process are generated, including control parameters for water flow rate and bed thickness.
[0007] Optionally, the sensor data includes: particle size distribution and density distribution of the feed ore, water flow rate of the jig water pulse, bed thickness, and buoy height distribution of minerals of different densities in the bed.
[0008] Optionally, by freezing the pre-trained weight matrix of the pedestal model using a low-rank adaptation algorithm and adding a low-rank matrix to the linear layer weight matrix of the pedestal model, a preliminary fine-tuned model is obtained, specifically including: Based on the following formula, the pre-trained weight matrix of the pedestal model is frozen using a low-rank adaptation algorithm, and a low-rank matrix is added to the linear layer weight matrix of the pedestal model: ; in, A This is the first low-rank matrix. B It is the second low-rank matrix. W 0 represents the pre-trained weight matrix, Δ W Incremental weights; The first low-rank matrix is initialized with random Gaussian initialization, and the second low-rank matrix is initialized as a zero matrix so that the incremental weights of the base model during the fine-tuning process are zero, thus obtaining the preliminary fine-tuning model.
[0009] Optionally, it also includes evaluating the initial fine-tuning effect of the base model using the cross-entropy loss function based on the following formula: ; in, L o ss Let cross-entropy be the loss function. These are authentic labels derived from manually labeled ore particle size distribution data and bed looseness data. The predicted probability of jigging data; The training learning rate and gradient accumulation steps of the base model during the initial fine-tuning process are dynamically adjusted based on the changing trend of the cross-entropy loss function curve.
[0010] Optionally, a quality score is assigned to the jigging data prediction results using a reward model, and the quality score is used as a reward signal. The parameters of the initially fine-tuned model are then updated using a proximal policy optimization algorithm to obtain the trained large model, specifically including: Based on the following formula, the forward propagation calculation of historical structured prompts is performed through preliminary fine-tuning of the model to obtain the jigging data prediction results: ; in, A This is the first low-rank matrix. B It is the second low-rank matrix. W 0 represents the pre-trained weight matrix, Δ W Incremental weights x These are structured prompt words. h The prediction results are based on jigging data; The prediction results of the jigging data are scored using a reward model, and a policy optimization objective function is constructed based on the following formula using a proximal policy optimization algorithm: ; in, To optimize the objective function of the strategy, For the expected value of the time step, For probability ratios, This is the estimated value of the dominance function. For hyperparameters; The objective function is optimized by minimizing the gradient descent strategy, and the parameters of the initially fine-tuned model are iteratively updated to obtain the trained large model.
[0011] This invention also provides an intelligent decision-making system for jigging sorting, comprising: The data acquisition module is used to acquire real-time sensor data during the mineral jigging process; The data storage module is used to store real-time sensor data and historical sensor data from the data acquisition module, as well as manually labeled ore particle size distribution data and bed looseness data. The model processing module is used to fuse the physical rules of jigging with real-time sensor data, manually labeled ore particle size distribution data and bed looseness data to obtain real-time structured prompts; the real-time structured prompts are then input into the trained large model to obtain jigging data analysis results; The decision output module is used to generate decision suggestions for the jigging sorting process based on the jigging data analysis results. The decision suggestions include control parameters for water flow rate and bed thickness.
[0012] Optionally, it also includes an interactive control module, which is used to display the real-time sensor data of the data acquisition module and the decision suggestions of the decision output module, as well as to receive the operator's correction parameters for the decision suggestions, and to feed back the decision suggestions after parameter correction and the corresponding decision execution results to the model processing module to continuously optimize the trained large model.
[0013] Optionally, the data acquisition module includes: a particle size sensor, a density sensor, a flow sensor, a bed thickness sensor, and a buoy height sensor; Particle size sensor, used to monitor the particle size distribution of the feed ore; Density sensors are used to monitor the density distribution of the feed ore; A flow sensor is used to monitor the water flow rate of the jig's water flow pulses; Bed thickness sensor, used to monitor bed thickness; A buoy height sensor is used to monitor the buoy height distribution of minerals of different densities in the bed.
[0014] Optionally, the data storage module includes: a time-series database and a relational database; The time-series database is used to store real-time sensor data uploaded by the data acquisition module and historical sensor data during the jigging decision-making process. A relational database used to store manually labeled ore particle size distribution and bed looseness data.
[0015] Optionally, the pre-trained large model is also used to match and infer the buoy height distribution from the buoy height sensor with historical sensor data from the time-series database to obtain control strategies for wind pressure and water flow and suggestions for modifying bed state parameters.
[0016] The intelligent decision-making method and system for jigging provided in this invention have the following advantages compared with the prior art: This invention uses DeepSeek-R1-1.5B as the base model, employs a low-rank adaptive algorithm for lightweight fine-tuning, and updates the model parameters through a reward model and a proximal policy optimization algorithm to construct a dedicated large-scale model adapted to the jigging field. This model can integrate multi-dimensional sensor data and manually labeled data to achieve intelligent analysis and decision-making regarding jigging conditions. By replacing manual experience-based judgment with data-driven approaches, it fundamentally avoids instability caused by manual decision-making in traditional mineral sorting processes, such as excessively tight or loose bed conditions. This technique not only significantly reduces reliance on manual intervention and control lag but also maintains the bed in the optimal loose range through real-time dynamic optimization, thereby significantly improving stratification accuracy and reducing concentrate loss, comprehensively enhancing the efficiency and stability of mineral jigging sorting. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating an intelligent decision-making method for jigging sorting provided in one embodiment; Figure 2 This is a system module diagram of an intelligent decision-making system for jigging sorting provided in one embodiment; Figure 3 This is a diagram illustrating the model fine-tuning process of an intelligent decision-making system for jigging sorting provided in one embodiment. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] This invention provides an intelligent decision-making method for jigging sorting, such as... Figure 1 As shown, the method includes: Historical sensor data from the mineral jigging process, along with manually labeled ore particle size distribution and bed looseness data, are acquired. The jigging physical rules are then fused with the historical sensor data, manually labeled ore particle size distribution data, and bed looseness data to obtain historical structured prompts.
[0020] The large model is trained using historical structured cue words, including: using DeepSeek-R1-1.5B as the base model, freezing the pre-trained weight matrix of the base model through a low-rank adaptation algorithm, and adding low-rank matrices to the linear layer weight matrix of the base model to obtain a preliminary fine-tuned model. Forward propagation calculations are performed on the historical structured cue words using the preliminary fine-tuned model to obtain jigging data prediction results. A reward model is used to assign a quality score to the jigging data prediction results, and the quality score is used as a reward signal. The parameters of the preliminary fine-tuned model are then updated using a proximal policy optimization algorithm to obtain the trained large model.
[0021] Real-time sensor data from the mineral jigging process is combined with jigging physical rules to obtain real-time structured prompts. These prompts are then input into a trained large model to obtain jigging data analysis results. Based on these results, decision-making suggestions for the jigging process are generated, including control parameters for water flow rate and bed thickness.
[0022] Preferably, the sensor data includes: the particle size distribution and density distribution of the feed ore, the water flow rate of the jig water pulse, the bed thickness, and the buoy height distribution of minerals of different densities in the bed.
[0023] Preferably, a preliminary fine-tuned model is obtained by freezing the pre-trained weight matrix of the base model using a low-rank adaptation algorithm and adding a low-rank matrix to the linear layer weight matrix of the base model, specifically including: Based on the following formula, the pre-trained weight matrix of the pedestal model is frozen using a low-rank adaptation algorithm, and a low-rank matrix is added to the linear layer weight matrix of the pedestal model: ; in, A This is the first low-rank matrix. B It is the second low-rank matrix. W0 represents the pre-trained weight matrix, Δ W Incremental weights; The first low-rank matrix is initialized with random Gaussian initialization, and the second low-rank matrix is initialized as a zero matrix so that the incremental weights of the base model during the fine-tuning process are zero, thus obtaining the preliminary fine-tuning model.
[0024] Preferably, the initial fine-tuning effect of the base model is evaluated using the cross-entropy loss function based on the following formula: ; in, L o ss Let cross-entropy be the loss function. These are authentic labels derived from manually labeled ore particle size distribution data and bed looseness data. This represents the predicted probability of the jigging data. The training learning rate and gradient accumulation steps of the base model are dynamically adjusted during the initial fine-tuning process based on the changing trend of the cross-entropy loss function curve.
[0025] Preferably, the prediction results of the jigging data are scored using a reward model, and the quality score is used as a reward signal. The parameters of the initially fine-tuned model are then updated using a proximal policy optimization algorithm to obtain the trained large model. Specifically, this includes: Based on the following formula, the forward propagation calculation of historical structured prompts is performed through preliminary fine-tuning of the model to obtain the jigging data prediction results: ; in, A This is the first low-rank matrix. B It is the second low-rank matrix. W 0 represents the pre-trained weight matrix, Δ W Incremental weights x These are structured prompt words. h This is a prediction result based on jigging data.
[0026] The prediction results of the jigging data are scored using a reward model, and a policy optimization objective function is constructed based on the following formula using a proximal policy optimization algorithm: ; in, To optimize the objective function of the strategy, For the expected value of the time step, For probability ratios, This is the estimated value of the dominance function. This is a hyperparameter.
[0027] The objective function is optimized by minimizing the gradient descent strategy, and the parameters of the initially fine-tuned model are iteratively updated to obtain the trained large model.
[0028] Based on the same inventive concept, embodiments of the present invention also provide an intelligent decision-making system for jigging sorting, such as... Figure 2 As shown, the system includes: The data acquisition module is used to acquire real-time sensor data during the mineral jigging process.
[0029] The data storage module is used to store real-time and historical sensor data from the data acquisition module, as well as manually labeled ore particle size distribution data and bed looseness data.
[0030] The model processing module is used to fuse the physical rules of jigging with real-time sensor data, manually labeled ore particle size distribution data and bed looseness data to obtain real-time structured prompts; the real-time structured prompts are then input into the trained large model to obtain jigging data analysis results.
[0031] The decision output module is used to generate decision suggestions for the jigging sorting process based on the jigging data analysis results. The decision suggestions include control parameters for water flow rate and bed thickness.
[0032] The interactive control module is used to display real-time sensor data from the data acquisition module and decision suggestions from the decision output module, as well as receive correction parameters from the operator for the decision suggestions. It then feeds back the corrected decision suggestions and the corresponding decision execution results to the model processing module to continuously optimize the trained large model.
[0033] Preferably, the data acquisition module includes: a particle size sensor, a density sensor, a flow sensor, a bed thickness sensor, and a buoy height sensor.
[0034] Particle size sensor is used to monitor the particle size distribution of the feed ore.
[0035] Density sensors are used to monitor the density distribution of the feed ore.
[0036] A flow sensor is used to monitor the water flow rate of the jig's water flow pulses.
[0037] Bed thickness sensor, used to monitor bed thickness.
[0038] A buoy height sensor is used to monitor the buoy height distribution of minerals of different densities in the bed.
[0039] Preferably, the data storage module includes a time-series database and a relational database.
[0040] The time-series database is used to store real-time sensor data uploaded by the data acquisition module and historical sensor data during the jigging decision-making process.
[0041] A relational database used to store manually labeled ore particle size distribution and bed looseness data.
[0042] Preferably, the pre-trained large model is also used to match and infer the buoy height distribution of the buoy height sensor with historical sensor data from the time series database to obtain the control strategy for wind pressure and water flow and the modification suggestions for bed state parameters.
[0043] Another specific embodiment of an intelligent decision-making system for jigging sorting is provided: The data acquisition module is used to collect real-time and historical data such as ore characteristic parameters, process operating parameters, and sensor timing signals during the jigging process.
[0044] The data storage module, connected to the data acquisition module, is used to store real-time data, historical data, and manually labeled ore particle size distribution and bed looseness data, allowing the data source to be identified.
[0045] The data processing module connects to the data storage module and has a built-in lightweight, fine-tuned large model. The large model is based on the DeepSeek-R1-1.5B base model and fine-tuned using the LoRA algorithm.
[0046] The decision output module, connected to the data processing module, is used to output intelligent decision suggestions for the jigging process. The decision suggestions include control parameters for water flow and bed thickness, and intuitively display the fluctuation pattern of process parameters within a specified period.
[0047] The interactive control module is connected to the data acquisition module and the decision output module respectively, and is used to realize data visualization and manual parameter correction, so as to output more accurate and professional answers and a more complete thinking process.
[0048] The datasets used for fine-tuning were generated in two ways: one was through manual supplementation, and the other was through a dataset building platform that generated professional content, converted the content into a dataset, and finally exported all the generated datasets in JSON format.
[0049] The process of manually supplementing the database involves manually inputting the rules of experience and cases obtained from engineers or operators in relevant coal preparation plants, based on paper records, verbal communications, or scattered electronic documents, into the corresponding fields in a pre-designed simplified database table structure using a code editor (such as Visual Studio Code, VSC). The entire dataset is then exported in JSON format and added to the database.
[0050] The process of supplementing the database by the dataset construction platform is as follows: Select papers on mineral sorting, output the papers as Markdown mode using a PDF parsing tool (such as MinerU), import the Markdown papers into the dataset construction platform (such as Easy Dataset, EDS) to generate multiple text blocks and generate questions in batches, select the corresponding questions suitable for coal preparation plants, observe the correctness of the answers and modify the answers as appropriate, and finally integrate all the questions and answers and export the entire generated dataset in JSON mode.
[0051] The decision output module simultaneously employs LoRa fine-tuning of basic knowledge in the jigging field and RAG retrieval function. The reasoning latency of the problem is less than 200 milliseconds, the error of the output decision suggestions in sorting accuracy is less than or equal to 2%, and the matching rate with the optimal strategy exceeds 90%.
[0052] The interactive control module includes a visual interface developed based on PyQt, which can display sensor data curves and model decision suggestions in real time, respond to operator feedback, and use the corrected parameters and results for continuous optimization and learning of the model.
[0053] This invention can dynamically analyze and recommend the best control strategies for wind pressure and water flow based on real-time collected data on buoy height changes, combined with historical operating condition matching and model reasoning. It provides jigging operators with suggestions for modifying bed state parameters, thereby improving mineral sorting efficiency and accuracy.
[0054] This invention utilizes a LoRa fine-tuning model, which can quickly analyze sensor data and dynamically adjust sorting parameters, such as water flow pulse frequency, thereby increasing the clean coal recovery rate by 3%-5%. Furthermore, it integrates intelligent decision-making functions, allowing operators to compare and select the optimal solution from multiple model-provided solutions, improving model accuracy and achieving closed-loop optimization of the entire jigging process.
[0055] Another specific embodiment of the present invention is provided: This invention comprises four parts: a data acquisition module, a data processing module, a decision output module, and an interactive control module.
[0056] The software's homepage displays a visual page built using a local database of the factory.
[0057] The top card displays real-time statistics of key jigging process parameters, including average feeding frequency, belt scale detection volume, etc., helping operators quickly grasp the production overview. The feeding trend analysis chart, through bar charts of feed rate and belt scale detection and line charts of feeding frequency, can show the fluctuation pattern of process parameters within a specified period, helping to identify changes in production rhythm and predict equipment load, such as the synergistic relationship between feed rate and feeding frequency. The buoy height distribution is displayed using stacked bar charts to show the buoy heights in different sections of middlings and gangue, reflecting the looseness of the bed and the stratification effect. The height distribution differences at different times can help guide operators to adjust the air-water system and improve sorting accuracy. The feeding record form stores historical feeding parameters, frequency, belt weighing volume, and feed amount by time. It supports production review for tracing abnormal working conditions, process optimization by comparing the sorting effect of different parameter combinations, and equipment maintenance related to parameter fluctuations and potential faults.
[0058] Another specific embodiment of the present invention is provided: This invention employs local deployment with Ollam calling a customized model, while also supporting remote model calling via OpenAI, and supports the application of mainstream models such as Deepseek and Doubao.
[0059] The local model employs the LoRa algorithm. Based on the inherent low-rank characteristic of large models, LoRa adds a bypass matrix to simulate full parameter fine-tuning, making it one of the most universal and effective fine-tuning methods currently available. Furthermore, it can be effectively combined with other efficient parameter fine-tuning methods. Using this method for 1.5B deepseek fine-tuning, the number of parameters requiring training and updating can be as small as 0.001% of the total parameters in the full fine-tuning, assuming that the weight updates also have low intrinsic rank during the adaptation process.
[0060] For the pre-trained weight matrix The update is constrained by low-rank decomposition. ,in , , and rank During training, the pre-trained weight matrix W 0 is frozen and does not receive gradient updates, while the first low-rank matrix... A Second low-rank matrix B These are trainable parameters. Note W 0 and They share the same input, and their output vectors are summed by coordinates.
[0061] for The modified forward propagation is as follows: This demonstrates parameterization methods, such as... Figure 3 As shown, for A Random Gaussian initialization is used. B Initialize to zero, so that in the early stages of training It is zero. Then... Scale by a coefficient α (α is a constant in r). When using Adam optimization, adjusting the initial scale appropriately has an effect similar to adjusting the learning rate.
[0062] Therefore, setting α directly to the initial r value and not adjusting it further reduces the need for repeated hyperparameter tuning due to changes in r. LoRa fine-tuning addresses the drawback of high hardware requirements for model operation.
[0063] The model's fine-tuning effect is achieved by L o ss Curve observation involves inputting the model's training data. After lateral propagation, the model begins calculating the loss function. The loss function is the cross-entropy loss function (used for classification problems to measure the difference between two probability distributions), and its formula is:
[0064] ; according to L o ss The learning rate and gradient accumulation steps are adjusted by changing the curve trend to continuously improve the model performance.
[0065] After initial model fine-tuning, reinforcement learning fine-tuning is employed, using both a reward model and a reinforcement learning algorithm. The reward model is a classifier that learns human preferences, mapping generated text to a scalar reward value reflecting text quality. For each cue, multiple responses (e.g., four) are generated, and annotators then rank these responses (from best to worst). As a core component of the system's reinforcement learning fine-tuning, it utilizes a lightweight MLP or fine-tuned BERT-based approach to adapt to industrial deployment needs. It can evaluate the quality of large model predictions, output quantified reward signals, and support PPO algorithm optimization of the model. The model outputs a quantitative reward value of -1 to 1 through a weighted logic of "prediction accuracy + physical rule compliance". An error of ≤2% corresponds to 0.8 to 1.0, and an error >10% or violation corresponds to -1 to 0. The training is based on 1000-5000 sets of "prediction result-real label-human score" triplet data, which mainly comes from sensor data of mineral jigging and sorting, and manually labeled ore particle size distribution and bed looseness data, and the corresponding prediction results generated by the model after preliminary fine-tuning.
[0066] The reinforcement learning algorithm is the PPO algorithm, and the objective function of the PPO algorithm is: ; The key to this function is the pruning mechanism and minimization operation, which calculates advantage estimates by collecting data and then updating the policy. The combination of reward model and PPO algorithm allows the model to learn complex human preferences while maintaining the stability and controllability of the training process.
[0067] The model can be accessed from the cloud, where it can be accessed via the OpenAI API. The vLLM inference engine and FlashAttention-2 effectively solve the problems of memory fragmentation and waste, while ensuring the accuracy of the calculation results.
[0068] The core algorithmic breakthrough of vLLM is PagedAttention, an optimization scheme for the inference process of large language models. Its core idea is to divide the key-value cache of each sequence into blocks, each block containing a fixed number of key-value tensors for each token. In this way, the algorithm can more flexibly manage keys and values in GPU memory, achieving efficient memory utilization and sharing.
[0069] Modern GPUs (such as the NVIDIA A100) have extremely high throughput for matrix multiplication in their Tensor Cores, but much lower throughput for non-matrix multiplication FLOPs (such as floating-point operations). FlashAttention-2 reduces non-matrix multiplication FLOPs such as rescaling, boundary checks, and causal masking operations by improving online softmax techniques. This allows the GPU to spend more time on efficient matrix multiplication FLOPs. Through algorithm optimization, improved parallelization strategies, and optimized work partitioning between warps, FlashAttention-2 significantly improves the speed and efficiency of attention computation, making it particularly suitable for processing long sequences.
[0070] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. An intelligent decision-making method for jigging sorting, characterized in that, include: Acquire historical sensor data of the mineral jigging process, as well as manually labeled ore particle size distribution data and bed looseness data; By fusing the jigging physical rules with historical sensor data, manually labeled ore particle size distribution data, and bed looseness data, historical structured prompt words are obtained. The large model is trained using historical structured cue words, including: using DeepSeek-R1-1.5B as the base model, freezing the pre-trained weight matrix of the base model and adding a low-rank matrix to the linear layer weight matrix of the base model through a low-rank adaptation algorithm to obtain a preliminary fine-tuned model; performing forward propagation calculations on the historical structured cue words using the preliminary fine-tuned model to obtain jigging data prediction results; performing quality scoring on the jigging data prediction results using a reward model, using the quality score results as reward signals, and updating the parameters of the preliminary fine-tuned model in reverse using a proximal policy optimization algorithm to obtain the trained large model; Real-time sensor data from the mineral jigging process is combined with jigging physical rules to obtain real-time structured prompts. These prompts are then input into the trained large model to obtain jigging data analysis results. Based on these results, decision recommendations for the jigging process are generated, including control parameters for water flow rate and bed thickness.
2. The intelligent decision-making method for jigging as described in claim 1, characterized in that, The sensor data includes: particle size distribution and density distribution of the feed ore, water flow rate of the jig water pulse, bed thickness, and buoy height distribution of minerals of different densities in the bed.
3. The intelligent decision-making method for jigging as described in claim 1, characterized in that, The process of freezing the pre-trained weight matrix of the base model using a low-rank adaptation algorithm and adding a low-rank matrix to the linear layer weight matrix of the base model to obtain a preliminary fine-tuned model specifically includes: Based on the following formula, the pre-trained weight matrix of the pedestal model is frozen using a low-rank adaptation algorithm, and a low-rank matrix is added to the linear layer weight matrix of the pedestal model: ； in, A This is the first low-rank matrix. B It is the second low-rank matrix. W 0 represents the pre-trained weight matrix, Δ W Incremental weights; The first low-rank matrix is initialized with random Gaussian initialization, and the second low-rank matrix is initialized as a zero matrix so that the incremental weights of the base model during the fine-tuning process are zero, thus obtaining the preliminary fine-tuning model.
4. The intelligent decision-making method for jigging as described in claim 3, characterized in that, It also includes evaluating the initial fine-tuning effect of the base model using the cross-entropy loss function based on the following formula: ； in, L o ss Let cross-entropy be the loss function. These are authentic labels derived from manually labeled ore particle size distribution data and bed looseness data. The predicted probability of jigging data; The training learning rate and gradient accumulation steps of the base model during the initial fine-tuning process are dynamically adjusted based on the changing trend of the cross-entropy loss function curve.
5. The intelligent decision-making method for jigging as described in claim 1, characterized in that, The process involves scoring the prediction results of jigging data using a reward model, using the quality score as a reward signal, and then using a proximal policy optimization algorithm to back-update the parameters of the initially fine-tuned model to obtain the trained large model. Specifically, this includes: Based on the following formula, the forward propagation calculation of historical structured prompts is performed through preliminary fine-tuning of the model to obtain the jigging data prediction results: ； in, A This is the first low-rank matrix. B It is the second low-rank matrix. W 0 represents the pre-trained weight matrix, Δ W Incremental weights x These are structured prompt words. h The prediction results are based on jigging data; The prediction results of the jigging data are scored using a reward model, and a policy optimization objective function is constructed based on the following formula using a proximal policy optimization algorithm: ； in, To optimize the objective function of the strategy, For the expected value of the time step, For probability ratios, This is the estimated value of the dominance function. For hyperparameters; The objective function of the strategy is minimized by gradient descent, and the parameters of the initially fine-tuned model are iteratively updated to obtain the trained large model.
6. An intelligent decision-making system for jigging based on the intelligent decision-making method for jigging according to any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire real-time sensor data during the mineral jigging process; The data storage module is used to store real-time sensor data and historical sensor data from the data acquisition module, as well as manually labeled ore particle size distribution data and bed looseness data. The model processing module is used to fuse the jigging physical rules with real-time sensor data, manually labeled ore particle size distribution data and bed looseness data to obtain real-time structured prompt words; By inputting real-time structured prompts into the trained large model, the jigging data analysis results are obtained. The decision output module is used to generate decision suggestions for the jigging sorting process based on the jigging data analysis results. The decision suggestions include control parameters for water flow rate and bed thickness.
7. The intelligent decision-making system for jigging as described in claim 6, characterized in that, It also includes an interactive control module, which is used to display the real-time sensor data of the data acquisition module and the decision suggestions of the decision output module, as well as to receive the operator's correction parameters for the decision suggestions, and to feed back the decision suggestions after parameter correction and the corresponding decision execution results to the model processing module to continuously optimize the trained large model.
8. The intelligent decision-making system for jigging as described in claim 6, characterized in that, The data acquisition module includes: a particle size sensor, a density sensor, a flow sensor, a bed thickness sensor, and a buoy height sensor; The particle size sensor is used to monitor the particle size distribution of the feed ore; The density sensor is used to monitor the density distribution of the feed ore; The flow sensor is used to monitor the water flow rate of the jig's water flow pulse; The bed thickness sensor is used to monitor the bed thickness; The buoy height sensor is used to monitor the buoy height distribution of minerals of different densities in the bed.
9. The intelligent decision-making system for jigging as described in claim 6, characterized in that, The data storage module includes: a time-series database and a relational database; The time-series database is used to store real-time sensor data uploaded by the data acquisition module and historical sensor data during the jigging decision-making process. The relational database is used to store the manually labeled ore particle size distribution and bed looseness data.
10. The intelligent decision-making system for jigging as described in any one of claims 8 or 9, characterized in that, The pre-trained large model is also used to match and infer the buoy height distribution of the buoy height sensor with historical sensor data from the time-series database to obtain control strategies for wind pressure and water flow and suggestions for modifying bed state parameters.