A method and device for predicting mineral gene index based on improved multi-task learning

Abstract

The purpose of the present invention is to study an iron ore selectivity evaluation model that can quantitatively evaluate the selectivity of iron ore throughout the entire process and guide ore blending production, and to provide an ore gene index prediction method and device based on improved multi-task learning. The method adopts an improved multi-task learning MMoE algorithm, with the ore gene index as the overall target, which is subdivided into ore time index target, ore quantity index target and ore location index target. A genetic selectivity prediction model for ore materials is constructed to effectively track and predict the changes in ore materials in various production processes, more effectively guide ore blending and mineral processing operations, rationally organize mineral processing production, reduce tailings grade on the basis of stabilizing concentrate grade, and at the same time improve equipment efficiency, reduce energy consumption, and create economic benefits.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine mining and dressing, and in particular relates to a method and device for predicting ore gene index. Background Art

[0002] During iron ore mining, limited by the objective work objectives and scenarios, production targets a complex, heterogeneous geological body composed of ore and rock. During exploration, drilling and blasting, loading, crushing, and multiple mining and processing stages, the mined material is continuously crushed and sorted. During production, sorting equipment must maintain a balanced ratio of ore to rock and maintain consistent ore properties. Extensive fluctuations in ore properties can impact continuous production across all stages of the sorting process. Blending and balancing the properties of the mined and processed ore play a crucial role in key production processes, including organizational management, production coordination and optimization, and equipment management. Therefore, before the ore enters each beneficiation stage, its selectivity for different beneficiation processes must be assessed based on its properties, and ore blending must be implemented accordingly.

[0003] Therefore, it is necessary to establish a new iron ore selectivity evaluation model that can quantitatively evaluate the selectivity of iron ore throughout the entire process and guide ore blending and beneficiation production.

[0004] Ore material genes are defined as an information structure for ore materials, used to describe ore material information that continuously changes, accumulates, and improves over time and space. Ore material genes are a five-tuple describing the ore material's time, space, quantity, properties (ore grade, particle size, and the basic composition, structure, and physical properties that enable processability in processes like grinding, magnetic separation, and flotation), and processing node equipment (equipment operating status and parameters). Ore material genes draw data from multiple sources. Within the various mining and processing processes, some information about a particular step in the ore material gene may be unknown or unmeasurable. However, by tracking ore properties, equipment operation, and processing procedures in upstream and downstream processes, some of the ore's properties at that step can be inferred or inferred.

[0005] The ore material processability factors and the processing factors of the processing equipment and processes comprise the double-stranded DNA of the ore material processing process. The process bases located on the process DNA are the basic units of the ore material processing gene. First, the ore material processability factors and the processing factors of the processing equipment and processes comprise the ore material processing factors, and their interrelationships form the ore material processing bases. Second, the processing bases of different types of ore materials interact with each other to form processing base pairs. Third, a certain number of process base pairs (i.e., processing base sequences) constitute the processing gene. The processing gene model can be used for the management and tracking of ore material processing.

[0006] The 2019 public patent "A Method for Extracting and Characterizing Mineral Gene Data" by Beijing Mining and Metallurgical Technology Group Co., Ltd. proposed a method for extracting and characterizing mineral gene data, but did not study how to predict mineral genes through processing process gene models. Summary of the Invention

[0007] The purpose of the present invention is to study an iron ore selectivity evaluation model that can quantitatively evaluate the selectivity of iron ore throughout the entire process and guide ore blending production, and to provide an ore gene index prediction method and device based on improved multi-task learning. The method adopts an improved multi-task learning MMoE algorithm, with the ore gene index as the overall target, which is subdivided into ore time index target, ore quantity index target and ore location index target. A genetic selectivity prediction model for ore materials is constructed to effectively track and predict the changes in ore materials in various production processes, more effectively guide ore blending and mineral processing operations, rationally organize mineral processing production, reduce tailings grade on the basis of stabilizing concentrate grade, and at the same time improve equipment efficiency, reduce energy consumption, and create economic benefits.

[0008] One of the technical solutions of the present invention is a method for predicting mineral gene index based on improved multi-task learning, comprising the following steps:

[0009] 1) Collect genetic data of ore materials and label the data;

[0010] 2) Calculation of mineral gene index;

[0011] The ore gene index model can be expressed as the following nonlinear function:

[0012] M i =f(x1, x2, ..., x i )

[0013] Where M is the mineral gene index, f is a nonlinear function, x1, x2,…, x i Parameters such as the nature, time, quantity, and location of the ore;

[0014] 3) Training data feature mining;

[0015] 4) Using the labeled data and features, we construct the sample set required for ore gene index prediction and establish an improved multi-task learning ore gene index prediction model;

[0016] 5) Use the trained model to predict the genetic index of ore production data, so that the model can be continuously iterated and optimized.

[0017] In the above-mentioned improved multi-task learning-based ore gene index prediction method, in step 1), ore material gene data is collected, including: ore grade, particle size, and relevant data collection of grinding, magnetic separation, and flotation; the data is labeled including: ore time index label, ore quantity index label, and ore location index label, which are manually labeled by domain experts.

[0018] In the above-mentioned improved multi-task learning-based ore gene index prediction method, step 3) training data feature mining includes mining the ore property characteristics, ore quantity characteristics, ore time characteristics, and ore space characteristics of the data in step 1).

[0019] In the above-mentioned improved multi-task learning-based mineral gene index prediction method, after preparing the feature and label data in step 3), the sample set is constructed by association, and the sample set is divided into a training sample set, a validation set, and a test set, and the data is preprocessed; the preprocessing includes normalization of continuous features and processing of categorical features;

[0020] The normalization formula for the continuous feature part is as follows:

[0021]

[0022] Where mean represents the mean of the current continuous feature, std represents the variance of the current continuous feature, and ε is a very small number used to smooth the feature distribution;

[0023] Category feature processing uses an incremental training method; the ID features used in the first training of the model are randomly initialized using the normal distribution supported by the deep learning framework. Subsequent training will read the features extracted from the data saved by the last model from the saved model file (CKPT) to speed up the model convergence.

[0024] In the above-mentioned method for predicting mineral gene index based on improved multi-task learning, the structure of the mineral gene index prediction model based on improved multi-task learning in step 4) includes:

[0025] 4.1) The first layer is the input layer, which receives the initial features of the ore bottom layer;

[0026] 4.2) The second layer is the feature layer extracted from the original data, which represents the information of the ore;

[0027] 4.3) The third layer is the expert network layer. This layer uses a fully connected neural network (DNN) to construct several different experts. Each expert contributes to the final task prediction from a different perspective. Furthermore, a gating network is designed for the expert network, multiplying each expert by a different weight. The gating network weights are continuously updated as the model trains.

[0028] 4.4) The fourth layer is the Tower layer. The number of Towers corresponds to the number of tasks.

[0029] 4.5) The last layer is the output layer. Different tasks output different prediction scores, which are multiplied by their respective task weight coefficients to calculate the final total estimated score;

[0030] 4.6) The labels for model training are the ore time index label, ore quantity index label and ore location index label recorded in the historical data.

[0031] In the above-mentioned improved multi-task learning-based mineral gene index prediction method, step 4) uses an incremental training method to train the improved multi-task learning-based mineral gene index prediction model, and the process includes:

[0032] 1) Daily data: Use data from t-5 as the training set and data from t-4 as the validation set, where t is the current date;

[0033] 2) Incremental model training: The total amount of training data and the proportion of each label are judged. If they do not exceed a certain threshold, the data is considered normal and training is started. Based on the high-traffic model, the model is trained using daily data sets;

[0034] 3) Point selection strategy: Automatically select the checkpoint with the maximum validation accuracy (validate_auc), save and replace the previous day-level model;

[0035] 4) Model push strategy: The validation accuracy of the selected model is compared with the test accuracy of the base model at the validation point. If the former is better, the model is pushed; otherwise, it is not saved or pushed.

[0036] 5) Model parameters: Since the training data for the daily model is small, offline verification requires debugging parameters such as the learning rate, number of iterations, and step size on the previous base model.

[0037] 6) Automatic point selection strategy optimization: imitate manual point selection and force automatic point selection to be after 0.5 steps;

[0038] 7) Fault-tolerance mechanism: Before model training, determine whether the data is normal, including the lower limit of the total daily data volume and the comparison of the proportion of each label with the upper and lower limits of the basic model;

[0039] 8) Incremental training monitoring: The monitoring model is trained using offline data from the past week, sent to the daily model directory, and participates in the push of daily model candidates;

[0040] In the above-mentioned improved multi-task learning-based ore gene index prediction method, step 5) iterating and optimizing the ore gene index prediction model includes:

[0041] 5.1) Use the model trained in step 4 to predict new ores used in the production environment, obtain genetic indexes, and guide mineral processing;

[0042] 5.2) The data obtained from guiding mineral processing will be manually annotated by domain experts as new samples;

[0043] 5.3) Iteratively train the model using new samples until the model reaches the set accuracy;

[0044] 5.4) Use the model that achieves the set accuracy to perform online real-time prediction.

[0045] The second technical solution of the present invention is a mineral gene index prediction device based on improved multi-task learning, which consists of four modules: material gene collection module, feature engineering module, mineral gene index construction model, and model evaluation module;

[0046] 1) Material gene collection module, used to collect material genes;

[0047] 2) Feature engineering mining module, used to mine feature data of collected material genes;

[0048] 3) Mineral Gene Index (MMoE) prediction model construction module, used to train and construct a mineral gene index prediction model;

[0049] 4) Model evaluation module, which is used to comprehensively evaluate the matching degree between the model prediction value and the true value.

[0050] Furthermore, the feature engineering mining module includes:

[0051] The feature storage unit stores the features of ores of different particle sizes in designated folders, sets the storage upper limit of the ore gene index features, and realizes the circular storage of ore features;

[0052] The tag storage unit collects statistics on the ore data every day and stores it in the database. According to the database, you can query the ore of each gene index by "month", "day" or "shift".

[0053] Furthermore, the above prediction device also includes a mineral gene prediction interface visualization module;

[0054] Among them, the ore gene prediction interface visualization module includes:

[0055] Material gene collection unit, which collects material genes from ore streams in industrial production environments;

[0056] The statistical analysis unit collects statistics on the mineral gene characteristic data every day and stores them in the database. The database can be used to query the distribution of mineral genes in each period by "month", "day" or "shift";

[0057] Alarm reminder unit, used to issue alarm reminders for ores with abnormal genetic conditions;

[0058] The data query unit is used to conduct detailed queries and statistical queries on the spatial data distribution of mineral genes in each period according to "month", "day" and "shift";

[0059] Data export unit, used to export statistical analysis results to Excel reports.

[0060] Compared with the prior art, the advantages of the present invention are:

[0061] The present invention proposes a method and device for constructing ore material genes based on multi-task learning, which can accurately track and predict the changes of ore materials in various processes. It is beneficial for mineral processing enterprises to organize production, ore blending and mineral processing operations, fine-tune resource allocation, effectively improve the core competitiveness of enterprises, and have great economic benefits.

[0062] Figures in the specification

[0063] Figure 1 , flow chart of the overall method.

[0064] Figure 2 , a flowchart of training the improved multi-task learning-based mineral gene index prediction model using the incremental training method.

[0065] Figure 3 , mineral gene index prediction device based on improved multi-task learning. DETAILED DESCRIPTION

[0066] Example 1

[0067] To facilitate understanding, several terms involved in the disclosure of the present invention are first explained below.

[0068] 1) Deep Learning: It learns the inherent laws and representation levels of sample data, mainly using a neural network approach, including input layer, hidden layer, and output layer, and gradually transforms the initial low-order features into high-order features. It is a complex machine learning algorithm.

[0069] 2) Multi-task learning: This is a very promising field in machine learning. The goal is to use the useful information contained in multiple learning tasks to help each task obtain a more accurate learner. The present invention mainly achieves this through a gating mechanism and an expert network.

[0070] 3) Daily incremental training: To enable the model to better exploit changes in data distribution, the offline MMoE high-traffic model is corrected using the most recent daily data to improve the model's timeliness.

[0071] 4) Ore gene matching: The main application scenario of the present invention, ore gene matching plays an important role in important production links and processes such as the organization and management of mining and production, production coordination and optimization, and equipment management.

[0072] This embodiment takes a specific mine as an example:

[0073] A method for predicting ore material genetic index based on improved multi-task learning (MMoE) is proposed. This model uses an improved multi-task learning (MMoE) algorithm, with the ore genetic index as the overall objective, subdivided into ore time index, ore quantity index, and ore location index objectives. A training dataset is constructed, and the positive and negative samples of the three tasks are strictly controlled within each task using label weights (label_weight). The loss function weights (loss_weight) are empirically configured to have a ratio of 0.3:0.3:0.4 for ore time index loss:0.3:0.4 for ore location index loss. The optimal linear weighted coefficients for different tasks are searched in a parameter space of 500 groups using a grid search method. The model includes the following steps:

[0074] Step 1: Collect the genetic data of the ore material and label the data;

[0075] Collect genetic data of ore materials, including ore grade, particle size, and related genetic data collection of grinding, magnetic grinding, and flotation; label the data including ore time index label, ore quantity index label, and ore location index label, which are manually labeled by domain experts.

[0076] in:

[0077] The composition structure of the bases of the mineral materials defined in this study can be described and represented in matrix form:

[0078]

[0079] Where:

[0080] Ad / i / k———Attributes of data / information / knowledge;

[0081] Aht——properties of ore material processing equipment;

[0082] Ast——Attributes of ore material processing organization.

[0083] Ore materials are constantly processed during the mining and dressing process. On the one hand, the properties of the ore materials are constantly changing. On the other hand, the properties of the ore materials are constantly expressed during the processing process, and the genetic information of the ore materials is constantly refined and verified. The attribute information is constantly supplemented and improved with the data.

[0084] There are multiple data sources for ore material genes. In the mining and processing processes, some information of the ore material genes in a certain process may be unknown or cannot be directly measured. However, by tracking the ore properties, equipment operation and processing in the upstream and downstream processes, some properties of the ore in that process can be inferred or inverted.

[0085] Collection of genetic data of mineral materials

[0086] The genetic data of ore materials were collected, and the specific data are shown in Table 1.

[0087] Ore material genetic data collection list

[0088]

[0089]

[0090] Specific examples of collected data:

[0091] Measurement of the genetics of mined ore materials

[0092] After blasting, the mineral processing quality inspectors will take samples at each shovel position to test the properties of the ore. The ore material genes are: inspection position, total iron grade, ferrous grade, pipe fineness, number of revolutions, and selectivity index.

[0093] The small laboratory currently collects samples from once a day, Monday through Friday, to one to three times a week. Sampling is done near the shovel site, with representative samples typically taken from five locations. The purpose of sampling is to reflect the ore properties at that location over the next few days. After analysis, the ore properties are entered into the quality inspection system.

[0094] Automobile transport data measurement

[0095] Due to the significant differences in ore properties at different scooping positions, irrational ore blending in mines and the instability of the properties of the incoming ore are the main reasons for the large fluctuations in various production and technical indicators of the beneficiation plant and the relatively passive production. This is not only detrimental to resource recovery and the full utilization of the beneficiation plant's production capacity, but also brings many inconveniences to production control. The GPS truck dispatch system can record the ore blending status of different scooping positions in real time. Based on the ore selectivity data and ore blending ratios at different scooping positions, it can track the genetic status of the ore material entering the coarse crushing after blending.

[0096] The ore of Qidashan Ore Dressing Plant mainly comes from Qidashan Iron Mine and Anqian Ore. The following data are accessed from the "Qidashan GPS Truck Dispatching System": loading time, arrival time, unloading time, shovel position, destination location, and ore weight.

[0097] Raw ore warehouse data detection

[0098] The Qidashan Ore Processing Plant has two ore bins: the old and new. Bin 1 only accepts ore from Qi Mining, which may also come from the auxiliary material plant. Bin 1 has six hoppers; Bin 2 accepts ore from Anqian and has four hoppers, with Bin 3 being the designated hopper. When the #0 belt is activated, Qi Mining ore is transported to Bin 2; otherwise, it is transported to Bin 1. Each hopper in the two bins corresponds to a collection belt, and the ore's movement is tracked by monitoring the start and stop status of the collection belts in the integrated management and control system. The discharge rate of the ore bin can be calculated using a level meter. Crusher 2 corresponds to the new crushing equipment for the ore processing plant, while crusher 1 corresponds to the old crushing equipment.

[0099] Genetic measurement of coarse and medium crushed materials

[0100] The material genes for coarse and medium-fine crushing are: start time, end time, ore weight, ore particle size, and crushability. During the coarse and medium-fine crushing process, the ore particle size changes. The medium-fine crushing belt scale reweighs the ore and mixes the different ores.

[0101] (1) Particle size detection

[0102] Ore particle size is one of the main factors affecting crusher operation. In order to track the ore particle size, the developed ore particle size image analysis system is used. A camera is installed on the belt conveyor to collect real-time images of the ore on the belt, which are then transmitted to the server through the ore Internet of Things for remote online analysis of the ore particle size.

[0103] The camera before coarse crushing is located at the entrance of the coarse crusher. It mainly collects the particle size of the ore before crushing. The particle size levels are +1500, -1500 to +500, and -500, and the proportion of them is counted.

[0104] Cameras after coarse crushing are located on the No. 2 belt and the saddle belt. The No. 2 belt camera captures the particle size after coarse crushing, while the saddle belt camera captures the particle size after coarse crushing. The particle size levels are divided into three levels: +300, -300, +44, and -44, and the percentage of each is calculated.

[0105] The cameras after medium and fine crushing are located on the 9# and new 9# belts, which monitor the ore particle size of the old medium and fine crushing and the new medium and fine crushing respectively. The particle size levels are divided into three levels: +24, -24 ~ +12, -12.

[0106] (2) Ore weight measurement

[0107] Calculate the ore weight based on the records of 9#, new 9#, 4#, and new 4# belt scales in the "Integrated Management and Control System."

[0108] (3) Crushing properties of ore

[0109] In mineral processing plants, the vast majority of electrical energy is used to crush ore. Understanding the crushability of materials provides a deeper understanding of the crushing process, allows evaluation of the efficiency of crushing machinery, and helps identify more effective crushing methods. Currently, the rock and mineral strength classification commonly used in my country is based on the Proctor hardness scale, denoted by f, which has 10 grades. The f value ranges from 0.3 to 20. The larger the f value, the stronger and more difficult the rock and mineral to crush. This is also reflected in the work index measurement. Basalt, classified as the strongest rock, has an f value of 20. Similarly, its work index is as high as 20.41 kWh / short ton. This clearly demonstrates the consistency between rock and mineral crushability and the work index.

[0110] It can be generally considered that: ores with a work index of less than 10 are brittle ores; ores with a work index of 10-14 are moderately brittle ores; and ores with a work index greater than 14 are collectively referred to as difficult-to-break ores.

[0111] Ore crushability Pugh hardness (f) Crushing work index Wi brittle ore <8 <10 Medium friable ore 8-16 10-14 Refractory ores 16-20 >14

[0112] Powder ore silo measurement

[0113] Genetic measurement of abrasive materials

[0114] The quality inspection center will test a number of indicators during the grinding process, which can be obtained through the "Quality Measurement System". The main indicator parameters are:

[0115] The analysis indicators are: ferrous grade, mixed magnetic fine grade, heavy fine north 1 grade, heavy fine north 2 grade, heavy fine south grade, heavy fine comprehensive grade, strong magnetic tail grade, sweep magnetic tail 1# grade, sweep magnetic tail 2# grade, sweep magnetic tail 3# grade, sweep magnetic comprehensive magnetic tail grade, coarse and fine overflow particle size 1#, coarse and fine overflow particle size 2#, coarse and fine overflow particle size 3#, coarse and fine overflow particle size comprehensive

[0116] Ore grade, comprehensive magnetic tail, primary particle size 1#, primary particle size 2#, primary particle size 3#, primary particle size 4#, primary particle size 5#, comprehensive primary particle size, overflow particle size 1#, overflow particle size 2#, overflow particle size 3#, comprehensive overflow particle size, ferrous grade, strong magnetic tail 1#, strong magnetic tail 2#, strong magnetic tail 3#, comprehensive strong magnetic tail, sweep tail 1#, sweep tail 2#, sweep tail 3#, comprehensive sweep tail, comprehensive mixed concentrate, 1# mixed magnetic concentrate, 2# mixed magnetic concentrate, 3# mixed magnetic concentrate.

[0117] Genetic measurement of flotation materials

[0118] The quality inspection center will test multiple indicators during the flotation process, which can be obtained through the "Quality Measurement System".

[0119] The test indicators include: flotation concentrate grade 1#, flotation concentrate grade 2#, flotation concentrate grade 3#, comprehensive flotation concentrate grade, flotation tailings grade 1#, flotation tailings grade 2#, flotation tailings grade 3#, comprehensive flotation tailings grade, one-sweep concentrate comprehensive, flotation concentrate comprehensive, flotation tailings comprehensive, flotation concentrate 1#, flotation concentrate 2#, flotation concentrate 3#, flotation tailings 1#, flotation tailings 2#, flotation tailings 3#, 1# concentrate, 1# one-sweep concentrate, 2# concentrate, 2# one-sweep concentrate, 3# concentrate, 3# one-sweep concentrate.

[0120] Step 2: Calculate the mineral gene index;

[0121] The ore genes detected in each process are incomplete and require supplementary information from ore genes in other processes. Furthermore, ore genes between processes are interconnected. The ore genes of upstream processes can be used to predict or guide the operation of downstream processes, and the ore genes of downstream processes serve as refinements and feedback to upstream ore genes. Therefore, modeling or the development of standard specifications (such as alignment algorithms for ore genes based on time, location, and weight) is essential to reflect and demonstrate the connections between material genes in each process.

[0122] Calculation of ore gene index:

[0123] The ore gene index model can be expressed as the following nonlinear function:

[0124] M i =f(x1, x2, ..., x i )

[0125] Where M is the mineral gene index, f is a nonlinear function, x1, x2,…, x i Parameters such as the nature, time, quantity, and location of the ore.

[0126] Step 3: Mining training data features;

[0127] Feature engineering mining is performed on the collected ore material genetic data, including ore property characteristics, ore quantity characteristics, ore time characteristics, and ore space characteristics.

[0128] After preparing the feature and label data, we build a training set by association and preprocess the data. Preprocessing includes normalization of continuous features and processing of categorical features.

[0129] For the continuous feature part, we tried min_max normalization and mean variance normalization. Finally, the present invention combined the real data distribution of the ore and used a custom normalization method. The normalization formula is as follows:

[0130]

[0131] Where mean represents the mean of the current continuous feature, std represents the variance of the current continuous feature, and ε is a very small number used to smooth the feature distribution.

[0132] For the embedding of categorical features, we tried random initialization using a truncated normal distribution and experimented with different embedding dimensions (embedding_dim). Ultimately, this paper used incremental embedding training. The ID feature embeddings used in the first training phase of the model were randomly initialized using a normal distribution supported by the deep learning framework. Subsequent training read the saved embeddings from the last model from the saved ckpt file, accelerating model convergence.

[0133] Step 4: Use the labeled data and features to construct the sample set required for ore gene index prediction and establish an improved multi-task learning ore gene index prediction model;

[0134] The structure of the ore gene index prediction model based on improved multi-task learning includes:

[0135] 4.1) The first layer is the input layer, which receives the initial features of the ore bottom layer;

[0136] 4.2) The second layer is the feature layer extracted from the original data, which represents the information of the ore;

[0137] 4.3) The third layer is the expert network layer. This layer uses a fully connected neural network (DNN) to construct several different experts. Each expert contributes to the final task prediction from a different perspective. Furthermore, a gating network is designed for the expert network, multiplying each expert by a different weight. The gating network weights are continuously updated as the model trains.

[0138] 4.4) The fourth layer is the Tower layer. The number of Towers corresponds to the number of tasks.

[0139] 4.5) The last layer is the output layer. Different tasks output different prediction scores, which are multiplied by their respective task weight coefficients to calculate the final total estimated score;

[0140] 4.6) The labels for model training are the ore time index label, ore quantity index label and ore location index label recorded in the historical data.

[0141] The effectiveness of the improved multi-task learning ore gene index prediction model is closely related to the quality of the training data. To enable the model to learn the distribution information of the latest data, the present invention performs incremental training on the MMoE model. The background is to use the recent daily data to correct the offline MMoE high-traffic model to improve the timeliness of the model. This helps the model keep up with the latest data dynamics, including data fluctuations caused by environmental factors. The process includes:

[0142] 1) Daily data: Use data from t-5 as the training set and data from t-4 as the validation set (t is the current date);

[0143] 2) Incremental model training: The total amount of training data and the proportion of each label are judged. If they do not exceed a certain threshold, the data is considered normal and training is started. Based on the large-scale traffic model, the model is trained using daily data sets;

[0144] 3) Point selection strategy: Automatically select the checkpoint with the maximum validation accuracy (validate_auc), save and replace the previous day-level model;

[0145] 4) Push model strategy: Compare the validation accuracy (validate_auc) of the selected model with the test accuracy (test_auc) of the base model on the validation set (validate_set). If the former is better, push the model; otherwise, do not save or push the model.

[0146] 5) Model parameters: Since the training data for the daily model is small, offline verification requires debugging parameters such as the learning rate, number of iterations (validate_step), and epoch size on the previous base model.

[0147] 6) Automatic point selection strategy optimization: imitate manual point selection and force automatic point selection to be after 0.5 steps (epoch);

[0148] 7) Optimization Basis: Problem 1: The model file (model.ckpt-1) was selected at the automatic point selection of the original daily model, which is equivalent to not using new data to revise the model at all. Speculated cause: The daily data is relatively small, and the step size (valid_step) in the model parameters is large, resulting in fewer saved model parameters (ckpt), and the verification results may omit the model file (ckpt) with good performance. The amount of daily data is relatively small, the initial model file (model.ckpt) is already a trained basic model, and the learning rate in the model parameters is relatively large, which may cause the model to fluctuate too much during updates and deviate from the optimal solution. Solution: Explore appropriate model hyperparameters such as learning rate, number of steps, and step size. Observe the locations of manually selected points and let the automatic point selection strategy simulate manual point selection. Offline exploration: Reduce the validation step size (validate_step) to 1 / 4 of the original value, and the number of verifiable model files (model.ckpt) is four times the original value. Increasing the validation step (validate_step) mainly increases the number of model files (model.ckpt) to choose from, without affecting the model file (model.ckpt) itself. Adjusting the learning rate to 1 / 2 of the original value corresponds to increasing the number of epochs to 3 or 4.

[0149] 8) Fault-tolerance mechanism: Before model training, determine whether the data is normal, including the lower limit of the total daily data volume and the comparison of the proportion of each label with the upper and lower limits of the basic model.

[0150] 9) Incremental training monitoring: In order to prevent training drift caused by daily model updates, the present invention has developed a monitoring model for daily updates. The monitoring model is trained using offline data from the past week, sent to the daily model directory, and participates in the push candidate of the daily model.

[0151] Step 5: Use the trained model to predict the genetic index of ore production data, and continuously iterate and optimize the model.

[0152] In this step, if data with insufficient accuracy or prediction errors is found in actual applications, it is screened and expanded into the data set of step 1. The previous steps are repeated until the set accuracy is achieved. The model training is completed and used for online real-time prediction.

[0153] 5.1) Use the model trained in step 4 to predict new ores used in the production environment, obtain genetic indexes, and guide mineral processing;

[0154] 5.2) The offline evaluation indicators of the model in this invention mainly focus on AUC (Area Under Curve, the area under the ROC curve, a professional term used to evaluate model indicators) and loss function. AUC is a comprehensive evaluation of the matching degree between the model prediction value and the true value. Loss is the model loss value calculated according to the scoring function, which can be used to determine whether the model is overfitting.

[0155] 5.3) Training data preprocessing technology, continuously expanding the ability to mine features and labels in the training set;

[0156] 5.4) The selected samples are manually labeled by domain experts;

[0157] 5.5) Iteratively train the model using new samples annotated by domain experts until the model reaches the set accuracy;

[0158] 5.6) Use the trained model for online real-time prediction. The gene index predicts the ore properties (such as grade, particle size, etc.) at a certain time point and location.

[0159] Example 2

[0160] An improved multi-task learning-based ore gene index prediction device consists of five modules: material gene collection module, feature engineering module, ore gene index construction model, model evaluation module, and ore gene prediction interface visualization module.

[0161] in,

[0162] 1) Material gene collection module, used to collect material genes;

[0163] 2) Feature engineering mining module, used to mine feature data of collected material genes;

[0164] The feature engineering mining module includes:

[0165] The feature storage unit stores the features of ores of different particle sizes in designated folders, sets the storage upper limit of the ore gene index features, and realizes the circular storage of ore features;

[0166] The tag storage unit collects statistics on the ore data every day and stores it in the database. According to the database, you can query the ore of each gene index by "month", "day" or "shift";

[0167] 3) Mineral gene index construction module, used to train and construct the mineral gene index prediction module;

[0168] 4) Model evaluation module, used to comprehensively evaluate the matching degree between the model prediction value and the true value;

[0169] 5) Ore gene prediction interface visualization module, used for storage, analysis, data query and alarm notification of ore genes in each period;

[0170] The visualization module of the ore gene prediction interface includes:

[0171] The statistical analysis unit collects statistics on the mineral gene characteristic data every day and stores them in the database. The database can be used to query the distribution of mineral genes in each period by "month", "day" or "shift";

[0172] Alarm reminder unit, used to issue alarm reminders for ores with abnormal genetic conditions;

[0173] The data query unit is used to conduct detailed queries and statistical queries on the spatial data distribution of mineral genes in each period according to "month", "day" and "shift";

[0174] Data export unit, used to export statistical analysis results to Excel reports.

[0175] Example 3

[0176] A mineral gene index prediction system based on improved multi-task learning. The model training module is abstracted into an offline training framework. The core components are as follows:

[0177] 1) The entry point of the entire project: run_dnn.py, which includes two methods: GPU single-machine multi-GPU training and distributed PS training;

[0178] 2) Save the model: Save_online_model.sh, used to save the online model;

[0179] 3) Daily update: update_model.sh, used for automatic update of online models.

[0180] 4) Configuration information: mainly includes model structure, training data, optimizer, GPU information, etc., as follows:

[0181] Network structure (model structure related)

[0182] Model_type: Model type. This invention uses multi-task learning (MMoE).

[0183] Feature_dimention: feature dimension.

[0184] Output_units: The number of output neurons.

[0185] Hidden_units: The number of hidden layer neurons.

[0186] Data feed (read data related)

[0187] Epoch_num: The number of rounds of training data iteration.

[0188] Batch_size: The amount of data in each batch of the training set.

[0189] Test_batch_size: The amount of data in each batch of the test set.

[0190] Valid_batch_size: The amount of data in each batch of the validation set.

[0191] Validate_step: How many steps to save a checkpoint file for validation.

[0192] Total_example_num: The total number of samples in the training data.

[0193] Optimize (optimizer related)

[0194] Optimizer: Optimizer category.

[0195] Dropout: To prevent overfitting, randomly drop the proportion of neurons.

[0196] Is_bn: whether to perform batch normalization.

[0197] Is_dropout: Whether to randomly drop neurons.

[0198] Wnd_wd: L2 regularization coefficient.

[0199] L2_emb_lambda: embedding l2 regularization coefficient.

[0200] Iteration

[0201] Max_iter_step: The maximum number of iteration steps. When all training data are iterated or the number of iteration steps is exceeded, the training stops.

[0202] Learning_rate: Supports changing the learning rate at different time steps.

[0203] Step_boundary: The step at which the learning rate changes.

[0204] GPU (gpu setting information)

[0205] Gpu_visible: The gpu machine we are using has 4 cards.

[0206] Path (path information)

[0207] Output_path: All outputs are in this path.

[0208] Train_data_path: training data path.

[0209] Valid_data_path: validation set data path.

[0210] Test_data_path: test set data information.

[0211] Task_info: multi-task learning parameters

[0212] Label_weight: The weight of different labels when calculating loss. You can weight a certain label.

[0213] Loss_weight: refers to the weight when different task loss functions are weighted.

[0214] Mask_index: refers to the column corresponding to the task in the tfrecord data.

[0215] Export_weight: The score of the multi-task learning model is the weighted sum of the scores of multiple tasks. This parameter is the weight when the scores of different tasks are weighted.

[0216] Export_weight_search: export_weight is generally adjusted based on the overall AUC and offline indicators during the test process. The export_weight parameter search space is relatively large, and manual parameter adjustment is time-consuming.

[0217] export_weight_search is used to set the export_weight parameter to automatically search, 'search_type' is used to specify the search mode, and 'weight_space' is used to specify the search space. Currently, three search modes are supported: 'grid' mode: grid search, for example, in the above example, all possible parameter combinations [10.0,5.0,10.0][10.0,5.0,15.0][20.0,5.0,10.0][20.0,5.0,15.0] will be exhausted. 'specify' mode: Specific combination search is performed. The weight spaces of different tasks are aligned. Therefore, in this mode, the number of elements in the weight spaces of different tasks must be the same. For example, in the following example, all tasks have 2 candidate weights: 'weight_space':{'click':[10.0,20.0],'detail':[5.0,10.0],'order':[10.0,15.0]}, and the parameter combination is [10.0,5.0,10.0][20.0,10.0,15.0]. 'off' mode: No parameter search is performed. By default, the export_weight set in task_info is used for testing.

[0218] The computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device.

[0219] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0220] Program code embodied on a computer-readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0221] Computer program code for performing the operations of the present invention can be written in one or more programming languages, or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0222] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computing device. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Alternatively, they can be implemented using program code executable by a computer device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module. Thus, the present invention is not limited to any specific combination of hardware and software.

[0223] Note that the above are only preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for predicting mineral gene index based on improved multi-task learning, characterized in that: The steps include: 1) Collect genetic data of mineral materials and label the data; The collection of ore material genetic data includes: geological exploration and testing, drilling, blasting, blast pile testing, truck transport particle size, coarse crushing, medium and fine crushing, grinding and magnetic separation, and flotation related data collection; 2) Calculation of mineral gene index; The ore gene index model is expressed as the following nonlinear function: ； Where M is the mineral gene index, f is the nonlinear function, The nature, time, quantity and location parameters of the ore; 3) Training data feature mining; 4) Using the labeled data and features, we construct the sample set required for ore gene index prediction and establish an improved multi-task learning ore gene index prediction model; Among them, the structure of the ore gene index prediction model based on the improved multi-task learning includes: 4.1) The first layer is the input layer, which receives the initial features of the ore bottom layer; 4.2) The second layer is the feature layer extracted from the original data, which represents the information of the ore; 4.3) The third layer is the expert network layer. This layer uses a fully connected neural network to construct several different experts. Each expert contributes to the final task prediction from a different perspective. Furthermore, a gating network is designed for the expert network, multiplying each expert by a different weight. The gating network weights are continuously updated as the model trains. 4.4) The fourth layer is the Tower layer. The number of Towers corresponds to the number of tasks. 4.5) The last layer is the output layer. Different tasks output different prediction scores, which are multiplied by their respective task weight coefficients to calculate the final total estimated score; 4.6) The labels used for model training are the ore time index label, ore quantity index label, and ore location index label recorded in the historical data; 5) Use the trained model to predict the genetic index of ore production data, so that the model can be continuously iterated and optimized.

2. The improved multi-task learning-based ore gene index prediction method according to claim 1 is characterized in that: In step 1), the data is labeled including: ore time index label, ore quantity index label and ore location index label, which are manually labeled by domain experts.

3. The improved multi-task learning-based ore gene index prediction method according to claim 1 is characterized in that: Step 3) Mining the training data features, including mining the ore property features, ore quantity features, ore time features, and ore space features of the data in step 1).

4. The improved multi-task learning-based ore gene index prediction method according to claim 1 is characterized in that: Step 3) After preparing the feature and label data, we associate them to construct a sample set, which is divided into a training sample set, a validation set, and a test set. We also preprocess the data. Preprocessing includes normalizing continuous features and processing categorical features. The normalization formula for the continuous feature part is as follows: ； in, represents the normalized eigenvalue, represents the current eigenvalue, Represents the mean of the current continuous feature, std represents the variance of the current continuous feature, Used to smooth feature distribution; For category feature processing, an incremental training method is used. The ID features used in the first training of the model are randomly initialized using the normal distribution supported by the deep learning framework. After that, the training will read the features extracted from the data saved by the last model from the saved model file to speed up the model convergence.

5. The improved multi-task learning-based ore gene index prediction method according to claim 1 is characterized in that: Step 4) The improved multi-task learning-based mineral gene index prediction model is trained using an incremental training method. The process includes: 1) Daily data: Use data from t-5 as the training set and data from t-4 as the validation set, where t is the current date. 2) Incremental model training: The total amount of training data and the proportion of each label are judged. If they do not exceed a certain threshold, the data is considered normal and training is started. Based on the high-traffic model, the model is trained using daily data sets; 3) Point selection strategy: Automatically select the detection point with the highest verification accuracy, save it, and replace the previous day-level model; 4) Model push strategy: The validation accuracy of the selected model is compared with the test accuracy of the base model at the validation point. If the former is better, the model is pushed; otherwise, it is not saved or pushed. 5) Model parameters: Since the training data for the daily model is small, offline verification requires debugging the learning rate, number of iterations, and step size parameters on the previous base model. 6) Automatic point selection strategy optimization: imitate manual point selection and force automatic point selection to be after 0.5 steps; 7) Fault-tolerance mechanism: Before model training, determine whether the data is normal, including the lower limit of the total daily data volume and the comparison of the proportion of each label with the upper and lower limits of the basic model; 8) Incremental training monitoring: The monitoring model is trained using offline data from the past week, sent to the daily model directory, and participates in the push of daily model candidates.

6. The improved multi-task learning-based ore gene index prediction method according to claim 1 is characterized in that: Step 5) Iteration and optimization of the ore gene index prediction model, including: 5.1) Use the model trained in step 4 to predict new ores used in the production environment, obtain genetic indexes, and guide mineral processing; 5.2) The data obtained from guiding mineral processing will be manually annotated by domain experts as new samples; 5.3) Iteratively train the model using new samples until the model reaches the set accuracy; 5.4) Use the model that achieves the set accuracy to perform online real-time prediction.

7. A device for predicting mineral gene index based on improved multi-task learning, characterized in that: Used to implement the improved multi-task learning-based mineral gene index prediction method according to claim 1, consisting of four modules, namely, a material gene collection module, a feature engineering module, a mineral gene index construction model, and a model evaluation module; 1) Material gene collection module, used to collect material genes; 2) Feature engineering mining module, used to mine feature data of collected material genes; 3) Mineral gene index construction module, used to train and construct the mineral gene index prediction model; 4) Model evaluation module, which is used to comprehensively evaluate the matching degree between the model prediction value and the true value.

8. The improved multi-task learning-based mineral gene index prediction device according to claim 7 is characterized in that: The feature engineering mining module includes: The feature storage unit stores the features of ores of different particle sizes in designated folders, sets the storage upper limit of the ore gene index features, and realizes the circular storage of ore features; The tag storage unit collects statistics on the ore data every day and stores it in the database; the ore with each gene index can be queried according to the database by "month", "day" or "shift".

9. The improved multi-task learning-based mineral gene index prediction device according to claim 7, characterized in that: It also includes a visualization module for the ore gene prediction interface; The ore gene prediction interface visualization module includes: Material gene collection unit, which collects material genes from ore streams in industrial production environments; The statistical analysis unit collects statistics on the mineral gene characteristic data on a daily basis and stores them in the database; the distribution of mineral genes in each period can be queried based on the database by "month", "day" or "shift"; Alarm reminder unit, used to issue alarm reminders for ores with abnormal genetic conditions; The data query unit is used to conduct detailed and statistical queries on the spatial data distribution of mineral genes in each period according to "month", "day" and "shift"; Data export unit, used to export statistical analysis results to Excel reports.

Images