Coal mine hydraulic system particle identification filtering method and system based on deep learning

Abstract

The application discloses a kind of coal mine hydraulic system particle identification filtering method and system based on deep learning, comprising: S1: the image of coal mine hydraulic oil is collected, and background difference and histogram equalization are carried out, to obtain the preprocessed hydraulic oil image;S2: calculate multi-scale feature map, spatial weight matrix and fusion feature map, finally calculate the fusion feature map of key area enhancement;S3: the fusion feature map of key area enhancement is carried out local response normalization, to obtain preliminary standardized feature map, final standardized feature map;S4: according to final standardized feature map, calculate space-channel joint gate, refine feature, calculate pollution degree thermodynamic diagram;S5: calculate pollution degree global statistical value, then image is evenly divided into 4 regions and the weighted pollution index of each region is calculated, finally calculate the probability distribution of hydraulic oil filtering strategy.The application can solve the problem of insufficient precision of traditional method in identifying complex particle contaminants.

Description

Technical Field

[0001] This invention relates to the field of coal mining technology, and in particular to a particle identification and filtering method and system for coal mine hydraulic systems based on deep learning. Background Technology

[0002] The hydraulic system in coal mines is a core component of coal mine machinery and equipment, and its operational stability directly affects production safety and efficiency. However, hydraulic systems are often affected by particulate contaminants during operation. These contaminants originate from equipment wear, external dust, and deteriorated lubricating oil, and are characterized by their complexity, small size, and uneven distribution, easily leading to wear, blockage, and failure of hydraulic components, thereby causing system malfunctions.

[0003] Traditional hydraulic filtration methods primarily rely on mechanical filter elements. While they can remove some larger particles, their filtration efficiency for fine particles is limited, and they cannot identify the type of contaminant. Furthermore, traditional methods lack real-time monitoring and intelligent processing capabilities, making it impossible to dynamically adjust filtration strategies to cope with complex operating conditions, thus increasing maintenance costs and the risk of failure.

[0004] In recent years, deep learning technology has demonstrated outstanding performance in image recognition and classification, enabling it to extract features from complex data and achieve accurate identification. However, the application of deep learning in the identification of particulate contaminants in coal mine hydraulic systems faces many challenges: the particulate contaminants in hydraulic oil are diverse in type and form, and exhibit low contrast and high noise imaging characteristics in the oil, which places higher demands on the feature extraction capabilities and anti-interference ability of deep learning models. Summary of the Invention

[0005] In view of this, the present invention aims to provide a particle identification and filtering method and system for coal mine hydraulic systems based on deep learning, so as to solve the problem of insufficient accuracy in identifying complex particulate pollutants by traditional methods.

[0006] A deep learning-based particle identification and filtering method and system for coal mine hydraulic systems includes:

[0007] S1: Acquire images of hydraulic oil in the coal mine hydraulic system, and perform background subtraction and histogram equalization to obtain preprocessed hydraulic oil images;

[0008] S2: Based on the preprocessed hydraulic oil image, calculate the multi-scale feature map, then generate the spatial weight matrix and fused feature map, and finally enhance the key regions to obtain the enhanced fused feature map of the key regions, including:

[0009] The gradient of the fused feature map is calculated, and the gradient is added to the fused feature map and then a spatial attention mask is generated by passing it through a convolutional layer and a nonlinear activation function.

[0010] The spatial attention mask is multiplied element-wise with the sharpened convolutional fusion feature map, and then superimposed onto the original fusion feature map to obtain a fusion feature map with enhanced key regions.

[0011] S3: Perform local response normalization calculation on the enhanced fusion feature map of the key region to obtain a preliminary standardized feature map, then calculate the adaptive correction factor to obtain the final standardized feature map;

[0012] S4: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution degree heatmap based on the refined features, including:

[0013] After multiplying the space-channel joint gate element-by-element with the final standardized feature map, it is concatenated with the hidden state of the previous time step. The hidden state of the current time step is updated by weight matrix and hyperbolic tangent function. After weighted summation of the hidden states of all time steps, a refined feature is generated by linear transformation.

[0014] S5: Based on the pollution heat map, calculate the global pollution statistics, then divide the image into four regions and calculate the weighted pollution index for each region. Finally, calculate the probability distribution of the hydraulic oil filtration strategy.

[0015] Furthermore, step S1 also includes:

[0016] S11: Acquire images of coal mine hydraulic oil using a high-speed microscopic imaging device;

[0017] S12: Based on the image of hydraulic oil in coal mines, the oil flow interference is separated by the background difference method to obtain the denoised hydraulic oil image;

[0018] S13: Based on the denoised hydraulic oil image, the pixel distribution is adjusted using an adaptive histogram equalization algorithm to obtain the preprocessed hydraulic oil image.

[0019] Furthermore, step S2 also includes:

[0020] The specific method for generating the fused feature map includes: using parallel three-way convolution, and employing convolution kernels of the first, second, and third sizes respectively to extract features from the preprocessed hydraulic oil image, generating a first-scale feature map, a second-scale feature map, and a third-scale feature map;

[0021] The first scale feature map, the second scale feature map, and the third scale feature map are concatenated, and a spatial weight matrix is ​​generated through convolutional layers and normalization. The spatial weight matrix is ​​then sliced ​​by channels to obtain the first spatial weight matrix, the second spatial weight matrix, and the third spatial weight matrix, respectively.

[0022] The first spatial weight matrix is ​​multiplied element-wise with the first scale feature map, the second spatial weight matrix is ​​multiplied element-wise with the second scale feature map, and the third spatial weight matrix is ​​multiplied element-wise with the third scale feature map. The three weighted results are then summed to obtain the fused feature map.

[0023] Furthermore, step S2 also includes:

[0024] S21: Based on the preprocessed hydraulic oil image, a multi-scale feature map is generated through parallel three-way convolution. The calculation method is as follows: in, This is the first-scale feature map. This is the second-scale feature map. This is a third-scale feature map. This is a preprocessed image of the hydraulic oil. For convolution, The first-size convolution kernel, For the second-sized convolution kernel, The third-sized convolution kernel;

[0025] S22: Based on the multi-scale feature map, a spatial weight matrix and a fused feature map are generated through a gating mechanism. The calculation method is as follows: in, This is the spatial weight matrix. For the softmax function, It is a convolutional layer. For splicing operations, These are the weight matrices for the first, second, and third spaces, respectively. For channel slicing operations, To fuse feature maps, For Hadamah accumulation;

[0026] S23: Based on the fused feature map, the key regions are enhanced using a residual masking mechanism to obtain the enhanced fused feature map of the key regions. The calculation method is as follows: in, For spatial attention mask, For the sigmoid function, For the feature map gradient, For Sobel horizontal operators, For Sobel vertical operators, For matrix transpose, This is the transpose of the Sobel horizontal operator. Enhanced fusion feature maps for key regions To sharpen the convolution kernel. It should be further explained that, in order to address the problems of missed detection of fine particles and insufficient accuracy in identifying blurred edge regions caused by single-scale feature extraction and fixed fusion weights in traditional hydraulic oil particle recognition methods, this invention achieves feature optimization in step S2 through multi-scale dynamic weighted fusion and gradient-driven residual enhancement;

[0027] In terms of technical features, this invention firstly extracts the macroscopic distribution, mesoscopic contour, and microscopic texture features of particles simultaneously through parallel multi-size convolutional kernels, covering the characteristics of particles of different sizes. Secondly, it uses a gating mechanism to perform spatially adaptive weighted fusion of multi-scale features, dynamically allocating the contribution of each scale feature in spatial location through a normalized weight matrix, overcoming the adaptability defects of traditional fixed-weight fusion for complex particle distributions. Finally, it combines gradient magnitude and residual masking mechanisms to superimpose sharpening and enhancement features in areas of significant gradient in the feature map (such as particle edges and texture abrupt changes), using the physical properties of gradients to enhance the recognizability of low-contrast particles. From a mathematical perspective, multi-scale convolution constructs feature complementarity through parallel computation of different receptive fields, the Softmax function of the gating mechanism constrains the sum of weights to 1, avoiding feature magnitude imbalance, and the square root operation of the sum of squares of gradient magnitudes is essentially a vector space magnitude calculation, quantifying the intensity of feature abrupt changes to guide attention focus.

[0028] Compared to the previous algorithm, this invention can improve the feature separation degree in the particle overlapping area and reduce the edge localization error. In the scenario of high noise and coexistence of multi-scale particles in hydraulic oil, this invention adapts to the changes in oil turbidity through dynamic weights and resists the edge weakening problem through gradient enhancement mechanism, thus solving the difficulty of insufficient generalization ability of traditional algorithms in complex industrial environments.

[0029] Furthermore, step S3 also includes:

[0030] S31: Based on the enhanced fused feature map of the key region, a preliminary standardized feature map is obtained through local response normalization layer calculation. The calculation method is as follows: in, For the initial standardization of feature maps, This is a local response normalization algorithm;

[0031] S32: Based on the preliminary standardized feature map, calculate the adaptive correction factor. The calculation method is as follows:

[0032] in, For adaptive correction factor, For global average pooling, For adaptive correction factor weight matrix;

[0033] S33: Based on the preliminary standardized feature map and the adaptive correction factor, the final standardized feature map is calculated as follows:

[0034] in, This is the final standardized feature map.

[0035] Furthermore, step S4 also includes:

[0036] The calculation method of the spatial-channel joint gate includes: performing 3×3 convolution and 1×1 convolution on the final standardized feature map respectively, activating the two convolution results through the Sigmoid function to generate a spatial gate and a channel gate respectively, and multiplying the spatial gate and the channel gate element by element to obtain the spatial-channel joint gate;

[0037] The method for generating the pollution heatmap includes: performing 3×3 convolutions on the refined features sequentially to generate a primary decoding feature map; adding the primary decoding feature map to the refined features and performing a second 3×3 convolution to generate an enhanced decoding feature map; and performing a 1×1 convolution on the enhanced decoding feature map and activating it with the Sigmoid function to obtain the pollution heatmap.

[0038] Furthermore, step S4 also includes:

[0039] S41: Based on the final standardized feature map, calculate the spatial-channel joint gate. The calculation method is as follows: in, For space gate, This is a convolutional layer with a kernel size of 3×3. For passageway doors, For a convolutional layer with a kernel size of 1×1, It is a combined space-passage door;

[0040] S42: Calculate the refined features based on the final standardized feature map and the space-channel joint gate. The calculation method is as follows: in, Let t be the hidden state at step t, where t is the time step index. It is the hyperbolic tangent function. The hidden state weight matrix is... This is the hidden state at step t-1. This is the final standardized feature map at step t. For refined features, To refine the feature weight matrix, This represents the total number of recursive steps.

[0041] S43: Calculate the pollution degree heat map based on the refined characteristics. The calculation method is as follows: in, This is a primary decoding feature map. To enhance the decoded feature map, This is a heatmap of contamination levels. It should be further explained that, addressing the problem of delayed dynamic contamination state response caused by insufficient utilization of temporal features and a simplistic decoding process in traditional hydraulic oil contamination identification methods, this invention achieves accurate contamination level modeling in step S4 through a spatial-channel joint gating mechanism and multi-level residual decoding.

[0042] In terms of technical features, this invention firstly extracts the spatial distribution pattern of contaminated areas and the channel-level contaminated features based on bi-branch convolution, and then fuses the two through gated multiplication to form a joint attention weight, thereby achieving collaborative screening of contaminated sensitive areas. Secondly, this invention introduces a temporal recursive mechanism to iteratively fuse the joint gated features with historical hidden states, and uses the nonlinear mapping of the hyperbolic tangent function to gradually fit the dynamic change law of contaminated features. The random error of single-step prediction is eliminated by weighted aggregation of multi-step hidden states. Finally, a two-level residual decoding structure is adopted. The core contaminated area features are captured through primary decoding, and then the original refined features are superimposed for secondary decoding. Cross-level feature complementarity is used to enhance the distinguishability of microparticles and diffuse contamination.

[0043] Compared with traditional static decoding algorithms, this invention improves the boundary intersection-over-union ratio of contaminated areas and reduces the false detection rate in low-contrast samples, thereby achieving holographic perception of contamination status under complex working conditions.

[0044] Furthermore, step S5 also includes:

[0045] S51: Calculate the global statistical value of pollution based on the pollution heat map. The calculation method is as follows: in, This is a global statistical value for pollution levels. The height of the pollution heatmap, The width of the pollution heatmap, For height indexing, For width index, For indicator functions, This represents the pixel value with height i and width j in the pollution intensity heatmap. For the pollution level adaptive threshold, It is a multilayer perceptron;

[0046] S52: Based on the pollution heatmap, the image is evenly divided into 4 regions, and the weighted pollution index for each region is calculated as follows: in, Here, k is the weighted pollution index for the k-th region, and k is the region index. For the k-th region, the region weight parameter is used.

[0047] S53: Calculate the probability distribution of hydraulic oil filtration strategies based on the global pollution statistics and the weighted pollution index of each region. The calculation method is as follows: in, The probability distribution of hydraulic oil filtration strategies. These are the weighted pollution indices for the first, second, third, and fourth regions, respectively.

[0048] This invention also discloses a deep learning-based particle identification and filtration system for coal mine hydraulic systems, comprising:

[0049] Coal Mine Hydraulic Oil Image Acquisition and Preprocessing Module: Acquires images of coal mine hydraulic oil and performs background subtraction and histogram equalization to obtain preprocessed hydraulic oil images;

[0050] Key region enhancement module: Based on the preprocessed hydraulic oil image, calculate multi-scale feature maps, generate spatial weight matrix and fusion feature map, and finally enhance key regions to obtain fusion feature map of key region enhancement;

[0051] Standardized feature map calculation module: Performs local response normalization calculation on the enhanced fusion feature map of key regions to obtain a preliminary standardized feature map, then calculates the adaptive correction factor, and finally calculates the final standardized feature map;

[0052] Pollution heat map calculation module: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution heat map based on the refined features;

[0053] Hydraulic oil filtration strategy calculation module: Based on the contamination heat map, calculate the global statistical value of contamination, then divide the image evenly into 4 regions and calculate the weighted contamination index of each region, and finally calculate the probability distribution of hydraulic oil filtration strategy.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] (1) To address the problem of insufficient accuracy in identifying complex particulate pollutants using traditional methods, this invention firstly achieves multi-level extraction of particulate physical properties through parallel multi-scale convolution and gradient-driven residual enhancement, thus solving the problem of detail loss caused by traditional single-scale feature extraction; secondly, it overcomes the technical bottleneck of the traditional static decoding mechanism's lagging response to dynamic pollution by using spatial-channel joint gating and recursive temporal refinement; finally, it improves the accuracy of identifying particulate pollutants in coal mine hydraulic oil by generating filtering decisions through the pollution degree heatmap.

[0056] (2) This invention innovatively introduces a parallel three-way convolution and gated spatial weight allocation mechanism. Through the synergistic effect of convolution kernels of different sizes, it simultaneously captures the macroscopic distribution, mesoscopic contour and microscopic texture features of particles. It also utilizes the gradient of the fused feature map and combines it with the residual mask to achieve adaptive enhancement of key regions, effectively solving the problem of missed detection of fine particles caused by the fixed receptive field of traditional single-way convolutional networks.

[0057] (3) This invention proposes a spatial-channel dual-dimensional attention gate and multi-step hidden state fusion method. The spatial gate focuses on the spatial distribution of the pollution area, the channel gate strengthens the pollution type characteristics, and the recursive network is used to model the diffusion law of pollution concentration with time step. This overcomes the lag problem of traditional static decoders in tracking dynamic pollution state and enhances the distinguishability of micro particles and diffuse pollution. Attached Figure Description

[0058] Figure 1 A schematic flowchart of the particle identification and filtering method for coal mine hydraulic systems based on deep learning provided by the present invention;

[0059] Figure 2 A schematic diagram of the algorithm flow for enhancing key regions provided by this invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0061] Example 1: A particle identification and filtering method for coal mine hydraulic systems based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0062] S1: Acquire images of hydraulic oil in the coal mine hydraulic system, and perform background subtraction and histogram equalization to obtain preprocessed hydraulic oil images;

[0063] S11: Acquire images of coal mine hydraulic oil using a high-speed microscopic imaging device;

[0064] S12: Based on the image of hydraulic oil in coal mines, the oil flow interference is separated by the background difference method to obtain the denoised hydraulic oil image;

[0065] S13: Based on the denoised hydraulic oil image, the pixel distribution is adjusted using an adaptive histogram equalization algorithm to obtain the preprocessed hydraulic oil image.

[0066] S2: Based on the preprocessed hydraulic oil image, calculate the multi-scale feature map, then generate the spatial weight matrix and fused feature map, and finally enhance the key regions to obtain the enhanced fused feature map of the key regions, such as... Figure 2 As shown, it includes:

[0067] The gradient of the fused feature map is calculated, and the gradient is added to the fused feature map and then a spatial attention mask is generated by passing it through a convolutional layer and a nonlinear activation function.

[0068] The spatial attention mask is multiplied element-wise with the sharpened convolutional fusion feature map, and then superimposed onto the original fusion feature map to obtain a fusion feature map with enhanced key regions.

[0069] S21: Based on the preprocessed hydraulic oil image, a multi-scale feature map is generated through parallel three-way convolution. The calculation method is as follows: in, This is the first-scale feature map. This is the second-scale feature map. This is a third-scale feature map. This is a preprocessed image of the hydraulic oil. For convolution, The first-size convolution kernel, For the second-sized convolution kernel, The third-sized convolution kernel;

[0070] S22: Based on the multi-scale feature map, a spatial weight matrix and a fused feature map are generated through a gating mechanism. The calculation method is as follows: in, This is the spatial weight matrix. For the softmax function, It is a convolutional layer. For splicing operations, These are the weight matrices for the first, second, and third spaces, respectively. For channel slicing operations, To fuse feature maps, For Hadamah accumulation;

[0071] S23: Based on the fused feature map, the key regions are enhanced using a residual masking mechanism to obtain the enhanced fused feature map of the key regions. The calculation method is as follows: in, For spatial attention mask, For the sigmoid function, For the feature map gradient, For Sobel horizontal operators, For Sobel vertical operators, For matrix transpose, This is the transpose of the Sobel horizontal operator. Enhanced fusion feature maps for key regions To sharpen the convolution kernel.

[0072] It should be further explained that, in order to address the problems of missed detection of fine particles and insufficient accuracy in identifying blurred edge regions caused by single-scale feature extraction and fixed fusion weights in traditional hydraulic oil particle identification methods, this invention achieves feature optimization in step S2 through multi-scale dynamic weighted fusion and gradient-driven residual enhancement.

[0073] In terms of technical features, this invention firstly extracts the macroscopic distribution, mesoscopic contour, and microscopic texture features of particles simultaneously through parallel multi-size convolutional kernels, covering the characteristics of particles of different sizes. Secondly, it uses a gating mechanism to perform spatially adaptive weighted fusion of multi-scale features, dynamically allocating the contribution of each scale feature in spatial location through a normalized weight matrix, overcoming the adaptability defects of traditional fixed-weight fusion for complex particle distributions. Finally, it combines gradient magnitude and residual masking mechanisms to superimpose sharpening and enhancement features in areas of significant gradient in the feature map (such as particle edges and texture abrupt changes), using the physical properties of gradients to enhance the recognizability of low-contrast particles. From a mathematical perspective, multi-scale convolution constructs feature complementarity through parallel computation of different receptive fields, the Softmax function of the gating mechanism constrains the sum of weights to 1, avoiding feature magnitude imbalance, and the square root operation of the sum of squares of gradient magnitudes is essentially a vector space magnitude calculation, quantifying the intensity of feature abrupt changes to guide attention focus.

[0074] Compared to the previous algorithm, this invention can improve the feature separation degree in the particle overlapping area and reduce the edge localization error. In the scenario of high noise and coexistence of multi-scale particles in hydraulic oil, this invention adapts to the changes in oil turbidity through dynamic weights and resists the edge weakening problem through gradient enhancement mechanism, thus solving the difficulty of insufficient generalization ability of traditional algorithms in complex industrial environments.

[0075] S3: Perform local response normalization calculation on the enhanced fusion feature map of the key region to obtain a preliminary standardized feature map, then calculate the adaptive correction factor to obtain the final standardized feature map;

[0076] S31: Based on the enhanced fused feature map of the key region, a preliminary standardized feature map is obtained through local response normalization layer calculation. The calculation method is as follows: in, For the initial standardization of feature maps, This is a local response normalization algorithm;

[0077] S32: Based on the preliminary standardized feature map, calculate the adaptive correction factor. The calculation method is as follows:

[0078] in, For adaptive correction factor, For global average pooling, For adaptive correction factor weight matrix;

[0079] S33: Based on the preliminary standardized feature map and the adaptive correction factor, the final standardized feature map is calculated as follows: in, This is the final standardized feature map.

[0080] S4: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution degree heatmap based on the refined features, including:

[0081] After multiplying the space-channel joint gate element-by-element with the final standardized feature map, it is concatenated with the hidden state of the previous time step. The hidden state of the current time step is updated by weight matrix and hyperbolic tangent function. After weighted summation of the hidden states of all time steps, a refined feature is generated by linear transformation.

[0082] S41: Based on the final standardized feature map, calculate the spatial-channel joint gate. The calculation method is as follows: in, For space gate, This is a convolutional layer with a kernel size of 3×3. For passageway doors, For a convolutional layer with a kernel size of 1×1, It is a combined space-passage door;

[0083] S42: Calculate the refined features based on the final standardized feature map and the space-channel joint gate. The calculation method is as follows: in, Let t be the hidden state at step t, where t is the time step index. It is the hyperbolic tangent function. The hidden state weight matrix is... This is the hidden state at step t-1. This is the final standardized feature map at step t. For refined features, To refine the feature weight matrix, This represents the total number of recursive steps.

[0084] S43: Calculate the pollution degree heat map based on the refined characteristics. The calculation method is as follows:

[0085] in, This is a primary decoding feature map. To enhance the decoded feature map, This is a heat map of pollution levels.

[0086] It should be further explained that, in order to address the problem of delayed dynamic pollution state response caused by insufficient utilization of temporal features and the simplistic decoding process in traditional hydraulic oil pollution identification methods, this invention achieves accurate pollution degree modeling in step S4 through a spatial-channel joint gating mechanism and multi-level residual decoding.

[0087] In terms of technical features, this invention firstly extracts the spatial distribution pattern of contaminated areas and the channel-level contaminated features based on bi-branch convolution, and then fuses the two through gated multiplication to form a joint attention weight, thereby achieving collaborative screening of contaminated sensitive areas. Secondly, this invention introduces a temporal recursive mechanism to iteratively fuse the joint gated features with historical hidden states, and uses the nonlinear mapping of the hyperbolic tangent function to gradually fit the dynamic change law of contaminated features. The random error of single-step prediction is eliminated by weighted aggregation of multi-step hidden states. Finally, a two-level residual decoding structure is adopted. The core contaminated area features are captured through primary decoding, and then the original refined features are superimposed for secondary decoding. Cross-level feature complementarity is used to enhance the distinguishability of microparticles and diffuse contamination.

[0088] Compared with traditional static decoding algorithms, this invention improves the boundary intersection-over-union ratio of contaminated areas and reduces the false detection rate in low-contrast samples, thereby achieving holographic perception of contamination status under complex working conditions.

[0089] S5: Based on the pollution heat map, calculate the global pollution statistics, then divide the image into 4 regions and calculate the weighted pollution index of each region, and finally calculate the probability distribution of hydraulic oil filtration strategy.

[0090] S51: Calculate the global statistical value of pollution based on the pollution heat map. The calculation method is as follows: in, This is a global statistical value for pollution levels. The height of the pollution heatmap, The width of the pollution heatmap, For height indexing, For width index, For indicator functions, This represents the pixel value with height i and width j in the pollution intensity heatmap. For the pollution level adaptive threshold, It is a multilayer perceptron;

[0091] S52: Based on the pollution heatmap, the image is evenly divided into 4 regions, and the weighted pollution index for each region is calculated as follows: in, Here, k is the weighted pollution index for the k-th region, and k is the region index. For the k-th region, the region weight parameter is used.

[0092] S53: Calculate the probability distribution of hydraulic oil filtration strategies based on the global pollution statistics and the weighted pollution index of each region. The calculation method is as follows:

[0093] in, The probability distribution of hydraulic oil filtration strategies. These are the weighted pollution indices for the first, second, third, and fourth regions, respectively.

[0094] For example, after a coal mine hydraulic support has been working continuously for 40 hours, the system detected an abnormal hydraulic oil contamination level. After processing in steps S1-S4, the contamination heat map showed that there were a large number of metal wear particles with a particle size of 15-40μm in the oil, mainly distributed in the oil circulation area, while local areas showed the accumulation of oxide precipitates with a particle size of less than 10μm.

[0095] At this point, the adaptive threshold dynamically calculated by the multilayer perceptron based on the heat map... Global pollution level ;

[0096] The heatmap is divided into four evenly distributed regions:

[0097] Area 1 (top left): near the oil inlet, weight w1=0.4 (sensitive area for oil cleanliness)

[0098] Region 2 (top right): Oil return zone, weight w2=0.3

[0099] Region 3 (bottom left): Filter neighborhood, weight w3=0.2

[0100] Region 4 (bottom right): Settlement area at the bottom of the fuel tank, weight w4=0.1

[0101] Weighted pollution index for each region:

[0102] z1=0.72 (dense packing of metal particles)

[0103] z2=0.35 (small amount of suspended particles)

[0104] z3=0.58 (risk of filter clogging)

[0105] z4=0.81 (enrichment of bottom oxide precipitates)

[0106] The input features of the multilayer perceptron are:

[0107] The output policy probability distribution is calculated using Softmax: Corresponding strategy:

[0108] Strategy A (probability 15%): Maintain the current cycle of filtering (suitable for s < 0.3 and local contamination equilibrium).

[0109] Strategy B (probability 10%): Initiate centrifugal filtration (for precipitated contaminants with z4 > 0.6)

[0110] Strategy C (probability 25%): Trigger high-pressure backwashing (preferred when z3>0.5)

[0111] Strategy D (probability 50%): Immediately shut down and change the oil (triggered by s>0.6 and z1>0.7)

[0112] Ultimately, strategy D was chosen as the execution strategy.

[0113] Example 2: This invention also discloses a deep learning-based particle identification and filtration system for coal mine hydraulic systems, comprising:

[0114] Coal Mine Hydraulic Oil Image Acquisition and Preprocessing Module: Acquires images of coal mine hydraulic oil and performs background subtraction and histogram equalization to obtain preprocessed hydraulic oil images;

[0115] Key region enhancement module: Based on the preprocessed hydraulic oil image, calculate multi-scale feature maps, generate spatial weight matrix and fusion feature map, and finally enhance key regions to obtain fusion feature map of key region enhancement;

[0116] Standardized feature map calculation module: Performs local response normalization calculation on the enhanced fusion feature map of key regions to obtain a preliminary standardized feature map, then calculates the adaptive correction factor, and finally calculates the final standardized feature map;

[0117] Pollution heat map calculation module: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution heat map based on the refined features;

[0118] Hydraulic oil filtration strategy calculation module: Based on the contamination heat map, calculate the global statistical value of contamination, then divide the image evenly into 4 regions and calculate the weighted contamination index of each region, and finally calculate the probability distribution of hydraulic oil filtration strategy.

[0119] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0121] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A particle identification and filtering method for coal mine hydraulic systems based on deep learning, characterized in that, Includes the following steps: S1: Acquire images of hydraulic oil in the coal mine hydraulic system, and perform background subtraction and histogram equalization to obtain preprocessed hydraulic oil images; S2: Based on the preprocessed hydraulic oil image, calculate the multi-scale feature map, then generate the spatial weight matrix and fused feature map, and finally enhance the key regions to obtain the enhanced fused feature map of the key regions, including: The specific method for generating the fused feature map includes: using parallel three-way convolution, and employing convolution kernels of the first, second, and third sizes respectively to extract features from the preprocessed hydraulic oil image, generating a first-scale feature map, a second-scale feature map, and a third-scale feature map; The first scale feature map, the second scale feature map, and the third scale feature map are concatenated, and a spatial weight matrix is ​​generated through convolutional layers and normalization. The spatial weight matrix is ​​then sliced ​​by channels to obtain the first spatial weight matrix, the second spatial weight matrix, and the third spatial weight matrix, respectively. The first spatial weight matrix is ​​multiplied element-wise with the first scale feature map, the second spatial weight matrix is ​​multiplied element-wise with the second scale feature map, and the third spatial weight matrix is ​​multiplied element-wise with the third scale feature map. The three weighted results are summed to obtain the fused feature map. The gradient of the fused feature map is calculated, and the gradient is added to the fused feature map and then a spatial attention mask is generated by passing it through a convolutional layer and a nonlinear activation function. The spatial attention mask is multiplied element-wise with the sharpened convolutional fusion feature map, and then superimposed onto the original fusion feature map to obtain a fusion feature map with enhanced key regions. S3: Perform local response normalization calculation on the enhanced fusion feature map of the key region to obtain a preliminary standardized feature map, then calculate the adaptive correction factor to obtain the final standardized feature map; S4: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution degree heatmap based on the refined features, including: The calculation method of the spatial-channel joint gate includes: performing 3×3 convolution and 1×1 convolution on the final standardized feature map respectively, activating the two convolution results through the Sigmoid function to generate a spatial gate and a channel gate respectively, and multiplying the spatial gate and the channel gate element by element to obtain the spatial-channel joint gate; After multiplying the space-channel joint gate element-by-element with the final standardized feature map, it is concatenated with the hidden state of the previous time step. The hidden state of the current time step is updated by weight matrix and hyperbolic tangent function. After weighted summation of the hidden states of all time steps, a refined feature is generated by linear transformation. The method for generating the pollution heatmap includes: performing 3×3 convolutions on the refined features sequentially to generate a primary decoding feature map; adding the primary decoding feature map to the refined features and performing a second 3×3 convolution to generate an enhanced decoding feature map; and performing a 1×1 convolution on the enhanced decoding feature map and activating it with the Sigmoid function to obtain the pollution heatmap. S5: Based on the pollution heat map, calculate the global pollution statistics, then divide the image into 4 regions and calculate the weighted pollution index of each region, and finally calculate the probability distribution of hydraulic oil filtration strategy. Step S5 includes: S51: Calculate the global statistical value of pollution based on the pollution heat map. The calculation method is as follows: in, This is a global statistical value for pollution levels. The height of the pollution heatmap, The width of the pollution heatmap, For height indexing, For width index, For indicator functions, This represents the pixel value with height i and width j in the pollution intensity heatmap. For the pollution level adaptive threshold, It is a multilayer perceptron; S52: Based on the pollution heatmap, the image is evenly divided into 4 regions, and the weighted pollution index for each region is calculated as follows: in, Here, k is the weighted pollution index for the k-th region, and k is the region index. For the k-th region, the region weight parameter is used. S53: Calculate the probability distribution of hydraulic oil filtration strategies based on the global pollution statistics and the weighted pollution index of each region. The calculation method is as follows: in, The probability distribution of hydraulic oil filtration strategies. These are the weighted pollution indices for the first, second, third, and fourth regions, respectively.

2. The particle identification and filtering method for coal mine hydraulic systems based on deep learning according to claim 1, characterized in that, Step S1 includes: S11: Acquire images of coal mine hydraulic oil using a high-speed microscopic imaging device; S12: Based on the image of hydraulic oil in coal mines, the oil flow interference is separated by the background difference method to obtain the denoised hydraulic oil image; S13: Based on the denoised hydraulic oil image, the pixel distribution is adjusted using an adaptive histogram equalization algorithm to obtain the preprocessed hydraulic oil image.

3. The particle identification and filtering method for coal mine hydraulic systems based on deep learning according to claim 1, characterized in that, Step S2 includes: S21: Based on the preprocessed hydraulic oil image, a multi-scale feature map is generated through parallel three-way convolution. The calculation method is as follows: in, This is the first-scale feature map. This is the second-scale feature map. This is a third-scale feature map. This is a preprocessed image of the hydraulic oil. For convolution, The first-size convolution kernel, For the second-sized convolution kernel, The third-sized convolutional kernel; S22: Based on the multi-scale feature map, a spatial weight matrix and a fused feature map are generated through a gating mechanism, calculated as follows: in, This is the spatial weight matrix. For the softmax function, It is a convolutional layer. For splicing operations, These are the weight matrices for the first, second, and third spaces, respectively. For channel slicing operations, To fuse feature maps, For Hadamah accumulation; S23: Based on the fused feature map, the key regions are enhanced using a residual masking mechanism to obtain the enhanced fused feature map of the key regions. The calculation method is as follows: in, For spatial attention mask, For the sigmoid function, For the feature map gradient, For Sobel horizontal operators, For Sobel vertical operators, For matrix transpose, This is the transpose of the Sobel horizontal operator. Enhanced fusion feature maps for key regions To sharpen the convolution kernel.

4. The particle identification and filtering method for coal mine hydraulic systems based on deep learning according to claim 3, characterized in that, Step S3 includes: S31: Based on the enhanced fused feature map of the key region, a preliminary standardized feature map is obtained through local response normalization layer calculation. The calculation method is as follows: in, For the initial standardization of feature maps, This is a local response normalization algorithm; S32: Based on the preliminary standardized feature map, calculate the adaptive correction factor. The calculation method is as follows: in, For adaptive correction factor, For global average pooling, For adaptive correction factor weight matrix; S33: Based on the preliminary standardized feature map and the adaptive correction factor, the final standardized feature map is calculated as follows: in, This is the final standardized feature map.

5. The particle identification and filtering method for coal mine hydraulic systems based on deep learning according to claim 1, characterized in that, Step S4 includes: S41: Based on the final standardized feature map, calculate the spatial-channel joint gate. The calculation method is as follows: in, For space gate, This is a convolutional layer with a kernel size of 3×3. For passageway doors, For a convolutional layer with a kernel size of 1×1, It is a combined space-passage door; S42: Calculate the refined features based on the final standardized feature map and the space-channel joint gate. The calculation method is as follows: in, Let t be the hidden state at step t, where t is the time step index. It is the hyperbolic tangent function. The hidden state weight matrix is... This is the hidden state at step t-1. This is the final standardized feature map at step t. For refined features, To refine the feature weight matrix, This represents the total number of recursive steps. S43: Calculate the pollution degree heat map based on the refined characteristics. The calculation method is as follows: in, This is a primary decoding feature map. To enhance the decoded feature map, This is a heat map of pollution levels.

6. A deep learning-based particle identification and filtering system for coal mine hydraulic systems, to implement the deep learning-based particle identification and filtering method for coal mine hydraulic systems as described in any one of claims 1-5, characterized in that, include: Coal Mine Hydraulic Oil Image Acquisition and Preprocessing Module: Acquires images of coal mine hydraulic oil and performs background subtraction and histogram equalization to obtain preprocessed hydraulic oil images; Key region enhancement module: Based on the preprocessed hydraulic oil image, calculate multi-scale feature maps, generate spatial weight matrix and fusion feature map, and finally enhance key regions to obtain fusion feature map of key region enhancement; Standardized feature map calculation module: Performs local response normalization calculation on the enhanced fusion feature map of key regions to obtain a preliminary standardized feature map, then calculates the adaptive correction factor, and finally calculates the final standardized feature map; Pollution heat map calculation module: Based on the final standardized feature map, calculate the spatial-channel joint gate and refined features, and then calculate the pollution heat map based on the refined features; Hydraulic oil filtration strategy calculation module: Based on the contamination heat map, calculate the global statistical value of contamination, then divide the image evenly into 4 regions and calculate the weighted contamination index of each region, and finally calculate the probability distribution of hydraulic oil filtration strategy.

Images