Dust detection method and system based on multi-task learning model

Abstract

The present application relates to a dust detection method and system based on a multi-task learning model, the system comprising a perception measurement module, a model training module, a model inference module, a visualization and over-limit warning module that cooperate with each other, the perception measurement module being used to collect image data and dust concentration data in a mine environment; the model training module being used to build and train a multi-task learning model, containing data loading, model definition, loss function setting and model optimization functions; the model inference module being used to run the trained model in a deployment environment, to perform dust detection and concentration prediction on the real-time collected images, and to optimize the inference speed of the model; the visualization and over-limit warning module being used to visually display the detection results and concentration values, and to trigger an alarm mechanism when the dust concentration exceeds a preset threshold, to remind relevant personnel, to enable real-time and accurate dust concentration monitoring, and to provide important technical support and decision basis for mine dust prevention and control.

Description

Technical Field

[0001] This invention relates to the field of mine dust detection, specifically to a dust detection method based on a multi-task learning model and its application. Background Technology

[0002] Dust concentration in the mine environment is directly related to mine safety and the health of workers. High concentrations of dust can not only cause occupational diseases such as pneumoconiosis in workers, but also potentially lead to major safety accidents such as explosions. Therefore, effective monitoring and control of mine dust is a crucial link in ensuring safe mine production.

[0003] Traditional dust detection methods mainly include filter membrane sampling, light scattering dust meters, and beta-ray methods. With the rapid development of computer vision and deep learning technologies, image-based dust detection methods have gradually gained attention. The above methods can meet engineering needs to a certain extent, but still have many limitations: (1) Traditional detection methods usually require regular manual sampling and offline analysis, which cannot achieve continuous detection and real-time data updates. (2) Many detection methods can only detect dust concentration at fixed points or in local areas, and cannot achieve comprehensive coverage of the entire mine space. This limitation may lead to high concentrations of dust in some key areas not being detected and dealt with in a timely manner. (3) Most image detection methods can only identify the presence or absence of dust, but cannot accurately quantify the dust concentration in the detection area, making it difficult to meet the needs of refined dust concentration detection.

[0004] Therefore, a new dust detection method is urgently needed to solve the above problems. Summary of the Invention

[0005] This invention proposes a dust detection method and system based on a multi-task learning model.

[0006] The dust detection method based on a multi-task learning model includes the following steps:

[0007] (1) Install and debug data acquisition equipment, specifically including: arranging camera equipment and dust concentration sensors in the mine roadway; debugging and calibrating the camera equipment and sensors;

[0008] (2) Data collection and dataset construction, specifically including:

[0009] (2.1) Using the installed camera equipment, continuously collect image data in the mine roadway, covering the dust distribution in different time periods, different areas and different working conditions; synchronously record the data of the dust concentration sensor in the area corresponding to the collected image, and obtain the dust concentration value at the corresponding time and location;

[0010] (2.2) Use image annotation tools to manually annotate the images acquired in step (2.1), perform manual semantic segmentation on the dust areas in the images, and generate pixel-level annotation masks; divide the tunnel images into two categories: background and dust, and label each pixel as "Background" or "Dust" to form semantic segmentation labels; for each annotated image, associate the dust concentration value of the image at the corresponding acquisition time with its dust area label, where Background represents the background and Dust represents the dust.

[0011] (2.3) Constructing the dataset includes: performing random augmentation on the images in step (2.2) to increase data diversity and improve the robustness of the model; organizing and converting the labeled image data into a format that the model can read; constructing a multi-task learning dataset containing images, semantic segmentation labels and dust concentration values; and preprocessing the dataset.

[0012] (3) Construct and train a multi-task learning model for dust detection, specifically including:

[0013] (3.1) Constructing a multi-task learning model for dust detection involves constructing a feature extraction network, namely a Backbone, a multi-scale feature fusion module, a segmentation head, and a regression head. The Backbone is responsible for extracting general features of the image; the multi-scale feature fusion module is used to fuse and enhance features of different scales in the image; the segmentation head is used to perform pixel-level classification and locate dust regions; and the regression head is used to predict the concentration value of the dust segmentation region. In the multi-task learning model, the features after multi-scale feature fusion can be shared by multiple subsequent task branches.

[0014] (3.2) The model constructed in step (3.1) is trained and optimized, specifically including: the preprocessed image dataset in step (2) is randomly divided into a 70% training set, a 20% validation set, and a 10% test set; the training set is used for training the model, and the dust detection multi-task learning model constructed in step (3.1) is trained; during the training process, the model optimizes the segmentation task and the regression task at the same time, and obtains the optimal performance of the model in these two tasks by combining the segmentation loss and the regression loss; the validation set is used to evaluate the performance of the model and perform hyperparameter tuning during the training process; the test set is used to evaluate the final performance of the model after training to ensure the generalization ability of the model;

[0015] (4) Dust concentration detection using an optimized model, specifically including: using the camera equipment installed in step (1), acquiring the original image data of the mine roadway through step (2), and constructing an labeled image dataset. The constructed image dataset is then imported into the optimized dust detection multi-task learning model established in step (3) to obtain multi-scale feature maps. The segmentation head upsamples and outputs a segmentation probability map for the learned feature maps at different scales, sets a threshold of 0.5, and generates a binary mask based on the set threshold. The prediction head extracts dust area features based on the generated binary mask, pools the features of the dust area, and generates a feature vector. The feature vector is then input into the regression head module for dust concentration prediction, thereby realizing real-time dust identification and concentration detection in the image.

[0016] (5) Visualize and over-limit warning of the dust detection results detected in step (4).

[0017] Preferably, the image annotation tool in step (2.2) is LabelImg.

[0018] Preferably, the random enhancement operations in step (2.3) include image rotation, translation, scaling, and color jitter.

[0019] Preferably, the preprocessing operations for the dataset in step (2.3) include resizing the images and normalizing their colors.

[0020] Preferably, the construction of the feature extraction network in step (3.1) specifically includes:

[0021] Feature extraction is performed using ResNet50, and the extracted features consist of: a two-layer input branch consisting of one 7×7 convolutional layer and one 3×3 max-pooling layer; and 48 layers of residual blocks in four stages.

[0022] Stage 1: Contains 3 residual blocks;

[0023] Stage 2: Contains 4 residual blocks, and the output feature map is denoted as C3;

[0024] Stage 3: Contains 6 residual blocks, and the output feature map is denoted as C4;

[0025] Stage 4: Contains 3 residual blocks, and the output feature map is denoted as C5.

[0026] Each residual block consists of 3 convolutional layers; C3 has 512 channels; C4 has 1024 channels; and C5 has 2048 channels.

[0027] Preferably, the construction of the multi-scale feature fusion module in step (3.1) includes:

[0028] A BiFPN architecture is adopted for multi-scale feature fusion. BiFPN enhances multi-scale features through bidirectional feature fusion, employing both top-down and bottom-up approaches. BiFPN includes:

[0029] (a1) Generate the initial feature pyramid layer

[0030] Generate the initial feature pyramid layers, which are arranged from bottom to top as feature layers P3, P4, P5, P6, and P7.

[0031] P3, P4, and P5 are obtained from feature layers C3, C4, and C5 through 1×1 convolution. The specific calculation formula is: P i =Conv 1x1 (C i ), i = 3, 4, 5, Conv 1x1 It is a 1×1 convolutional layer;

[0032] Add convolution and downsampling operations to P5 to generate feature layers P6 and P7. The specific calculation formula is as follows:

[0033] P6 = Conv 3x3,s＝2 (P5)

[0034] P7 = ReLU(Conv) 3x3,s＝2 (P6)

[0035] Among them, Conv 3x3，s＝2 The kernel is 3x3 with a stride of s = 2; ReLU is the activation function, ReLU = max(0,x);

[0036] (a2) Top-down path fusion

[0037] Feature fusion is performed from high-level features to low-level features. In the top-down path, the calculation formula for feature fusion is:

[0038]

[0039]

[0040] in, P represents the feature map of the i-th layer in the top-down path; i The feature layer is the initial input; This is the top-down feature map of the previous layer, i.e., the (i+1)th layer; The weights are learnable non-negative weights, which are guaranteed to be non-negative by the ReLU activation function; ∈ is a small constant to prevent division by zero, usually taken as ∈ = 10. -4 Upsample is an upsampling operation that enlarges the feature map size by a factor of 2; Conv 3x3 It is a 3×3 convolutional layer used for smoothing the fused features;

[0041] (a3) Bottom-up path fusion

[0042] Feature fusion is performed from low-level features to high-level features. In the bottom-up path, the calculation formula for feature fusion is:

[0043]

[0044] in, This is the feature map of the i-th layer in the bottom-up path; This represents the feature map of the i-th layer in the top-to-bottom path; This is the bottom-up feature map of the next layer, i.e., the (i-1)th layer; The weights are learnable non-negative weights, which are guaranteed to be non-negative by the ReLU activation function; ∈ is a small constant to prevent division by zero, usually taken as ∈ = 10. -4 Downsample is a downsampling operation that reduces the feature map size by a factor of 2; Conv 3x3 It is a 3×3 convolutional layer used for smoothing the fused features;

[0045] (a4) Learnable fusion weights

[0046] In steps (b) and (c), at each fusion node, learnable weights are introduced to adjust the contributions of different feature layers; initial weights are set. During training, the weight parameters are automatically learned through backpropagation, enabling the BiFPN multi-scale feature fusion module to output feature maps at different scales.

[0047] Preferably, step (3.1) of constructing the segmentation head includes: receiving feature maps from the multi-scale feature fusion module, upsampling the low-resolution feature map, and fusing feature maps of different scales to finally generate a feature map with the same size as the input image, classifying each pixel of the feature map to determine whether it belongs to "Dust" or "Background", specifically including the following steps:

[0048] (b1) Perform feature layer adjustments, specifically including:

[0049] For the multi-scale feature layer of BiFPN from step (d1): Channel alignment is performed, and a 1×1 convolution is applied to each feature layer to obtain... to

[0050] in, The adjusted feature layer;

[0051] (b2) Perform multi-scale feature fusion and upsampling, specifically including:

[0052] A bottom-up upsampling strategy is adopted to progressively upsample the low-resolution feature map and then fuse it with the high-resolution feature map. Initially, the samples are progressively upsampled and fused with higher resolution feature layers, resulting in fused features F7, F6, F5, F4, and F3.

[0053] For i=7, perform initial sampling.

[0054] For i = 6, 5, 4, 3, perform the following steps:

[0055] (b21) Fuse the previous layer's features F i+1 Upsampled to the current adjusted feature layer The size, upsampling operation uses bilinear interpolation or deconvolution: U i =Upsample(F i+1 ), U i These are the features after upsampling;

[0056] (b22) Upsample the feature U i Compared with the current adjustment The fusion process is performed using a splicing method:

[0057]

[0058] Where conv3x3 is a 3×3 convolutional layer with C output channels;

[0059] The highest resolution feature layer F3 was obtained, with a size of

[0060] Where H is the height of the input image; W is the width of the input image;

[0061] (b22) Gradually upsample F3 to the size H×W of the input image, extracting features through a convolutional layer after each upsampling; upsample to... F up1 =Upsample(F3), then perform convolution processing: F up1 =Conv3x3(F up1 Upsampling to F up2 =Upsample(F up1 Then, convolution is performed: F up2 =Conv3x3(F up2 Upsampled to H×W:F up3 =Upsample(F up2 Then, convolution is performed: Fup3 =Conv3x3(F up3 ), F up3 It is an image feature map with the same size as the original image; F up1 The upsampling operation parameters for the highest resolution feature layer F3, upsampled one layer; F up2 The upsampling operation parameters for the highest resolution feature layer F3, which upsamples two feature layers; F up3 The upsampling operation parameters for upsampling 3 layers of the highest resolution feature layer F3;

[0062] (b3) Split the output layer, specifically including:

[0063] Pixel-level classification is performed on the upsampled features from step (b22), and the segmentation results and fused feature maps are output, specifically including:

[0064] (b31) Using the sampled features F at the 1×1 convolutional layer up3 Perform a transformation to adjust the number of channels in the feature map, and output the adapted feature map O:

[0065] O = Conv 1x1 (F up3 )

[0066] (b32) Using the Sigmoid activation function, the output is mapped to the (0,1) range, representing the probability that a pixel belongs to "dust":

[0067] S=σ(O)

[0068] Where σ is the Sigmoid activation function.

[0069] (b33) Generate a binary segmentation mask based on the probability from step (b32):

[0070]

[0071] Where τ is the threshold, set to 0.5; S(x,y) represents the probability value that any pixel (x,y) on feature map O belongs to the "dust" class. A binary segmentation mask is generated based on the threshold to obtain the final segmentation result.

[0072] Preferably, step (3.1) of constructing the regression head specifically includes:

[0073] (c1) Extract features of the dust region:

[0074] The process involves calling up the fused and upsampled high-resolution feature maps from the segmentation head, focusing on dust-related regions in the image, and filtering out background and other irrelevant information. Specifically, this includes the following steps:

[0075] (c1) Obtain the feature map O from the segmentation head, denoted as F. seg The dimensions are H×W×C;

[0076] (c12) Apply the segmentation mask M(x,y) obtained in step (b32) to the feature map F seg A masking operation is performed on each pixel location, retaining only the features of the dust area and setting the features of the background area to zero. The specific calculation formula is as follows:

[0077] F masked (x,y)=F seg (x,y)×M(x,y)

[0078] Among them, F seg (x,y) represents the feature map O, F from the segmentation head after one-dimensional convolution. masked (x,y) is the feature map after the masking operation, retaining only the features of the "dust" region and setting the background region to zero;

[0079] (c2) Perform feature aggregation and global representation:

[0080] The feature map of the dust region is converted into a fixed-length feature vector to provide input for subsequent regression calculations. The feature vector f of the dust region is then calculated.

[0081]

[0082] Where, N dust The total number of pixels in the dusty area is calculated as follows:

[0083]

[0084] Where f has dimension C, which is the number of channels in the feature map, if N dust =0 indicates that no dust area was detected;

[0085] (c3) Calculation of regression layer and dust concentration output:

[0086] A fully connected layer is used to perform regression calculation on the feature vector, outputting the dust concentration value D:

[0087] D = W²h + b², where W² is the weight matrix; b² is the bias scalar; and h is calculated as follows:

[0088] h = ReLU(W1f + b1)

[0089] Where h is the hidden layer output; W1 is the weight matrix; and b1 is the bias scalar.

[0090] Preferably, step (3.2) specifically includes:

[0091] (d1) Define the total loss function L:

[0092] L=λ reg L reg +λ seg L reg

[0093] Among them, L reg The regression loss is used to measure the error in dust concentration prediction; L seg λ is the segmentation loss, used to measure the loss in dust image segmentation; reg and λ seg The loss weight coefficients control the contributions of the segmentation and regression tasks to the total loss;

[0094] (d2) Model training is performed, specifically including:

[0095] (d21) Perform forward propagation

[0096] The images and dust concentration data of the training set divided in step (2) are input into the model constructed in step (3.1) to perform joint inference of the segmentation head and the regression head, and the dust segmentation result and concentration prediction value are obtained.

[0097] (d22) Perform loss calculation

[0098] Based on the segmentation results predicted in step (d21) and the actual segmentation labels, the segmentation loss L is calculated. seg Based on the predicted dust concentration and the actual concentration, the regression loss L is calculated. reg Combining the two loss functions, we obtain the total loss L;

[0099] (d23) Perform backpropagation

[0100] Backpropagation is performed based on the total loss L obtained in step (d22) to calculate the gradient of each layer of the model;

[0101] (d23) Perform parameter updates

[0102] The optimizer updates the model parameters based on the gradient to minimize the total loss. During training, the learning rate is adaptively adjusted based on the model's performance on the validation set to avoid the model getting stuck in local optima.

[0103] (d24) Perform hyperparameter tuning, specifically including:

[0104] Based on the model convergence, a learning rate scheduler is used to gradually adjust the learning rate, and the batch size is adjusted according to hardware resources to balance training speed and model stability. λ is dynamically adjusted based on experimental feedback. reg and λ segThe weights are determined to ensure that the losses from the segmentation and regression tasks are in a reasonable proportion of the total loss.

[0105] (d3) Perform model validation, specifically including:

[0106] After model training is completed, the model performance is evaluated using a validation set, and the segmentation accuracy and concentration prediction error are calculated to ensure that the model can adapt to dust detection tasks in actual mining environments. The specific steps include the following:

[0107] (d31) Validation set preparation and preprocessing, specifically including:

[0108] The image data of the validation set divided in step (2) are preprocessed to ensure that the data input format is consistent with the training set;

[0109] (d32) Calculation performance metrics, specifically including:

[0110] (d321) Calculate the segmentation performance evaluation metrics, specifically including:

[0111] Accuracy refers to the proportion of correct predictions among all predictions, and is used to measure the overall accuracy in dust separation tasks.

[0112] For dust separation tasks, the formula for accuracy is:

[0113]

[0114] Where: TP is the number of pixels whose true category is "dust" and which are correctly predicted as "dust"; TN is the number of pixels whose true category is "background" and which are correctly predicted as "background"; FP is the number of pixels whose true category is "background" but are incorrectly predicted as "dust"; FN is the number of pixels whose true category is "dust" but are incorrectly predicted as "background".

[0115] Intersection over Union (IoU): This measures the degree of overlap between the predicted result and the true label. For segmentation tasks, it calculates the degree of overlap between the predicted dust region and the real dust region. The higher the IoU, the higher the model segmentation accuracy.

[0116]

[0117] Where A represents the predicted segmentation region; B represents the ground truth segmentation region; |A∩B| represents the number of pixels at the intersection of the predicted region and the ground truth region; |AUB| represents the number of pixels at the union of the predicted region and the ground truth region.

[0118] The F1 score is calculated by taking into account both precision and recall in the segmentation process to evaluate the model's segmentation performance.

[0119]

[0120]

[0121]

[0122] (d322) Calculate the concentration prediction error assessment index, specifically including:

[0123] Mean squared error (MSE): used to measure the difference between predicted and actual dust concentration values;

[0124]

[0125] Where: y i It is the i-th true value; is the i-th predicted value; n is the total number of samples;

[0126] Mean Absolute Error (MAE): This measures the average absolute error between the predicted and the actual values.

[0127]

[0128] (d33) Save the optimal model, specifically including:

[0129] Based on the performance metrics calculated in step (d32), check the model's segmentation and regression task fluctuations on the validation set to ensure that the model's prediction results are stable on different samples, with no obvious overfitting or underfitting. The model parameters that achieve the best performance on the validation set will be saved for use in dust detection and concentration prediction in practical applications.

[0130] (d4) Perform model testing. Use the test set in step (2) to test the validated model and test its performance.

[0131] This invention also provides a system based on the above method, characterized by comprising a sensing and measurement module, a model training module, a model inference module, and a visualization and over-limit early warning module that cooperate with each other. The sensing and measurement module is used to collect image data and dust concentration data in the mine environment; the model training module is used to construct and train a multi-task learning model, including data loading, model definition, loss function setting, and model optimization functions; the model inference module is used to run the trained model in the deployment environment, perform dust detection and concentration prediction on the real-time collected images, and optimize the model's inference speed; the visualization and over-limit early warning module is used to visualize the detection results and concentration values, and trigger an alarm mechanism to remind relevant personnel when the dust concentration exceeds a preset threshold. The sensing and measurement module is responsible for collecting environmental data and transmitting the environmental data to the model training module. The model training module receives the environmental data, trains the multi-task learning model, and outputs the trained multi-task learning model to the model inference module. The model inference module uses the multi-task learning model to infer the collected data and outputs the dust concentration prediction result to the visualization and over-limit early warning module. The visualization and over-limit early warning module visualizes the prediction result and issues an early warning broadcast based on the set dust concentration threshold.

[0132] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0133] (1) The present invention can perform sampling and online analysis on the area to be detected, and can achieve continuous detection and real-time data updates.

[0134] (2) The dust concentration detection method of the present invention can achieve full coverage of the entire mine space, thereby avoiding the failure to detect and deal with high concentrations of dust in certain key areas in a timely manner.

[0135] (3) This invention integrates semantic segmentation and regression prediction tasks into a unified deep learning framework using a multi-task learning model. By utilizing shared feature representations, it can achieve precise segmentation of dust dispersion areas and accurate detection of concentration, which can make up for the shortcomings of traditional detection methods and is of great significance for real-time, comprehensive and accurate monitoring of mine dust.

[0136] (4) This invention does not require pre-set monitoring points and can directly capture the entire view inside the alley through a smart camera, providing higher detection accuracy and timeliness, while achieving a higher degree of detection automation.

[0137] (5) The present invention uses image recognition technology to detect dust environment, which is more suitable for complex underground roadway environment. Attached Figure Description

[0138] Figure 1 This is a schematic diagram of the dust detection method according to Embodiment 1 of the present invention;

[0139] Figure 2 This is a schematic diagram of the multi-task learning model architecture for dust detection in Embodiment 1 of the present invention;

[0140] Figure 3 This is a schematic diagram of the dust detection system according to Embodiment 2 of the present invention. Specific implementation methods

[0141] To make the objectives, technical solutions, and optimizations of this embodiment clearer and more explicit, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0142] Example 1:

[0143] like Figure 1-3 As shown, this embodiment provides a dust detection method and system based on a multi-task learning model.

[0144] The dust detection method based on a multi-task learning model includes the following steps:

[0145] (1) Install and debug data acquisition equipment, specifically including: arranging camera equipment and dust concentration sensors in the mine roadway; debugging and calibrating the camera equipment and sensors;

[0146] (2) Data collection and dataset construction, specifically including:

[0147] (2.1) Using the installed camera equipment, continuously collect image data in the mine roadway, covering the dust distribution in different time periods, different areas and different working conditions; synchronously record the data of the dust concentration sensor in the area corresponding to the collected image, and obtain the dust concentration value at the corresponding time and location;

[0148] (2.2) Use an image annotation tool to manually annotate the images acquired in step (2.1). The image annotation tool is LabelImg. Perform manual semantic segmentation on the dust area in the image to generate pixel-level annotation masks. Divide the tunnel image into two categories: background and dust. Each pixel is marked as "Background" or "Dust" to form a semantic segmentation label. For each annotated image, associate the dust concentration value of the image at the corresponding acquisition time with its dust area label. Background represents the background and Dust represents the dust.

[0149] (2.3) Constructing the dataset specifically includes: performing random augmentation operations on the images in step (2.2) to increase data diversity and improve the robustness of the model. Random augmentation operations include image rotation, translation, scaling, and color jittering. The labeled image data is organized and converted into a format that the model can read. A multi-task learning dataset containing images, semantic segmentation labels, and dust concentration values ​​is constructed. The dataset is preprocessed. Preprocessing operations include image resizing and color normalization.

[0150] (3) Construct and train a multi-task learning model for dust detection, specifically including:

[0151] (3.1) Constructing a multi-task learning model for dust detection involves building a feature extraction network, namely a backbone, a multi-scale feature fusion module, a segmentation head, and a regression head. The backbone is responsible for extracting general features of the image; the multi-scale feature fusion module is used to fuse and enhance features at different scales in the image; the segmentation head is used for pixel-level classification and localization of dust regions; and the regression head is used to predict the concentration value of the segmented dust regions. In the multi-task learning model, the features after multi-scale feature fusion can be shared by multiple subsequent task branches. The construction of the feature extraction network specifically includes:

[0152] Feature extraction is performed using ResNet50, and the extracted features consist of: a two-layer input branch consisting of one 7×7 convolutional layer and one 3×3 max-pooling layer; and 48 layers of residual blocks in four stages.

[0153] Stage 1: Contains 3 residual blocks;

[0154] Stage 2: Contains 4 residual blocks, and the output feature map is denoted as C3;

[0155] Stage 3: Contains 6 residual blocks, and the output feature map is denoted as C4;

[0156] Stage 4: Contains 3 residual blocks, and the output feature map is denoted as C5.

[0157] Each residual block consists of 3 convolutional layers; C3 has 512 channels; C4 has 1024 channels; and C5 has 2048 channels.

[0158] Construct a multi-scale feature fusion module, including:

[0159] A BiFPN architecture is adopted for multi-scale feature fusion. BiFPN enhances multi-scale features through bidirectional feature fusion, employing both top-down and bottom-up approaches. BiFPN includes:

[0160] (a1) Generate the initial feature pyramid layer

[0161] Generate the initial feature pyramid layers, which are arranged from bottom to top as feature layers P3, P4, P5, P6, and P7.

[0162] P3, P4, and P5 are obtained from feature layers C3, C4, and C5 through 1×1 convolution. The specific calculation formula is: P i =Conv 1x1 (C i ), i = 3, 4, 5, Conv 1x1 It is a 1×1 convolutional layer;

[0163] Add convolution and downsampling operations to P5 to generate feature layers P6 and P7. The specific calculation formula is as follows:

[0164] P6 = Conv 3x3,s＝2 (P5)

[0165] P7 = ReLU(Conv) 3x3,s＝2 (P6)

[0166] Among them, Conv 3x3，s＝2 The kernel is 3x3 with a stride of s = 2; ReLU is the activation function, ReLU = max(0,x).

[0167] (a2) Top-down path fusion

[0168] Feature fusion is performed from high-level features to low-level features. In the top-down path, the calculation formula for feature fusion is:

[0169]

[0170]

[0171] in, P represents the feature map of the i-th layer in the top-down path; i The feature layer is the initial input; This is the top-down feature map of the previous layer, i.e., the (i+1)th layer; The weights are learnable non-negative weights, which are guaranteed to be non-negative by the ReLU activation function; ∈ is a small constant to prevent division by zero, usually taken as ∈ = 10. -4 Upsample is an upsampling operation that enlarges the feature map size by a factor of 2; Conv 3x3 It consists of 3×3 convolutional layers for smoothing the fused features; the BiFPN architecture is a neural network architecture well-known to those skilled in the art for object detection tasks.

[0172] (a3) Bottom-up path fusion

[0173] Feature fusion is performed from low-level features to high-level features. In the bottom-up path, the calculation formula for feature fusion is:

[0174]

[0175] in, This is the feature map of the i-th layer in the bottom-up path; This represents the feature map of the i-th layer in the top-to-bottom path; This is the bottom-up feature map of the next layer, i.e., the (i-1)th layer; The weights are learnable non-negative weights, which are guaranteed to be non-negative by the ReLU activation function; ∈ is a small constant to prevent division by zero, usually taken as ∈ = 10. -4 Downsample is a downsampling operation that reduces the feature map size by a factor of 2; Conv 3x3 It is a 3×3 convolutional layer used for smoothing the fused features;

[0176] (a4) Learnable fusion weights

[0177] In steps (b) and (c), at each fusion node, learnable weights are introduced to adjust the contributions of different feature layers; initial weights are set. During training, the weight parameters are automatically learned through backpropagation, enabling the BiFPN multi-scale feature fusion module to output feature maps at different scales.

[0178] Constructing the segmentation head includes: receiving feature maps from the multi-scale feature fusion module; upsampling the low-resolution feature map and fusing it between feature maps of different scales to finally generate a feature map of the same size as the input image; classifying each pixel in the feature map to determine whether it belongs to "Dust" or "Background"; specifically including the following steps:

[0179] (b1) Perform feature layer adjustments, specifically including:

[0180] For the multi-scale feature layer of BiFPN from step (d1): Channel alignment is performed, and a 1×1 convolution is applied to each feature layer to obtain... to

[0181] in, The adjusted feature layer;

[0182] (b2) Perform multi-scale feature fusion and upsampling, specifically including:

[0183] A bottom-up upsampling strategy is adopted to progressively upsample the low-resolution feature map and then fuse it with the high-resolution feature map. Initially, the samples are progressively upsampled and fused with higher resolution feature layers, resulting in fused features F7, F6, F5, F4, and F3.

[0184] For i=7, perform initial sampling.

[0185] For i = 6, 5, 4, 3, perform the following steps:

[0186] (b21) Fuse the previous layer's features F i+1 Upsampled to the current adjusted feature layer The size, upsampling operation uses bilinear interpolation or deconvolution: U i =Upsample(F i+1 ), U i These are the features after upsampling;

[0187] (b22) Upsample the feature U i Compared with the current adjustment The fusion process is performed using a splicing method:

[0188]

[0189] Conv3x3 is a 3×3 convolutional layer with C output channels;

[0190] The highest resolution feature layer F3 was obtained, with a size of

[0191] Where H is the height of the input image and W is the width of the input image.

[0192] (b22) Gradually upsample F3 to the size H×W of the input image, extracting features through a convolutional layer after each upsampling; upsample to... F up1 =Upsample(F3), then perform convolution processing: F up1 =Conv3x3(F up1 Upsampling to F up2 =Upsample(F up1 Then, convolution is performed: F up2 =Conv3x3(F up2 Upsampled to H×W:F up3 =Upsample(F up2 Then, convolution is performed: F up3 =Conv3x3(F up3 ), F up3 It is an image feature map with the same size as the original image; Fup1 The upsampling operation parameters for the highest resolution feature layer F3, upsampled one layer; F up2 The upsampling operation parameters for the highest resolution feature layer F3, which upsamples two feature layers; F up3 The upsampling operation parameters for upsampling 3 layers of the highest resolution feature layer F3;

[0193] (b3) Split the output layer, specifically including:

[0194] Pixel-level classification is performed on the upsampled features from step (b22), and the segmentation results and fused feature maps are output, specifically including:

[0195] (b31) Using the sampled features F at the 1×1 convolutional layer up3 Perform a transformation to adjust the number of channels in the feature map, and output the adapted feature map O:

[0196] O = Conv 1x1 (F up3 )

[0197] (b32) Using the Sigmoid activation function, the output is mapped to the (0,1) range, representing the probability that a pixel belongs to "dust":

[0198] S=σ(O)

[0199] Where σ is the Sigmoid activation function.

[0200] (b33) Generate a binary segmentation mask based on the probability from step (b32):

[0201]

[0202] Among them, F seg (x,y) represents the feature map O, F from the segmentation head after one-dimensional convolution. masked (x,y) is the feature map after the masking operation, retaining only the features of the "dust" region and setting the background region to zero; τ is the threshold, set to 0.5; S(x,y) represents the probability value that any pixel (x,y) on the feature map O belongs to the "dust" class. A binary segmentation mask is generated based on the threshold to obtain the final segmentation result.

[0203] Constructing the regression head specifically includes:

[0204] (c1) Extract features of the dust region:

[0205] The process involves calling up the fused and upsampled high-resolution feature maps from the segmentation head, focusing on dust-related regions in the image, and filtering out background and other irrelevant information. Specifically, this includes the following steps:

[0206] (c11) Obtain the feature map O from the segmentation head, denoted as F. seg The dimensions are H×W×C;

[0207] (c12) Apply the segmentation mask M(x,y) obtained in step (b32) to the feature map F seg A masking operation is performed on each pixel location, retaining only the features of the dust area and setting the features of the background area to zero. The specific calculation formula is as follows:

[0208] F masked (x,y)=F seg (x,y)×M(x,y)

[0209] Among them, F seg (x,y) represents the feature map O, F from the segmentation head after one-dimensional convolution. masked (x,y) is the feature map after the masking operation, retaining only the features of the "dust" region and setting the background region to zero;

[0210] (c2) Perform feature aggregation and global representation:

[0211] The feature map of the dust region is converted into a fixed-length feature vector to provide input for subsequent regression calculations. The feature vector f of the dust region is then calculated.

[0212]

[0213] Where, N dust The total number of pixels in the dusty area is calculated as follows:

[0214]

[0215] Where f has dimension C, which is the number of channels in the feature map, if N dus t = 0 indicates that no dust was detected in the area;

[0216] (c3) Calculation of regression layer and dust concentration output:

[0217] A fully connected layer is used to perform regression calculation on the feature vector, outputting the dust concentration value D:

[0218] D = W²h + b²

[0219] y = W2h + b2

[0220] Where W2 is the weight matrix; b2 is the bias scalar; and h is calculated as follows:

[0221] h = ReLU(W1f + b1)

[0222] Where h is the hidden layer output; W1 is the weight matrix; and b1 is the bias scalar.

[0223] (3.2) The model constructed in step (3.1) is trained and optimized, specifically including: the preprocessed image dataset in step (2) is randomly divided into a 70% training set, a 20% validation set, and a 10% test set. The training set is used for training the model, and the dust detection multi-task learning model constructed in step (3.1) is trained. During the training process, the model optimizes both the segmentation and regression tasks. By combining the segmentation loss and the regression loss, the optimal performance of the model in these two tasks is obtained. The validation set is used to evaluate the performance of the model and perform hyperparameter tuning during the training process. The test set is used to evaluate the final performance of the model after training to ensure the generalization ability of the model.

[0224] Specifically, it includes:

[0225] (d1) Define the total loss function L:

[0226] L=λ reg L reg +λ reg L reg

[0227] Among them, L reg The regression loss is used to measure the error in dust concentration prediction; L seg λ is the segmentation loss, used to measure the loss in dust image segmentation; reg and λ reg The loss weight coefficients control the contributions of the segmentation and regression tasks to the total loss;

[0228] (d2) Model training is performed, specifically including:

[0229] (d21) Perform forward propagation

[0230] The images of the training set divided in step (2) are input into the model constructed in step (3.1) to perform joint inference of the segmentation head and the regression head, and the dust segmentation results and concentration prediction values ​​are obtained.

[0231] (d22) Perform loss calculation

[0232] Based on the segmentation results predicted in step (d21) and the actual segmentation labels, the segmentation loss L is calculated. seg Based on the predicted dust concentration and the actual concentration, the regression loss L is calculated. reg Combining the two loss functions, we obtain the total loss L;

[0233] (d23) Perform backpropagation

[0234] Backpropagation is performed based on the total loss L obtained in step (d22) to calculate the gradient of each layer of the model;

[0235] (d23) Perform parameter updates

[0236] The optimizer updates model parameters based on gradients to minimize the total loss. During training, the learning rate is adaptively adjusted based on the model's performance on the validation set to avoid the model getting trapped in local optima. In this embodiment, the optimizer uses stochastic gradient descent to calculate the gradient for each data sample during training and update the model parameters, iteratively approximating the minimum value of the objective function. The update rule is as follows:

[0237]

[0238] Where, θ t It is a parameter; α is the learning rate; It is calculated based on the current sample (x) i ,y i (d24) Gradient of the loss function; (d24) Hyperparameter tuning, specifically including:

[0239] Based on the model convergence, a learning rate scheduler is used to gradually adjust the learning rate, and the batch size is adjusted according to hardware resources to balance training speed and model stability. λ is dynamically adjusted based on experimental feedback. reg and λ seg The weights are adjusted to ensure a reasonable proportion of the losses from segmentation and regression tasks in the total loss. This embodiment uses StepLR as the learning rate scheduler. StepLR multiplies the current learning rate by a predetermined decay factor after a certain number of training iterations, thereby gradually reducing the learning rate to optimize model performance. The formula for adjusting the learning rate is:

[0240]

[0241] Among them, lr t : Current learning rate; lr0: Initial learning rate; γ: Learning rate decay factor (usually set to 0.1 or 0.5); epoch: Current training epoch; step_size: Step size of learning rate decay, indicating how many epochs the learning rate is reduced once; Round down;

[0242] (d3) Perform model validation, specifically including:

[0243] After model training is completed, the model performance is evaluated using a validation set, and the segmentation accuracy and concentration prediction error are calculated to ensure that the model can adapt to dust detection tasks in actual mining environments. The specific steps include the following:

[0244] (d31) Validation set preparation and preprocessing, specifically including:

[0245] The image data of the validation set divided in step (2) are preprocessed to ensure that the data input format is consistent with the training set;

[0246] (d32) Model evaluation performance metrics, specifically including:

[0247] (d321) Calculate the segmentation performance evaluation metrics, specifically including:

[0248] Accuracy refers to the proportion of correct predictions among all predictions, and is used to measure the overall accuracy in dust separation tasks.

[0249] For dust separation tasks, the formula for accuracy is:

[0250]

[0251] Where: TP is the number of pixels whose true category is "dust" and which are correctly predicted as "dust"; TN is the number of pixels whose true category is "background" and which are correctly predicted as "background"; FP is the number of pixels whose true category is "background" but are incorrectly predicted as "dust"; FN is the number of pixels whose true category is "dust" but are incorrectly predicted as "background".

[0252] Intersection over Union (IoU): This measures the degree of overlap between the predicted result and the true label. For segmentation tasks, it calculates the degree of overlap between the predicted dust region and the real dust region. The higher the IoU, the higher the model segmentation accuracy.

[0253]

[0254] Where A represents the predicted segmentation region; B represents the ground truth segmentation region; |A∩B| represents the number of pixels at the intersection of the predicted region and the ground truth region; |AUB| represents the number of pixels at the union of the predicted region and the ground truth region.

[0255] The F1 score is calculated by taking into account both precision and recall in the segmentation process to evaluate the model's segmentation performance.

[0256]

[0257]

[0258]

[0259] (d322) Calculate the concentration prediction error assessment index, specifically including:

[0260] Mean squared error (MSE): Used to measure the difference between predicted and actual dust concentration values;

[0261]

[0262] Where: y i It is the i-th true value; is the i-th predicted value; n is the total number of samples;

[0263] Mean Absolute Error (MAE): This measures the average absolute error between the predicted and the actual values.

[0264]

[0265] (d33) Save the optimal model, specifically including:

[0266] Based on the performance metrics calculated in step (d32), check the model's splitting and regression task fluctuations on the validation set to ensure that the model's prediction results are stable on different samples, with no obvious overfitting or underfitting. The model parameters that achieve the best performance on the validation set will be saved for use in dust detection and concentration prediction in practical applications.

[0267] (d4) Perform model testing. Use the test set in step (2) to test the validated model and test its performance.

[0268] (4) Utilizing an optimized model for dust concentration detection, specifically including: using the camera equipment installed in step (1), acquiring original image data of the mine roadway through step (2), and constructing an labeled image dataset. The constructed image dataset is then imported into the optimized multi-task learning model for dust detection established in step (3) to obtain multi-scale feature maps. The segmentation head upsamples and outputs segmentation probability maps for the learned feature maps at different scales, sets a threshold of 0.5, and generates a binary mask based on the set threshold. The prediction head extracts dust region features based on the generated binary mask, pools the features of the dust region, and generates feature vectors. The feature vectors are then input into the regression head module for dust concentration prediction, achieving real-time dust identification and concentration detection in the image.

[0269] (5) Visualize and issue an over-limit warning for the dust detection results obtained in step (4). Display the currently detected dust concentration value in real time on an external display screen, with the unit being mg / m³. 3 Different warning thresholds, such as 2 mg / m³, are set according to national or industry standards. 3 The system continuously monitors dust concentration, and triggers an early warning immediately once the threshold is exceeded.

[0270] This embodiment also provides a system based on the above method, characterized in that it includes a sensing and measurement module, a model training module, a model inference module, and a visualization and over-limit early warning module that cooperate with each other.

[0271] Perception and Measurement Module: This module is responsible for collecting image data and dust concentration data in real time in the mine roadway environment, providing data support for the subsequent multi-task learning model for dust detection.

[0272] Specifically, the sensing and measurement module includes: a self-cleaning mining camera for collecting image data of mine roadways; a dust concentration sensor for acquiring dust concentration data inside the roadways; and an industrial-grade Ethernet cable for transmitting high-definition images.

[0273] Model Training Module: This module is responsible for building and training a multi-task learning model for dust detection, and outputting the trained multi-task learning model for dust detection.

[0274] Specifically, the model training module includes: model building, using ResNet-50 as the backbone network and combining it with a bidirectional feature pyramid network (BiFPN) to achieve multi-scale feature fusion; a segmentation head for pixel-level segmentation of dust regions and a regression head for dust concentration prediction; model training and optimization, training the model using a labeled dataset, employing a multi-task loss function that combines segmentation and regression losses, and adjusting hyperparameters to optimize model performance; and model validation, evaluating model performance on a validation set, calculating segmentation accuracy and concentration prediction error to ensure the model's reliability in practical applications.

[0275] Model Inference Module: This module receives real-time images transmitted from the image acquisition module, uses a trained multi-task learning model to perform image segmentation and concentration detection of dust in the images, and then transmits the calculation results to the visualization and over-limit early warning module.

[0276] Visualization and Over-limit Early Warning Module: This module receives the output of the model inference module, visualizes the detection results, and provides an early warning when the dust concentration exceeds the limit.

[0277] Specifically, the visualization and over-limit early warning module includes: an external mine display screen that shows the dust concentration value; and an over-limit early warning device that triggers an early warning when the dust concentration exceeds the safety threshold.

[0278] In this embodiment, the deployment of the detection system is based on a mine roadway, but the application of this embodiment is not limited to underground mines. The system is also applicable to dust concentration detection in dust-producing places such as flour mills and construction sites.

[0279] This system features simple equipment deployment and low maintenance costs, facilitating rapid deployment and use in mining environments, and significantly improving the efficiency and safety of underground inspections. This embodiment utilizes a multi-task learning model-based dust detection technology to achieve end-to-end acquisition of the location and concentration of dust areas in roadways, greatly simplifying the workflow and improving detection efficiency and accuracy. Therefore, this embodiment not only enables real-time and accurate dust concentration monitoring but also provides crucial technical support and decision-making basis for dust control in mines.

[0280] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications and equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A dust detection method based on a multi-task learning model, characterized in that, Includes the following steps: (1) Install and debug data acquisition equipment, specifically including: arranging camera equipment and dust concentration sensors in the mine roadway; debugging and calibrating the camera equipment and sensors; (2) Data collection and dataset construction, specifically including: (2.1) Using the installed camera equipment, continuously collect image data in the mine roadway, covering the dust distribution in different time periods, different areas and different working conditions; synchronously record the data of the dust concentration sensor in the area corresponding to the collected image, and obtain the dust concentration value at the corresponding time and location; (2.2) Use image annotation tools to manually annotate the images acquired in step (2.1), perform manual semantic segmentation on the dust areas in the images, and generate pixel-level annotation masks; divide the tunnel images into two categories: background and dust, and mark each pixel as "Background" or "Dust" to form semantic segmentation labels; for each annotated image, associate the dust concentration value of the image at the corresponding acquisition time with its dust area label, where Background represents the background and Dust represents the dust. (2.3) Constructing the dataset includes: performing random augmentation on the images in step (2.2) to increase data diversity and improve the robustness of the model; organizing and converting the labeled image data into a format that the model can read; constructing a multi-task learning dataset containing images, semantic segmentation labels and dust concentration values; and preprocessing the dataset. (3) Construct and train a multi-task learning model for dust detection, specifically including: (3.1) Constructing a multi-task learning model for dust detection involves constructing a feature extraction network, namely a Backbone, a multi-scale feature fusion module, a segmentation head, and a regression head. The Backbone is responsible for extracting general features of the image; the multi-scale feature fusion module is used to fuse and enhance features of different scales in the image; the segmentation head is used to perform pixel-level classification and locate dust regions; and the regression head is used to predict the concentration value of the dust segmentation region. In the multi-task learning model, the features after multi-scale feature fusion can be shared by multiple subsequent task branches. The construction of the feature extraction network in step (3.1) specifically includes: Feature extraction is performed using ResNet50, and the extracted features consist of: a two-layer input branch consisting of one 7×7 convolutional layer and one 3×3 max-pooling layer; and 48 layers of residual blocks in four stages. Stage 1: Contains 3 residual blocks; Stage 2: Contains 4 residual blocks, and the output feature map is denoted as... ; Stage 3: Contains 6 residual blocks, and the output feature map is denoted as... ; Stage 4: Contains 3 residual blocks, and the output feature map is denoted as... , Each residual block consists of 3 convolutional layers; The number of channels is 512; The number of channels is 1024; The number of channels is 2048; Step (3.1) involves constructing a multi-scale feature fusion module, including: A BiFPN architecture is adopted for multi-scale feature fusion. BiFPN enhances multi-scale features through bidirectional feature fusion, employing both top-down and bottom-up approaches. BiFPN includes: (a1) Generate the initial feature pyramid layer; Generate the initial feature pyramid layers, and then arrange the feature pyramid levels from bottom to top as feature layers. , , , , ; , , It is composed of feature layers , , It is obtained through 1×1 convolution, and the specific calculation formula is as follows: , It is a 1×1 convolutional layer; Add convolution and downsampling operations to P5 to generate feature layers. and The specific calculation formula is as follows: ； ； Among them, Conv 3x3，s=2 The convolution kernel is 3x3 with a stride of s = 2; ReLU is the activation function, ReLU(x) = max(0,x); (a2) Top-down path fusion: Feature fusion is performed from high-level features to low-level features. In the top-down path, the calculation formula for feature fusion is: ； ； in, The first in the top-down path Feature map of the layer; The feature layer is the initial input; The next level, i.e. +1 layer top-down feature map; The weights are learnable non-negative and are guaranteed to be non-negative through the ReLU activation function. To prevent division by zero of small constants, it is usually taken as... ; For upsampling, the feature map size is increased by a factor of 2; It is a 3×3 convolutional layer used for smoothing the fused features; (a3) Bottom-up path fusion; Feature fusion is performed from low-level features to high-level features. In the bottom-up path, the calculation formula for feature fusion is: ； in, The first in the bottom-up path Feature map of the layer; The first in the top-down path Feature map of the layer; This is the bottom-up feature map of the next layer, i.e., the (i-1)th layer; The weights are learnable non-negative and are guaranteed to be non-negative through the ReLU activation function. To prevent division by zero of small constants, it is usually taken as... ; For downsampling operations, the feature map size is reduced by a factor of 2; It is a 3×3 convolutional layer used for smoothing the fused features; (a4) Learnable fusion weights; In each fusion node in steps (a2) and (a3), learnable weights are introduced to adjust the contributions of different feature layers; initial weights are set. During training, the weight parameters are automatically learned through backpropagation, enabling the BiFPN multi-scale feature fusion module to output feature maps at different scales. , , , , ; (3.2) The model constructed in step (3.1) is trained and optimized. Specifically, the image dataset after preprocessing in step (2) is randomly divided into a 70% training set, a 20% validation set, and a 10% test set. The training set is used for training the model. The multi-task learning model for dust detection constructed in step (3.1) is trained. During the training process, the model optimizes both the segmentation and regression tasks. By combining the segmentation loss and the regression loss, the optimal performance of the model in these two tasks is obtained. The validation set is used to evaluate the performance of the model and perform hyperparameter tuning during the training process. The test set is used to evaluate the final performance of the model after training. (4) Utilize the optimized model to detect dust concentration, specifically including: using the camera equipment installed in step (1), acquiring the original image data of the mine roadway through step (2), and constructing an labeled image dataset. Import the constructed image dataset into the optimized dust detection multi-task learning model established in step (3) to obtain multi-scale feature maps. The segmentation head upsamples and outputs a segmentation probability map for the learned feature maps of different scales, sets the threshold to 0.5, and generates a binary mask according to the set threshold. The prediction head extracts the dust area features based on the generated binary mask, pools the features of the dust area, generates a feature vector, and inputs the feature vector into the regression head module to predict the dust concentration, thereby realizing real-time identification and concentration detection of image dust. (5) Visualize and issue an over-limit warning for the dust detection results obtained in step (4); Step (3.1) constructing the segmentation head includes: receiving feature maps from the multi-scale feature fusion module, upsampling the low-resolution feature map, and fusing feature maps of different scales to finally generate a feature map with the same size as the input image. Each pixel in the feature map is then classified to determine whether it belongs to "Dust" or "Background". Specifically, this includes the following steps: (b1) Perform feature layer adjustments, specifically including: For the multi-scale feature layer of BiFPN from step (a4): , , , , Channel alignment is performed, and a 1×1 convolution is applied to each feature layer to obtain... to : ; in, The adjusted feature layer; (b2) Perform multi-scale feature fusion and upsampling, specifically including: A bottom-up upsampling strategy is adopted to progressively upsample the low-resolution feature map and then fuse it with the high-resolution feature map. Initially, the features are progressively upsampled and fused with higher-resolution feature layers, resulting in fused features. ; for Perform initial sampling. ; for Perform the following steps: (b21) Fuse features from the previous layer Upsampled to the current adjusted feature layer The size of the sample is determined by upsampling operations using bilinear interpolation or deconvolution. , These are the features after upsampling; (b22) Upsample the features Compared with the current adjustment The fusion process is performed using a splicing method: ； in, It is a 3×3 convolutional layer with C output channels; Obtain the highest resolution feature layer The size is ; Where H is the height of the input image; W is the width of the input image; (b22) will Stepwise upsampling to the size of the input image After each upsampling, features are extracted through a convolutional layer; upsampling to Then, convolution processing is performed: Upsampling Then, convolution processing is performed: Upsampling Then, convolution processing is performed: , It is an image feature map with the same size as the original image; F up1 For the highest resolution feature layer Feature layer upsampling operation parameters; F up2 For the highest resolution feature layer Upsampling operation parameters for feature layers with two upsampling layers; F up3 For the highest resolution feature layer Upsampling parameters for feature layers with 3 upsampling layers; (b3) Split the output layer, specifically including: Pixel-level classification is performed on the upsampled features from step (b22), and the segmentation results and fused feature maps are output, specifically including: (b31) Using the sampled features at the 1×1 convolutional layer Perform a transformation to adjust the number of channels in the feature map, and output the adapted feature map O: ； (b32) Using the Sigmoid activation function, the output is mapped to the (0,1) range, representing the probability that a pixel belongs to "dust": ； in, It is the Sigmoid activation function. ; (b33) Generate a binary segmentation mask based on the probability from step (b32): ； in, The threshold is set to 0.5; S(x,y) represents the probability value that any pixel (x,y) on feature map O belongs to the "dust" class. A binary segmentation mask is generated based on the threshold to obtain the final segmentation result. Step (3.1) of constructing the regression head specifically includes: (c1) Extract features from the dust area: The process involves calling up the fused and upsampled high-resolution feature maps from the segmentation head, focusing on dust-related regions in the image, and filtering out background and other irrelevant information. Specifically, this includes the following steps: (c1) Obtain the feature map O from the segmentation head, denoted as The size is ; (c12) The segmentation mask obtained in step (b33) Applied to feature maps A masking operation is performed on each pixel location, retaining only the features of the dust area and setting the features of the background area to zero. The specific calculation formula is as follows: ； in, The feature map O in the segmentation head after one-dimensional convolution is... This is the feature map after the masking operation, where only the features of the "dust" area are retained, and the background area is set to zero. (c2) Perform feature aggregation and global representation: The feature map of the dust region is converted into a fixed-length feature vector to provide input for subsequent regression calculations. The feature vector of the dust region is then calculated. : ； in, The total number of pixels in the dusty area is calculated as follows: ； in, The dimension is That is, the number of channels in the feature map, if This indicates that no dust was detected in the area. (c3) Calculation of regression layer and dust concentration output: A fully connected layer is used to perform regression calculation on the feature vector, outputting the dust concentration value D: D= ； in, This is the weight matrix; It is a biased scalar; The calculation method is as follows: ； in, Output for hidden layer; This is the weight matrix; It is a biased scalar; Step (3.2) specifically includes: (d1) Define the total loss function L: ； in, The regression loss is used to measure the error in dust concentration prediction. The segmentation loss is used to measure the loss in dust image segmentation. and The loss weight coefficients control the contributions of the segmentation and regression tasks to the total loss; (d2) Model training is performed, specifically including: (d21) Perform forward propagation; The images of the training set divided in step (2) are input into the model constructed in step (3.1) to perform joint inference of the segmentation head and the regression head, and the dust segmentation results and concentration prediction values ​​are obtained. (d22) Perform loss calculation; Based on the segmentation results predicted in step (d21) and the actual segmentation labels, the segmentation loss is calculated. Based on the predicted dust concentration and the actual concentration, the regression loss is calculated. Combining the two loss functions, we obtain the total loss. ; (d23) Perform backpropagation; The total loss obtained based on step (d22) Perform backpropagation to calculate the gradients of each layer of the model; (d23) Update the parameters; The optimizer updates the model parameters based on the gradient to minimize the total loss. During training, the learning rate is adaptively adjusted based on the model's performance on the validation set to avoid the model getting stuck in local optima. (d24) Perform hyperparameter tuning, specifically including: Based on the model convergence, a learning rate scheduler is used to gradually adjust the learning rate, and the batch size is adjusted according to hardware resources to balance training speed and model stability. Dynamic adjustments are also made based on experimental feedback. and The weights are determined to ensure that the losses from the segmentation and regression tasks are in a reasonable proportion of the total loss. (d3) Perform model validation, specifically including: After model training is completed, the model performance is evaluated using a validation set, and the segmentation accuracy and concentration prediction error are calculated to ensure that the model can adapt to dust detection tasks in actual mining environments. The specific steps include the following: (d31) Validation set preparation and preprocessing, specifically including: Preprocess the image data of the validation set divided in step (2) to ensure that the data input format is consistent with the training set; (d32) Calculation performance indicators, specifically including: (d321) Calculate the segmentation performance evaluation metrics, specifically including: Accuracy refers to the proportion of correct predictions among all predictions, and is used to measure the overall accuracy in dust separation tasks. For dust separation tasks, the formula for accuracy is: Accuracy= ; Where: TP is the number of pixels whose true category is "dust" and which are correctly predicted as "dust"; TN is the number of pixels whose true category is "background" and which are correctly predicted as "background"; FP is the number of pixels whose true category is "background" but are incorrectly predicted as "dust"; FN is the number of pixels whose true category is "dust" but are incorrectly predicted as "background". Intersection over Union (IoU): This measures the degree of overlap between the predicted result and the true label. For segmentation tasks, it calculates the degree of overlap between the predicted dust region and the real dust region. The higher the IoU, the higher the model segmentation accuracy. IoU= ; Where A represents the predicted segmentation region; Represents the actual segmented region; This represents the number of pixels that intersect the predicted region and the actual region. This represents the number of pixels in the union of the predicted region and the true region. Calculate the F1 score: It takes into account the precision and recall of the segmentation to evaluate the segmentation performance of the model. Precision= ; Recall= ; ； (d322) Calculate the concentration prediction error assessment index, specifically including: Mean squared error (MSE): used to measure the difference between predicted and actual dust concentration values; ； Where: y k It is the kth true value; It is the k-th predicted value; n is the total number of samples; Mean Absolute Error (MAE): This measures the average absolute error between the predicted and the actual values. ； (d33) Save the optimal model, specifically including: Based on the performance metrics calculated in step (d32), check the model's splitting and regression task fluctuations on the validation set to ensure that the model's prediction results are stable on different samples, with no obvious overfitting or underfitting. The model parameters that achieve the best performance on the validation set will be saved for use in dust detection and concentration prediction in practical applications. (d4) Perform model testing. Use the test set in step (2) to test the validated model and test its performance.

2. The dust detection method based on a multi-task learning model according to claim 1, characterized in that, The image annotation tool in step (2.2) is LabelImg.

3. The dust detection method based on a multi-task learning model according to claim 1, characterized in that, The random enhancement operations in step (2.3) include image rotation, translation, scaling, and color jitter.

4. The dust detection method based on a multi-task learning model as described in claim 1, characterized in that: The preprocessing operations for the dataset in step (2.3) include resizing the images and normalizing their colors.

5. A system based on the dust detection method based on the multi-task learning model as described in claim 1, characterized in that, The system comprises a sensory measurement module, a model training module, a model inference module, and a visualization and over-limit early warning module that work together. The sensory measurement module collects image data and dust concentration data from the mine environment. The model training module builds and trains a multi-task learning model, including data loading, model definition, loss function setting, and model optimization. The model inference module runs the trained model in the deployment environment, performs dust detection and concentration prediction on real-time acquired images, and optimizes the model's inference speed. The visualization and over-limit early warning module visualizes the detection results and concentration values, and triggers an alarm mechanism to alert relevant personnel when the dust concentration exceeds a preset threshold. The sensory measurement module is responsible for collecting environmental data and transmitting it to the model training module. The model training module receives environmental data, trains the multi-task learning model, and outputs the trained multi-task learning model to the model inference module. The model inference module processes the on-site collected data through a multi-task learning model and outputs the dust concentration prediction results to the visualization and over-limit early warning module. The visualization and over-limit early warning module visualizes the prediction results and issues early warnings based on the set dust concentration threshold.

Images