Deep learning-based intelligent identification system for polluted coal sample

Abstract

The invention provides a polluted coal sample intelligent identification system based on deep learning, and belongs to the technical field of polluted coal sample identification, and the system comprises a preprocessing module which obtains image data of a polluted coal sample and a non-polluted coal sample according to a high-resolution camera and carries out the preprocessing; the model training module is used for acquiring a plurality of features of the coal sample and labels corresponding to the image data, and training a deep learning model; the adjusting module is used for evaluating the performance of the trained model according to cross validation and adjusting hyper-parameters of the model; the optimization module is used for predicting the coal sample according to the adjusted model, obtaining the sample probability of prediction errors and influence factors of the prediction errors, and optimizing the model; and the recognition module is used for quickly recognizing the coal sample image acquired in real time according to the optimized model. The problems that manual recognition is needed, the misjudgment rate is high, the accuracy of coal sample quality detection is affected, meanwhile, the data size of a coal sample is too simple, and the applicability of a recognition result cannot be guaranteed are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of contaminated coal sample identification, and in particular to a contaminated coal sample intelligent identification system based on deep learning. Background Art

[0002] With the rapid development of my country's economy, energy demand continues to grow. As one of the main sources of energy, the quality and environmental protection issues of coal are receiving more and more attention.

[0003] However, the traditional method of identifying contaminated coal requires manual identification, which is not only time-consuming and labor-intensive, but also has a high misjudgment rate, affecting the environmental protection and quality detection accuracy of coal samples. At the same time, the amount of data from coal samples is too simple and small, which cannot guarantee the applicability of the identification results and the comprehensiveness of the identification system.

[0004] Therefore, the present invention proposes a contaminated coal sample intelligent identification system based on deep learning. Summary of the invention

[0005] The present invention provides a contaminated coal sample intelligent identification system based on deep learning, which is used to solve the problem that the traditional contaminated coal identification method in the prior art requires manual identification, which is not only time-consuming and labor-intensive, but also has a high misjudgment rate, affecting the environmental protection of the coal samples and the accuracy of quality detection. At the same time, the data volume of the coal samples is too simple and small, and the applicability of the identification results and the comprehensiveness of the identification system cannot be guaranteed.

[0006] In one aspect, the present invention provides a contaminated coal sample intelligent identification system based on deep learning, comprising: Preprocessing module: acquiring image data of polluted coal samples and non-polluted coal samples according to a high-resolution camera, and preprocessing the image data; Model training module: obtaining multiple features of coal samples and labels corresponding to the image data based on the preprocessed image data, and training the deep learning model according to the multiple features of coal samples and labels corresponding to the image data; Adjustment module: evaluate the performance of the trained model based on cross-validation and adjust the model hyperparameters based on the evaluation results; Optimization module: predict the coal samples according to the adjusted model, obtain the sample probability of prediction error and the influencing factors of prediction error, and optimize the model according to the sample probability of prediction error and the influencing factors of prediction error; Identification module: quickly identify the coal sample images collected in real time according to the optimized model, and identify the contaminated coal samples according to the identification results.

[0007] According to the present invention, a contaminated coal sample intelligent identification system based on deep learning, a preprocessing module includes: A first determination unit: obtaining ambient environment parameters of the polluted coal sample and the non-polluted coal sample, and determining a corresponding high-resolution camera according to the ambient environment parameters; A first acquisition unit: acquiring a preset number of image data of contaminated coal samples and non-contaminated coal samples according to the high-resolution camera; Preprocessing unit: preprocess the image data according to multiple filter functions of the PIL library and the OpenCV library.

[0008] According to the present invention, a contaminated coal sample intelligent identification system based on deep learning, a model training module includes: Extraction unit: extract multiple features of coal samples based on convolutional neural network according to preprocessed image data; A second acquisition unit: analyzing the multiple features, and acquiring a label corresponding to the coal sample in each image data according to the analysis result; Model training unit: converts the extracted multiple features and corresponding labels into a unified vector representation, and trains the deep learning model according to the vector representation.

[0009] According to a deep learning-based intelligent identification system for contaminated coal samples provided by the present invention, the extraction unit includes: Get subunits: Use a deep learning framework to build a convolutional neural network model and get multiple components of the convolutional neural network model; Training subunit: Use labeled image datasets to train the convolutional neural network model to obtain the optimal model configuration; Input subunit: input the preprocessed image data into the trained convolutional neural network model to obtain a feature map containing multi-level and multi-scale feature information of the image; Extraction subunit: Multiple feature maps are fused based on the attention mechanism, and multiple features of the coal sample are extracted according to the fusion results.

[0010] According to a contaminated coal sample intelligent identification system based on deep learning provided by the present invention, the adjustment module includes: Evaluation unit: evaluates the performance of the trained model based on K-fold cross validation; Sorting unit: Determine the performance index of the model according to the evaluation results, sort the models according to the performance index, and obtain the best model; Testing unit: After determining the best model, use the model to test on a data set that has not been cross-validated; Adjustment unit: adjust the model hyperparameters according to the test results, determine the best parameter combination through the grid search algorithm, and determine the adjusted model based on the best parameter combination.

[0011] According to the present invention, a contaminated coal sample intelligent identification system based on deep learning and an optimization module are provided, comprising: The third acquisition unit: acquires the actual label of the coal sample, predicts the coal sample according to the adjusted model, acquires the first prediction result, classifies the image according to the first prediction result, and acquires the probability distribution of each pixel in each image in each classification; The fourth acquisition unit: performs non-maximum suppression on the obtained probability distribution, obtains the confidence in each category, and uses the category with the highest confidence as the second prediction result; Comparison unit: comparing the second prediction result with the actual label of the coal sample to obtain a comparison result; The third acquisition unit: if the comparison results are inconsistent, the samples with incorrect predictions are determined according to the confusion matrix calculation, and the number of samples with incorrect predictions and the internal and external influencing factors of each sample with incorrect predictions are obtained; Optimization unit: obtains the probability of prediction errors according to the number of prediction errors, and optimizes the model by combining the internal and external influencing factors of the prediction errors.

[0012] According to the present invention, a contaminated coal sample intelligent identification system based on deep learning is provided, and the identification module includes: Input unit: takes the coal sample image collected in real time as input and inputs it into the optimized model; Post-processing unit: obtaining the predicted value output by the model and performing post-processing on the predicted value; Classification unit: determining the category of each pixel point according to the post-processing result, and classifying the pixel point based on the classification algorithm; Identification unit: quickly identify the coal sample images collected in real time according to the classification results, and identify the contaminated coal samples according to the identification results.

[0013] According to a contaminated coal sample intelligent identification system based on deep learning provided by the present invention, before obtaining image data of contaminated coal samples and non-contaminated coal samples using a high-resolution camera, the system further includes: The first determination module: obtains coal sample images of multiple coal sample types, performs feature analysis on the coal sample images to determine particle size feature parameters and appearance feature parameters of each type of coal sample; The second determination module: determines the state loss of each type of coal sample compared with the standard coal sample according to the particle size characteristic parameters and the appearance characteristic parameters; The third determination module: determines the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss of each type of coal sample according to the state loss; The fourth determination module: determining the volume shape critical difference parameters of each type of coal sample based on the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss; The first acquisition module: determines the image identification and differentiation parameters of each type of coal sample according to the critical difference parameter, and obtains the deep layer image information and the shallow layer image information of each type of coal sample based on the image identification and differentiation parameters; The second acquisition module: acquiring the point spread function of each type of coal sample according to the deep layer image information and the shallow layer image information; The fifth determination module: calculates the protrusion amplitude of each protrusion point of each type of coal sample through the point spread function, and determines the phase complex matrix according to the protrusion amplitude; The third acquisition module: performs non-blind deconvolution calculation on the phase complex matrix to obtain high-resolution image samples of various types of coal samples; The sixth determination module: determines the pixel array of each type of coal sample according to the high-resolution image sample, and determines the qualified judgment conditions of each type of coal sample image based on the pixel array; Judgment module: Based on the qualified judgment conditions, the image data of each type of coal sample, contaminated coal sample and non-contaminated coal sample are judged to be qualified, qualified images are screened out according to the judgment results, and re-collection reminders are issued for unqualified images.

[0014] Compared with the prior art, the present invention has the following beneficial effects: By obtaining the features and corresponding labels of the image data of preprocessed contaminated coal samples and non-contaminated coal samples, and training the deep learning model, the accuracy and robustness of the model can be guaranteed. Furthermore, the model can be optimized and real-time recognition of coal samples can be performed based on the optimized model, which can save a lot of manpower and time, and has a low misjudgment rate, thereby improving the accuracy of environmental protection and quality detection of coal samples. At the same time, the large amount of coal sample data can ensure the diversity of the samples, ensuring the applicability of the recognition results and the comprehensiveness of the recognition system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a structural schematic diagram of a contaminated coal sample intelligent identification system based on deep learning provided by an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a preprocessing module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Embodiment 1: The embodiment of the present invention provides a contaminated coal sample intelligent identification system based on deep learning, such as Figure 1 As shown, the system mainly includes the following modules: Preprocessing module: acquiring image data of polluted coal samples and non-polluted coal samples according to a high-resolution camera, and preprocessing the image data; Model training module: obtaining multiple features of coal samples and labels corresponding to the image data based on the preprocessed image data, and training the deep learning model according to the multiple features of coal samples and labels corresponding to the image data; Adjustment module: evaluate the performance of the trained model based on cross-validation and adjust the model hyperparameters based on the evaluation results; Optimization module: predict the coal samples according to the adjusted model, obtain the sample probability of prediction error and the influencing factors of prediction error, and optimize the model according to the sample probability of prediction error and the influencing factors of prediction error; Identification module: quickly identify the coal sample images collected in real time according to the optimized model, and identify the contaminated coal samples according to the identification results.

[0019] In this embodiment, the high-resolution camera refers to a photographic device with a relatively high image resolution, which can generate high-definition photos or videos during the shooting process.

[0020] In this embodiment, the contaminated coal sample refers to a coal sample that is affected by environmental pollution, such as air pollution, water pollution, and soil pollution.

[0021] In this embodiment, non-polluted coal samples generally refer to coal samples that are not affected by external pollutants. These samples may be collected from unpolluted coal fields, or some measures may be taken during the mining process to reduce the impact of external pollutants.

[0022] In this embodiment, the image data preprocessing is aimed at removing noise, eliminating interference and improving image quality, such as: Denoising: Use various filtering algorithms (such as Gaussian filtering and median filtering) to remove random noise from the image.

[0023] Grayscaling: Converting a color image to a grayscale image can usually be done by subtracting the color component of each pixel.

[0024] Histogram equalization: Adjust the histogram distribution of the grayscale image so that the grayscale levels of all pixels are evenly distributed within a certain range.

[0025] Edge detection: Perform edge detection on image data, such as using the Canny algorithm to determine the edges in an image.

[0026] In this embodiment, the multiple features of the coal sample include: color features, texture features, and shape features.

[0027] In this embodiment, the labels corresponding to the image data include: component, property, classification, and standard.

[0028] In this embodiment, the deep learning model is a computer learning algorithm that can automatically learn and recognize complex patterns, and is usually composed of a neural network composed of multiple levels.

[0029] In this embodiment, cross-validation is a method for evaluating the performance of a machine learning model by dividing the data set into a training set and a test set, then training the model on the training set and testing it on the test set, thereby avoiding overfitting and improving the generalization ability of the model, such as: K-fold cross-validation and leave-one-out cross-validation.

[0030] In this embodiment, performance evaluation is the process of evaluating whether a machine learning model meets certain specific standards or requirements. The machine learning performance evaluation indicators include: Precision: refers to the proportion of correctly classified samples. For example, in coal sample image recognition, if the model can correctly label 90% of non-polluted coal samples as non-polluted coal samples, then its accuracy reaches 90%.

[0031] Recall rate: refers to the proportion of correctly classified samples, which is similar to precision, but does not take into account those samples that are misclassified as negative. For example, in coal sample image recognition, if the model can correctly identify 80% of contaminated coal samples and does not misclassify non-contaminated coal samples as contaminated coal samples, then the recall rate is 80%.

[0032] F1 score: It is a weighted average of precision and recall, and can consider the importance of both precision and recall. For example, if in the sentiment analysis task, the model can correctly identify 80% of the polluted coal samples and correctly classify 70% of the non-polluted coal samples as non-polluted, then its F1 score is 85 points.

[0033] In this embodiment, the model hyperparameters are parameters set when training the machine learning model, which are used to adjust the calculation and operation mode of the model to optimize the performance of the model, such as: regularization coefficient, learning rate, batch normalization smoother, random seed.

[0034] In this embodiment, the probability of incorrectly predicted samples means that, for example, there are a total of 100 coal sample samples, of which 40 are contaminated coal sample samples and 60 are non-contaminated coal sample samples. During the prediction process, 10 contaminated coal sample samples are predicted to be non-contaminated coal sample samples, then the probability of incorrectly predicted samples is 25%.

[0035] In this embodiment, the factors affecting the prediction error may be: model parameter setting problems and unclear samples.

[0036] The beneficial effects of the above technical solution are: by obtaining the features and corresponding labels of the image data of the preprocessed contaminated coal samples and non-contaminated coal samples, and training the deep learning model, the accuracy and robustness of the model can be guaranteed. Furthermore, the model is optimized, and real-time recognition of coal samples is performed according to the optimized model, which can save a lot of manpower and time, and has a low misjudgment rate, thereby improving the accuracy of environmental protection and quality detection of coal samples. At the same time, the large amount of data of coal samples can ensure the diversity of samples, ensure the applicability of the recognition results and the comprehensiveness of the recognition system.

[0037] Embodiment 2: Based on Example 1, the preprocessing module of the embodiment of the present invention is as follows: Figure 2 As shown, including: A first determination unit: obtaining ambient environment parameters of the polluted coal sample and the non-polluted coal sample, and determining a corresponding high-resolution camera according to the ambient environment parameters; A first acquisition unit: acquiring a preset number of image data of contaminated coal samples and non-contaminated coal samples according to the high-resolution camera; Preprocessing unit: preprocess the image data according to multiple filter functions of the PIL library and the OpenCV library.

[0038] In this embodiment, the ambient environmental parameters of the coal sample include: temperature, humidity, oxygen concentration, and carbon dioxide concentration.

[0039] In this embodiment, the preset number refers to 100 samples.

[0040] In this embodiment, the PIL library is an image processing library in Python. The PIL library provides many filter functions to perform various operations on images, such as ImageFilter, blur, and Gaussian filtering.

[0041] In this embodiment, the image data preprocessing is aimed at removing noise, eliminating interference and improving image quality, such as: Denoising: Use various filtering algorithms (such as Gaussian filtering and median filtering) to remove random noise from the image.

[0042] Grayscaling: Converting a color image to a grayscale image can usually be done by subtracting the color component of each pixel.

[0043] Histogram equalization: Adjust the histogram distribution of the grayscale image so that the grayscale levels of all pixels are evenly distributed within a certain range.

[0044] Edge detection: Perform edge detection on image data, such as using the Canny algorithm to determine the edges in an image.

[0045] The beneficial effect of the above technical solution is: obtaining the corresponding high-resolution camera through the surrounding parameters of the coal sample, obtaining a preset number of image data of contaminated coal samples and non-contaminated coal samples, and processing them, which can ensure the quality and accuracy of the data.

[0046] Embodiment 3: Based on Example 2, the model training module of the embodiment of the present invention includes: Extraction unit: extract multiple features of coal samples based on convolutional neural network according to preprocessed image data; A second acquisition unit: analyzing the multiple features, and acquiring a label corresponding to the coal sample in each image data according to the analysis result; Model training unit: converts the extracted multiple features and corresponding labels into a unified vector representation, and trains the deep learning model according to the vector representation.

[0047] In this embodiment, the convolutional neural network is a deep learning algorithm that can automatically extract features and complete various complex pattern recognition and classification tasks. The convolution layer is one of the important components for realizing feature extraction. The function of the convolution layer is to perform dot multiplication of the input data with the local weight matrix and sum them to obtain a feature map.

[0048] In this embodiment, the multiple characteristics of the coal sample include: Physical characteristics: refers to the properties of coal samples that can be directly observed, such as density, hardness, specific surface area, moisture content, shape, and color.

[0049] Chemical characteristics: refers to the characteristics of coal samples that can only be obtained through chemical analysis, such as total moisture, ash, volatile matter, fixed carbon, unsaturated hydrocarbons, and elemental composition.

[0050] Structural characteristics: refers to the microscopic structural characteristics inside the coal sample, such as pore structure, crystal morphology, and degree of crystallization.

[0051] In this embodiment, the labels corresponding to the coal samples include: composition, properties, classification, and standard.

[0052] In this embodiment, in machine learning and artificial intelligence, we often need to convert data into numerical matrices and process them using mathematical operations. This process is called "vectorization". For example, suppose there is a text sample: "This is an anthracite coal". First, we need to determine the position of each word in the text. In this example, "this" is the first word, "is a" is the second word, and "anthracite coal" is the third word. These position information needs to be converted into numerical vectors. These positions can be encoded into a three-dimensional array, where the first dimension represents the position of the word, the second dimension represents the part of speech (0 represents stop words, 1 represents verbs), and the third dimension represents the word order. Therefore, the position of "this" is [1,1,3].

[0053] The beneficial effects of the above technical solution are: extracting and analyzing multiple features of the coal sample based on the preprocessed image data, obtaining the label corresponding to the coal sample, converting it into a vector representation for training the deep learning model, which can improve the generalization ability and robustness of the model, while reducing the complexity and computational cost of the model.

[0054] Embodiment 4: Based on Example 3, the extraction unit of the embodiment of the present invention includes: Get subunits: Use a deep learning framework to build a convolutional neural network model and get multiple components of the convolutional neural network model; Training subunit: Use labeled image datasets to train the convolutional neural network model to obtain the optimal model configuration; Input subunit: input the preprocessed image data into the trained convolutional neural network model to obtain a feature map containing multi-level and multi-scale feature information of the image; Extraction subunit: Multiple feature maps are fused based on the attention mechanism, and multiple features of the coal sample are extracted according to the fusion results.

[0055] In this embodiment, the deep learning framework includes: TensorFlow, PyTorch.

[0056] In this embodiment, the multiple components include: a convolutional layer, a pooling layer, and a fully connected layer.

[0057] In this embodiment, the optimal model configuration refers to a method of finding hyperparameters such as the model architecture, loss function, and optimizer that best suit a specific task in a deep learning task.

[0058] In this embodiment, the multi-level and multi-scale feature information of the image includes: local features and global features.

[0059] In this embodiment, the attention mechanism is a technology that enables the machine to understand the sequence information of the input data. It allocates the attention of the input data to different positions, so as to focus on the most important and relevant information for the current task. For each element, it will calculate a score, which represents the importance of this element. This score is calculated by other elements and is only related to the adjacent elements, not the element itself. In the end, these scores will be used for weighted averaging, such as: self-attention mechanism, dense connection.

[0060] The beneficial effects of the above technical solution are: using a labeled image data set to train the convolutional neural network model to obtain the optimal model configuration, and obtaining a feature map based on the preprocessed image data, thereby obtaining multiple features of the coal sample. It can automatically learn and extract multiple features of the image, avoiding the problems caused by too many features extracted manually. At the same time, real-time image feature recognition can be achieved, laying the foundation for subsequent training of deep learning models.

[0061] Embodiment 5: Based on the fourth embodiment, the adjustment module of the embodiment of the present invention includes: Evaluation unit: evaluates the performance of the trained model based on K-fold cross validation; Sorting unit: Determine the performance index of the model according to the evaluation results, sort the models according to the performance index, and obtain the best model; Testing unit: After determining the best model, use the model to test on a data set that has not been cross-validated; Adjustment unit: adjust the model hyperparameters according to the test results, determine the best parameter combination through the grid search algorithm, and determine the adjusted model based on the best parameter combination.

[0062] In this embodiment, k-fold cross validation is a method commonly used in machine learning training and model evaluation. The data set is divided into k parts, one of which is taken as the test set each time, and the remaining k-1 parts are used as the training set. This process is repeated k times to reduce overfitting and improve the generalization performance of the model.

[0063] In this embodiment, the performance evaluation indicators include: precision, recall, and F1 score.

[0064] In this embodiment, adjusting the model hyperparameters refers to, for example, increasing or decreasing the number of neurons, changing the activation function, and adjusting the weight of the loss function.

[0065] In this embodiment, the grid search algorithm is an algorithm widely used in the fields of machine learning and statistics, and finds the optimal one by enumerating all possible parameter combinations, including: Determine the parameter space: Determine the range of all hyperparameters you want to adjust, usually represented as an array in some way. For example, if you have a logistic regression model, you may want to adjust the hyperparameters including: the width of the positive and negative numbers (such as the slope of the positive and negative numbers), the penalty coefficient (such as the strength of the penalty term).

[0066] Define the objective function: Determine the objective function. Usually, you need to determine whether it is maximum likelihood or minimum error based on the task type (such as classification or regression).

[0067] Define hyperparameter ranges: Define the possible values ​​for each hyperparameter.

[0068] Initialize hyperparameters: During the grid search process, you need to initialize the hyperparameters to a specific range and record the current values ​​of the hyperparameters.

[0069] Run a grid search: For each possible value of the hyperparameter, create a new instance and feed that instance into the model to get the model’s performance.

[0070] In this embodiment, the best parameter combination is the optimal hyperparameter combination found in a grid search or similar optimization algorithm.

[0071] The beneficial effect of the above technical solution is: by evaluating the performance of the trained model and testing it on a data set that has not been cross-validated, and adjusting the model parameters and hyperparameters according to the test results, the characteristics of the model can be more effectively utilized to obtain the best performance and ensure the accuracy of the recognition results of coal samples.

[0072] Embodiment 6: Based on Example 5, the optimization module of this embodiment of the present invention includes: The third acquisition unit: acquires the actual label of the coal sample, predicts the coal sample according to the adjusted model, acquires the first prediction result, classifies the image according to the first prediction result, and acquires the probability distribution of each pixel in each image in each classification; The fourth acquisition unit: performs non-maximum suppression on the obtained probability distribution, obtains the confidence in each category, and uses the category with the highest confidence as the second prediction result; Comparison unit: comparing the second prediction result with the actual label of the coal sample to obtain a comparison result; The third acquisition unit: if the comparison results are inconsistent, the samples with incorrect predictions are determined according to the confusion matrix calculation, and the number of samples with incorrect predictions and the internal and external influencing factors of each sample with incorrect predictions are obtained; Optimization unit: obtains the probability of prediction errors according to the number of prediction errors, and optimizes the model by combining the internal and external influencing factors of the prediction errors.

[0073] In this embodiment, non-maximum suppression is used in target detection tasks to optimize and screen the detection frames. In practical applications, since there are multiple detection frames appearing at the same time, these detection frames need to be screened and the maximum value among them is retained to achieve more accurate target positioning.

[0074] In this embodiment, the confidence of each classification refers to the probability score when the model predicts that the object belongs to different categories for each image, which is usually represented by a value between 0 and 1. A high confidence indicates that the object is very likely to be at this position, while a low confidence indicates that the object is very unlikely to be at this position.

[0075] In this embodiment, confusion matrix calculation refers to using true labels and model prediction results to calculate various error indicators, such as precision, recall, and F1 score.

[0076] In this embodiment, the internal influencing factors of each prediction error sample refer to the factors affecting the prediction result of the model, such as parameter setting errors.

[0077] In this embodiment, the external influencing factors of each prediction error sample refer to the influencing factors of external factors on the prediction result, such as: the input image quality is not high.

[0078] The beneficial effects of the above technical solution are: by performing non-maximum suppression on the probability distribution, obtaining the confidence in each classification, obtaining the second prediction result, and comparing it with the actual label of the coal sample, obtaining the sample probability of prediction error and the influencing factors of the prediction error based on the comparison result, the specific cause of the model error can be quickly determined, and solved through corresponding solutions, thereby improving the overall performance of the model and prediction accuracy.

[0079] Embodiment 7: Based on Example 6, the identification module in this embodiment of the present invention includes: Input unit: takes the coal sample image collected in real time as input and inputs it into the optimized model; Post-processing unit: obtaining the predicted value output by the model and performing post-processing on the predicted value; Classification unit: determining the category of each pixel point according to the post-processing result, and classifying the pixel point based on the classification algorithm; Identification unit: quickly identify the coal sample images collected in real time according to the classification results, and identify the contaminated coal samples according to the identification results.

[0080] In this embodiment, post-processing the predicted value refers to performing a series of processing on the predicted value output by the model according to actual conditions in the actual application stage after the prediction model is completed, including: threshold setting, verification and validation, feature selection and dimensionality reduction.

[0081] In this embodiment, the categories of the pixels may be: Mineral type: Different mineral types in coal samples, such as anthracite, bituminous coal, lignite, etc.

[0082] Coal quality indicators: According to the chemical composition and other characteristics of the coal samples, the coal samples can be classified into different coal quality indicators, such as high volatile coal, low ash coal, high sulfur coal, etc.

[0083] Structural characteristics: characteristics that describe the internal structure of coal samples, such as porosity, degree of crack development, fracture, etc.

[0084] Shape and size: the shape and size of the coal sample, such as flakes, granules, blocks, etc.

[0085] The beneficial effect of the above technical solution is: by inputting the real-time collected coal sample images into the optimized model and preprocessing the predicted values, determining the category of each pixel point and classifying them, the contamination of the coal samples can be discovered in time, ensuring the real-time and accuracy of the identification results.

[0086] Embodiment 8: Based on Example 7, before obtaining image data of contaminated coal samples and non-contaminated coal samples using a high-resolution camera, the embodiment of the present invention also includes: The first determination module: obtains coal sample images of multiple coal sample types, performs feature analysis on the coal sample images to determine particle size feature parameters and appearance feature parameters of each type of coal sample; The second determination module: determines the state loss of each type of coal sample compared with the standard coal sample according to the particle size characteristic parameters and the appearance characteristic parameters; The third determination module: determines the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss of each type of coal sample according to the state loss; The fourth determination module: determining the volume shape critical difference parameters of each type of coal sample based on the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss; The first acquisition module: determines the image identification and differentiation parameters of each type of coal sample according to the critical difference parameter, and obtains the deep layer image information and the shallow layer image information of each type of coal sample based on the image identification and differentiation parameters; The second acquisition module: acquiring the point spread function of each type of coal sample according to the deep layer image information and the shallow layer image information; The fifth determination module: calculates the protrusion amplitude of each protrusion point of each type of coal sample through the point spread function, and determines the phase complex matrix according to the protrusion amplitude; The third acquisition module: performs non-blind deconvolution calculation on the phase complex matrix to obtain high-resolution image samples of various types of coal samples; The sixth determination module: determines the pixel array of each type of coal sample according to the high-resolution image sample, and determines the qualified judgment conditions of each type of coal sample image based on the pixel array; Judgment module: Based on the qualified judgment conditions, the image data of each type of coal sample, contaminated coal sample and non-contaminated coal sample are judged to be qualified, qualified images are screened out according to the judgment results, and re-collection reminders are issued for unqualified images.

[0087] In this embodiment, the particle size characteristic parameters of the coal sample refer to parameters describing the particle size distribution and composition of the coal, such as: Average particle size: refers to the cumulative percentage of coal samples that are smaller than a specified particle size, usually expressed in sieve hole diameter, such as 45μm, 90μm, etc.

[0088] Coarseness and fineness uniformity: reflects the influence of coarse particles in the coal sample on the overall particle size.

[0089] Main separation coefficient: It indicates the ratio of the average particle size of materials larger than the specified particle size to the average particle size of materials smaller than the specified particle size in the coal sample. The larger the Ks value, the more materials with larger particle sizes in the coal sample, and vice versa.

[0090] Particle size frequency curve: A graphical representation of the particle size distribution of a coal sample, including the number or percentage of particles in each particle size range, such as a bimodal curve or a multimodal curve.

[0091] In this embodiment, the appearance characteristic parameters of the coal sample refer to the information about the coal quality obtained by observing and analyzing the appearance of the coal sample, such as: color, fracture, specific gravity, glassy luster, and brittleness.

[0092] In this embodiment, the state loss of the coal sample is caused by the appearance of certain conditions, which leads to a decrease in coal quality or an increase in loss, such as: high temperature and pressure changes, changes in moisture content, and the presence of pollutants.

[0093] In this embodiment, for different types of coal samples, the main coarse-scale loss is mainly due to the influence of external factors on the coal samples during the sampling process, such as vibration, impact, and friction of the coal mining machine, which may cause wear and breakage of the coal sample surface.

[0094] In this embodiment, the coarse-scale auxiliary loss is mainly due to improper operation of the coal sample during preparation, such as insufficient grinding and uneven mixing, which may lead to uneven distribution of particle sizes in the coal sample.

[0095] In this embodiment, the main loss of high-scale is mainly due to the error accumulation of coal samples during high-resolution analysis. High-resolution analysis requires multiple cutting and grinding of samples, which will cause micro deformation and damage to the sample surface.

[0096] In this embodiment, the high-scale auxiliary loss is mainly due to errors caused by failure of the experimental equipment or improper operation, such as insufficient instrument accuracy, inadequate maintenance, environmental pollution and other factors.

[0097] In this embodiment, the critical difference parameter of the volume shape of the coal sample is an indicator representing the difference in density measurement results between the coal sample and the standard sample.

[0098] In this embodiment, in the identification of coal sample images, the distinguishing parameter is a quantitative method used to describe the similarity and difference between two images, such as: uniformity, contrast, structure.

[0099] In this embodiment, the deep image information includes: chemical composition, physical properties, internal structure, and porosity.

[0100] In this embodiment, shallow image information refers to two-dimensional images of the surface of an object obtained by machine vision technology. These images are obtained by scanning the object from different angles or under different light sources, including: surface texture and color, defects and cracks.

[0101] In this embodiment, the point spread function is used to describe the diffusion pattern formed on the receiver after the light passes through a point light source of an object.

[0102] In this embodiment, in coal sample detection, the protrusion amplitude of the protruding point refers to the grayscale difference between a specific target area and a background area during image processing, which can provide quantitative information about the size, shape and distribution of the object.

[0103] In this embodiment, the phase complex matrix is ​​a method for analyzing the optical properties of coal samples. By measuring the absorption and transmission spectrum data of coal samples at different wavelengths, information about the composition and structure of the coal samples can be provided. The phase complex matrix is ​​obtained by calculating the phase difference at each wavelength point. The phase difference reflects the changes in the absorption, transmission and reflection characteristics of a specific wavelength. The elements of the phase complex matrix represent the phase difference values ​​corresponding to each wavelength point.

[0104] In this embodiment, the non-blind deconvolution calculation is an unsupervised learning method that does not require the structure and parameters of the model to be specified in advance. Instead, it directly extracts meaningful information from the input data and automatically estimates the low-frequency signal from the continuous high-frequency signal, thereby achieving the purpose of removing noise.

[0105] In this embodiment, the coal sample pixel array is an array composed of many small pixels, each pixel being able to capture minute changes in coal within a certain range.

[0106] The beneficial effects of the above technical solution are: determining the qualified judgment conditions of each type of coal sample image through the pixel array of the coal sample, and performing qualified judgment on the image data of each type of coal sample, contaminated coal sample and non-contaminated coal sample obtained, and issuing a re-collection reminder based on the judgment result, which can ensure the quality of the coal sample image and avoid the problem of being unable to extract coal sample features due to image quality problems.

[0107] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A contaminated coal sample intelligent identification system based on deep learning, characterized in that: include: Preprocessing module: acquiring image data of polluted coal samples and non-polluted coal samples according to a high-resolution camera, and preprocessing the image data; Model training module: obtaining multiple features of coal samples and labels corresponding to the image data based on the preprocessed image data, and training the deep learning model according to the multiple features of coal samples and labels corresponding to the image data; Adjustment module: evaluate the performance of the trained model based on cross-validation and adjust the model hyperparameters based on the evaluation results; Optimization module: predict the coal samples according to the adjusted model, obtain the sample probability of prediction error and the influencing factors of prediction error, and optimize the model according to the sample probability of prediction error and the influencing factors of prediction error; Identification module: quickly identify the coal sample images collected in real time according to the optimized model, and identify the contaminated coal samples according to the identification results.

2. The deep learning-based intelligent identification system for contaminated coal samples according to claim 1 is characterized in that: Preprocessing modules include: A first determination unit: obtaining ambient environment parameters of the polluted coal sample and the non-polluted coal sample, and determining a corresponding high-resolution camera according to the ambient environment parameters; A first acquisition unit: acquiring a preset number of image data of contaminated coal samples and non-contaminated coal samples according to the high-resolution camera; Preprocessing unit: preprocess the image data according to multiple filter functions of the PIL library and the OpenCV library.

3. The contaminated coal sample intelligent identification system based on deep learning according to claim 1 is characterized in that: Model training module, including: Extraction unit: extract multiple features of coal samples based on convolutional neural network according to preprocessed image data; A second acquisition unit: analyzing the multiple features, and acquiring a label corresponding to the coal sample in each image data according to the analysis result; Model training unit: converts the extracted multiple features and corresponding labels into a unified vector representation, and trains the deep learning model according to the vector representation.

4. The deep learning-based intelligent identification system for contaminated coal samples according to claim 3 is characterized in that: Extraction unit, comprising: Get subunits: Use a deep learning framework to build a convolutional neural network model and get multiple components of the convolutional neural network model; Training subunit: Use labeled image datasets to train the convolutional neural network model to obtain the optimal model configuration; Input subunit: input the preprocessed image data into the trained convolutional neural network model to obtain a feature map containing multi-level and multi-scale feature information of the image; Extraction subunit: Multiple feature maps are fused based on the attention mechanism, and multiple features of the coal sample are extracted according to the fusion results.

5. The contaminated coal sample intelligent identification system based on deep learning according to claim 1 is characterized in that: Adjustment modules, including: Evaluation unit: evaluates the performance of the trained model based on K-fold cross validation; Sorting unit: Determine the performance index of the model according to the evaluation results, sort the models according to the performance index, and obtain the best model; Testing unit: After determining the best model, use the model to test it on a dataset that has not been cross-validated; Adjustment unit: adjust the model hyperparameters according to the test results, determine the best parameter combination through the grid search algorithm, and determine the adjusted model based on the best parameter combination.

6. The contaminated coal sample intelligent identification system based on deep learning according to claim 1 is characterized in that: Optimization modules, including: The third acquisition unit: acquires the actual label of the coal sample, predicts the coal sample according to the adjusted model, acquires the first prediction result, classifies the image according to the first prediction result, and acquires the probability distribution of each pixel in each image in each classification; The fourth acquisition unit: performs non-maximum suppression on the obtained probability distribution, obtains the confidence in each category, and uses the category with the highest confidence as the second prediction result; Comparison unit: comparing the second prediction result with the actual label of the coal sample to obtain a comparison result; The third acquisition unit: if the comparison results are inconsistent, the samples with incorrect predictions are determined according to the confusion matrix calculation, and the number of samples with incorrect predictions and the internal and external influencing factors of each sample with incorrect predictions are obtained; Optimization unit: obtains the probability of prediction errors according to the number of prediction errors, and optimizes the model by combining the internal and external influencing factors of the prediction errors.

7. The deep learning-based intelligent identification system for contaminated coal samples according to claim 1 is characterized in that: Identification module, including: Input unit: takes the coal sample image collected in real time as input and inputs it into the optimized model; Post-processing unit: obtaining the predicted value output by the model and performing post-processing on the predicted value; Classification unit: determining the category of each pixel point according to the post-processing result, and classifying the pixel point based on the classification algorithm; Identification unit: quickly identify the coal sample images collected in real time according to the classification results, and identify the contaminated coal samples according to the identification results.

8. The contaminated coal sample intelligent identification system based on deep learning according to claim 1 is characterized in that: Before obtaining image data of contaminated coal samples and non-contaminated coal samples using a high-resolution camera, it also includes: The first determination module: obtains coal sample images of multiple coal sample types, performs feature analysis on the coal sample images to determine particle size feature parameters and appearance feature parameters of each type of coal sample; The second determination module: determines the state loss of each type of coal sample compared with the standard coal sample according to the particle size characteristic parameters and the appearance characteristic parameters; The third determination module: determines the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss of each type of coal sample according to the state loss; The fourth determination module: determining the volume shape critical difference parameters of each type of coal sample based on the coarse-scale main loss and auxiliary loss and the high-scale main loss and auxiliary loss; The first acquisition module: determines the image identification and differentiation parameters of each type of coal sample according to the critical difference parameter, and obtains the deep layer image information and the shallow layer image information of each type of coal sample based on the image identification and differentiation parameters; The second acquisition module: acquiring the point spread function of each type of coal sample according to the deep layer image information and the shallow layer image information; The fifth determination module: calculates the protrusion amplitude of each protrusion point of each type of coal sample through the point spread function, and determines the phase complex matrix according to the protrusion amplitude; The third acquisition module: performs non-blind deconvolution calculation on the phase complex matrix to obtain high-resolution image samples of various types of coal samples; The sixth determination module: determines the pixel array of each type of coal sample according to the high-resolution image sample, and determines the qualified judgment conditions of each type of coal sample image based on the pixel array; Judgment module: Based on the qualified judgment conditions, the image data of each type of coal sample, contaminated coal sample and non-contaminated coal sample are judged to be qualified, qualified images are screened out according to the judgment results, and re-collection reminders are issued for unqualified images.

Images