Field medical sharp instrument recovery management method and system based on module anti-permeation structure

Abstract

The invention provides a field medical sharp instrument recovery management method and system based on a module anti-permeation structure, and relates to the technical field of image analysis, and the method comprises the steps: recognizing a field medical sharp instrument through multi-modal image fusion and a deep learning model, extracting key features, and carrying out the precise positioning, segmentation and semantic analysis, classifying by combining a causal reasoning network to obtain category and danger level information, and generating and optimizing a recovery scheduling strategy by utilizing an improved space-time diagram neural network and a hierarchical reinforcement learning algorithm and combining an improved auction mechanism and a security constraint model based on the classification information and the danger level information. An optimal recovery route is planned through an improved ant colony algorithm and a heuristic search strategy in combination with a Pareto optimization algorithm.

Description

Technical Field

[0001] The present invention relates to the field of image analysis technology, and in particular to a field medical sharps recovery management method and system based on a module anti-penetration structure. Background Art

[0002] The safe recycling and management of medical sharps in the wild is an important part of the medical and health field. If syringes, needles, surgical blades and other medical sharps are not handled properly, they will pose a serious threat to the environment and public health, and may cause accidental injuries and disease transmission. Therefore, how to effectively identify, classify and recycle medical sharps in the wild is crucial to ensuring public health safety. Existing medical sharps recycling management methods mainly rely on manual identification and classification, which is inefficient and prone to errors. Traditional image recognition technology has problems with insufficient recognition accuracy and robustness when processing medical sharps images in complex field environments. At the same time, existing recycling path planning methods lack consideration of the hazard level and real-time status of medical sharps, making it difficult to achieve efficient and safe recycling and transportation. Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for managing the recovery of sharp medical instruments in the field based on a modular anti-penetration structure, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of an embodiment of the present invention provides a method for managing field medical sharps recovery based on a modular anti-permeability structure, comprising: Construct a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features using an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain category information and hazard level information corresponding to the medical sharps; A classification scheduling system is constructed based on the category information and the hazard level information, a dynamic state representation is modeled through an improved spatiotemporal graph neural network and state representations at different spatiotemporal scales are extracted, the state representation is added to a time series prediction model and the state evolution trend is determined, an initial scheduling strategy is generated in combination with a hierarchical reinforcement learning algorithm, the initial scheduling strategy is optimized by an improved policy gradient algorithm in a pre-set two-stage optimization network, a dynamic allocation is performed in combination with an improved auction mechanism, a dynamic allocation result is obtained, and a safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeability structure, the dynamic allocation result is adjusted by an improved optimization algorithm and the adjusted dynamic allocation result is added to a path planning module, a transportation path set is generated in combination with an improved ant colony algorithm and a heuristic search strategy, and the optimal recovery route is obtained by solving the Pareto optimization algorithm.

[0005] In an optional embodiment, A medical sharps feature recognition model is constructed, and visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps are collected and combined into medical sharps images. A multimodal image fusion algorithm based on deep hashing is executed on the medical sharps image to obtain a fused image. The shape, material, texture and internal structure corresponding to the field medical sharps are extracted based on the fused image and combined to generate medical sharps features. A multi-scale feature pyramid is constructed through a hierarchical attention mechanism, and the medical sharps features are adaptively recalibrated by combining the combination mechanism of channel attention and spatial attention. The recalibrated medical sharps features are enhanced through a deformable convolutional network and a residual connection structure. The enhanced medical sharps features are decoupled by an improved adversarial training method to obtain independent features. The importance of the independent features is evaluated and a feature selection strategy is established to obtain the key features of medical sharps, including: Collect visible light images, infrared thermal imaging images and ultrasonic scanning images of medical sharp objects, and construct feature extraction branches for the visible light images, infrared thermal imaging images and ultrasonic scanning images respectively, wherein the feature extraction branches include multiple convolutional layers, each of which is connected to a batch normalization layer and a rectified linear unit activation function, and the extracted features are mapped into feature vectors through a fully connected layer; Constructing a hash coding layer, mapping the feature vector into a binary hash code through a fully connected layer and a hyperbolic tangent activation function, calculating the Hamming distance between the binary hash codes corresponding to the visible light image, the infrared thermal imaging image, and the ultrasonic scanning image, screening feature-similar images based on a preset distance threshold, using a softmax function to calculate the fusion weight coefficients of the feature-similar images, and linearly combining the weighted features to obtain fused features; Construct a multi-layer feature pyramid network, adjust the number of channels of the fused features through convolution, use the maximum pooling operation to downsample layer by layer to obtain feature maps of different scales, calculate the channel attention weight and the spatial attention weight for the feature map, the channel attention weight is calculated by global average pooling and multi-layer full connection, the spatial attention weight is calculated by convolution operation and sigmoid activation function, and the channel attention weight and the spatial attention weight are multiplied to obtain a comprehensive attention map; Construct an offset prediction branch including multiple convolution layers, predict an offset value for each position of the fused feature, limit the offset value to a preset range through a hyperbolic tangent function, adjust the sampling position of the standard convolution according to the offset value, calculate the feature value at the sampling point by bilinear interpolation and perform convolution operation to obtain enhanced features, and perform residual connection between the enhanced features and the input features to obtain optimized features; Constructing a generator network and a discriminator network, wherein the generator network adopts an encoder-decoder structure, the encoder comprises a plurality of downsampling blocks, each of which comprises convolution, instance normalization, and rectified linear unit activation, the decoder comprises a plurality of upsampling blocks, each of which comprises transposed convolution, instance normalization, and rectified linear unit activation, the discriminator network comprises a plurality of convolutional layers and adopts a rectified linear unit activation function with leakage, and the optimized features are adversarially trained through the generator network and the discriminator network to obtain independent features; Calculate the mutual information value between the independent feature and the medical sharps category label, normalize the independent feature, count the probability distribution of the independent feature and the category label, calculate the mutual information value based on the probability distribution, and select the feature with the highest mutual information value as the key feature.

[0006] In an optional embodiment, The Hamming distance between binary hash codes is calculated as follows: in, D ( H 1, H 2) Represents the first hash code H 1 and the second hash code H The Hamming distance between 2, nrepresents the number of sub-blocks the hash code is divided into, w (i) Indicates i The local sensitivity weight of each sub-block, m Indicates the number of bits of the hash code in each sub-block, C [ i , j ] indicates the i The hash code position and j The correlation between the hash code positions, Represents the first hash code H 1 in i The first j The value of a binary bit, Represents the second hash code H 2 in i The first j In an optional implementation, Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain the category information and hazard level information corresponding to the medical sharps, including: Constructing an improved dense prediction network to locate the medical sharp instrument based on the key features, the dense prediction network includes a main branch and an auxiliary branch, the main branch includes a plurality of densely connected blocks, each densely connected block includes a plurality of convolutional units, each convolutional unit consists of a batch normalization layer, a rectified linear unit activation function and a convolutional layer, the auxiliary branch includes a spatial pyramid pooling module and a feature pyramid attention module, and the position coordinates and confidence scores of the medical sharp instrument are output through the dense prediction network to generate positioning features; Determine a region of interest based on the position coordinates of the positioning feature, and construct a hierarchical attention-guided segmentation network in the region of interest, wherein the segmentation network includes an encoder and a decoder, wherein the encoder extracts multi-level features using an improved residual network structure according to the confidence score of the positioning feature, and the decoder performs feature fusion through jump connections according to the multi-level features, and introduces a channel attention module and a spatial attention module in the feature fusion process to enhance the features, wherein the weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence of the positioning feature, and an edge perception module is introduced to optimize the edge details of the enhanced features to generate segmentation features; Input the segmentation features into a graph convolutional network to analyze the component composition relationship of the medical sharps, and construct a graph structure based on the edge information of the segmentation features, in which nodes represent different component regions, and edges represent the connection relationship between adjacent components. The graph convolutional network includes multiple graph convolutional layers, and the feature conversion parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation; According to the spatial distribution of the positioning feature, the regional segmentation result of the segmentation feature and the component relationship of the semantic feature, an attention-guided adaptive fusion strategy is used to calculate the feature weight, and the weighted feature is input into the causal reasoning network. The temporal flow of the causal reasoning network models the temporal dependency relationship based on the temporal evolution law of the feature, and the spatial flow models the spatial dependency relationship based on the regional distribution characteristics of the feature. A probabilistic graph model is constructed based on the dependency relationship output by the causal reasoning network, and the dependency relationship is mapped into a node connection relationship of a directed acyclic graph. The graph structure is optimized and the conditional probability distribution is determined in combination with prior knowledge. Based on the positioning features, the segmentation features and the semantic features, the overall contour features, the surface color distribution features and the texture features of the medical sharp instrument are respectively extracted, and the global features corresponding to the medical sharp instrument are obtained by combining them, and the edge detail features, the hand-held part features and the functional features of the medical sharp instrument are extracted to obtain the local features corresponding to the medical sharp instrument; A multi-level classifier is constructed according to the output of the probabilistic graphical model, including a coarse classifier for category division and a fine classifier for fine classification, wherein the coarse classifier adopts a random forest structure to classify the global features, and the fine classifier adopts a gradient boosting tree structure to finely classify the local features. The classification results are fused by voting, and the multi-level classifier is used as a base classifier. An integrated learning strategy is used to construct a classifier set, and model parameters are optimized through cross-validation to output category information and hazard level information of medical sharps.

[0007] In an optional embodiment, The coarse classifier uses a random forest structure to classify the global features, the fine classifier uses a gradient boosting tree structure to finely classify the local features, and the voting method is used to fuse the classification results, including: Acquire the global features corresponding to the medical sharps, perform feature normalization processing, construct an improved random forest structure including multi-level decision units as a coarse classifier, set a feature selection module and a classification module in the improved random forest structure, perform adaptive feature sampling based on feature importance scores through the feature selection module, use the classification module to construct a decision tree based on information gain and impose regularization constraints, and output the basic category of the medical sharps; Acquire local features corresponding to the medical sharps, establish a cascaded gradient boosting tree structure as a detailed classifier, divide the cascaded gradient boosting tree structure into a feature extraction layer and a classification layer, use the feature extraction layer to perform attention weighting on the local features, perform residual learning through the classification layer and iteratively optimize the tree structure parameters, implement sample balancing and feature interaction operations, and output a detailed category of the medical sharps; An adaptive multimodal fusion mechanism is established, and a dynamic weight calculation module and a decision fusion module are configured. The dynamic weight calculation module is used to adjust the weight coefficient according to the classification effect. Multi-level voting is performed through the decision fusion module, and expert knowledge base arbitration is initiated when the confidence levels are inconsistent. The final classification results are output in combination with usage scenarios and security risk assessments.

[0008] In an optional embodiment, A classification scheduling system is constructed based on the category information and the hazard level information. The dynamic state representation is modeled by an improved spatiotemporal graph neural network and the state representation at different spatiotemporal scales is extracted. The state representation is added to the time series prediction model and the state evolution trend is determined. The initial scheduling strategy is generated by combining the hierarchical reinforcement learning algorithm. The initial scheduling strategy is optimized by the improved policy gradient algorithm in the pre-set two-stage optimization network. The improved auction mechanism is combined for dynamic allocation to obtain the dynamic allocation result and a safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeability structure. The dynamic allocation result is adjusted by the improved optimization algorithm and the adjusted dynamic allocation result is added to the path planning module. The transportation path set is generated by combining the improved ant colony algorithm and the heuristic search strategy. The optimal recycling route obtained by solving the Pareto optimization algorithm includes: Constructing a spatiotemporal relationship graph of medical sharps, taking the medical sharps as nodes in the spatiotemporal relationship graph, wherein the nodes contain location information, category information, and hazard level information of the medical sharps, analyzing the spatial distribution characteristics of the medical sharps in different regions by setting multiple spatial convolution layers, analyzing the temporal distribution law of the medical sharps by setting temporal convolution layers, and inputting the spatial distribution characteristics and the temporal distribution law into a feature fusion module to generate a state representation of the medical sharps; The state representation is input into a time series prediction model, the time series prediction model includes a time series encoder and a time series decoder, the time series encoder extracts features from historical distribution data and historical recovery data of medical sharps according to the time sequence to obtain periodic features and burst features, the time series decoder predicts the distribution trend of medical sharps based on the periodic features and the burst features, and performs weighted processing on the historical distribution data and the historical recovery data through an attention mechanism to generate a state evolution trend; Based on the state evolution trend, a hierarchical reinforcement learning network is constructed, wherein the hierarchical reinforcement learning network includes a high-level strategy network and a low-level execution network. The high-level strategy network prioritizes the recovery tasks of medical sharps according to the region and type according to the state evolution trend to obtain a subtask set. The low-level execution network determines the recovery time and recovery route based on the subtask set to obtain an initial scheduling strategy, and the initial scheduling strategy is stored in an experience replay pool and the scheduling experience is selected through priority sampling. The initial scheduling strategy is input into a two-stage optimization network, the first stage of the two-stage optimization network evaluates the value of the scheduling action through a policy gradient algorithm and introduces an entropy regularization term to adjust the policy parameters, and the second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates a dynamic allocation result according to the bidding results and constraints of the recycling point; A safety constraint model is constructed based on the dynamic allocation result. The safety constraint model selects an impermeable container according to the physical characteristics and hazard level of the medical sharps, monitors the loading status of the impermeable container in real time, and automatically adjusts the transportation plan when a safety hazard is detected. The transportation plan is input into a path planning module. The path planning module uses an ant colony algorithm to construct a path map between recycling points, performs path search based on transportation distance, time window and safety factors, optimizes the path sequence through pheromone update and local search, and generates an optimal path set that meets multi-objective constraints.

[0009] In an optional embodiment, The second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates dynamic allocation results based on the bidding results and constraints of the recycling points, including: Construct a bidding entity, set the medical sharps recycling point as the buyer entity, and set the medical sharps to be recycled as the seller entity. The medical sharps recycling point has a preset bidding budget and processing capacity constraint. The processing capacity constraint includes the number of medical sharps that can be processed per unit time and the processing capacity limit of different types of medical sharps. The bidding floor price is set based on the category information, danger level and urgency of the medical sharps to be recycled; The bidding rules are constructed by adopting a multi-round incremental bidding method. The medical sharps recycling point submits a bid that meets the maximum bid limit, budget limit and processing capacity requirements in each round of bidding. When the medical sharps recycling point submits a new bid, the price update mechanism is activated. The remaining medical sharps recycling points follow up with bids within the specified time, and the bidding process is promoted based on the minimum price increase limit. A bidding strategy optimization mechanism is built based on historical bidding data and current market conditions. The optimal bidding range is predicted by analyzing historical transaction prices, competitor bidding patterns, and market supply and demand relationships. Dynamic allocation constraints including time window constraints, path constraints, and capacity constraints are set. The time window constraints are used to ensure the time limit for completing the recycling of medical sharps, the path constraints are used to ensure the feasibility of the transportation routes, and the capacity constraints are used to limit the maximum processing capacity of the medical sharps recycling point. Construct a multi-objective evaluation system including the dimensions of allocation efficiency, cost control and risk management. The allocation efficiency dimension includes the timeliness index and balance index of medical sharps recovery, the cost control dimension includes the transportation cost index and the processing cost index, and the risk management dimension includes the safety hazard index and the emergency response capability index. A preliminary allocation plan is generated based on the bidding results and the dynamic allocation constraints, and the feasibility of the preliminary allocation plan is verified. When a constraint conflict occurs, a dynamic allocation result that meets actual needs is generated by adjusting the allocation quantity or replacing the medical sharps recycling point.

[0010] A second aspect of an embodiment of the present invention provides a field medical sharps recovery management system based on a modular anti-penetration structure, comprising: The first unit is used to build a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, build a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features by an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; The second unit is used to accurately locate the medical sharps based on the key features of the medical sharps through an improved dense prediction network to obtain positioning features, use a hierarchical attention-guided segmentation algorithm to construct a refined mask for the medical sharps to obtain segmentation features, use a graph convolutional network to analyze the component composition relationship of the medical sharps to obtain semantic features, fuse the positioning features, segmentation features and semantic features and input them into a causal reasoning network, establish a dependency relationship between features through an improved probabilistic graph model, build a multi-level classifier based on the dependency relationship, and classify the medical sharps in combination with an integrated learning strategy to obtain category information and hazard level information corresponding to the medical sharps; The third unit is used to build a classified scheduling system based on the category information and the hazard level information, model dynamic state representation through an improved spatiotemporal graph neural network and extract state representations at different spatiotemporal scales, add the state representation to a time series prediction model and determine the state evolution trend, generate an initial scheduling strategy in combination with a hierarchical reinforcement learning algorithm, optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network, dynamically allocate in combination with an improved auction mechanism, obtain dynamic allocation results and build a safety constraint model based on the geometric parameters and material parameters of the anti-permeability structure, adjust the dynamic allocation results through an improved optimization algorithm and add the adjusted dynamic allocation results to a path planning module, generate a transportation path set in combination with an improved ant colony algorithm and a heuristic search strategy, and obtain the optimal recovery route through a Pareto optimization algorithm.

[0011] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0012] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0013] In the present invention, through multimodal image fusion, hierarchical attention mechanism, deformable convolutional network and adversarial training and other technical means, the key features of medical sharps are extracted, the recognition accuracy of medical sharps in complex field environments is effectively improved, the category and danger level of medical sharps can be accurately identified, a classification scheduling system is constructed, and the improved spatiotemporal graph neural network, time series prediction model and hierarchical reinforcement learning algorithm are used in combination with the improved auction mechanism and optimization algorithm to realize the intelligent scheduling and dynamic allocation of medical sharps recycling, which can effectively improve the recycling efficiency and reduce the recycling cost. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-penetration structure, and the improved ant colony algorithm and heuristic search strategy are combined for path planning, which ensures the safety of the medical sharps recycling process and prevents secondary pollution and accidental injuries. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of a method for field medical sharps recovery management based on a module anti-penetration structure according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a field medical sharps recovery management system based on a modular anti-penetration structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0017] Figure 1 FIG. 1 is a flow chart of a method for managing field medical sharps recovery based on a modular anti-permeability structure according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Construct a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features by an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; The multimodal image fusion algorithm based on deep hashing is an image fusion method that combines deep learning and hashing technology. It maps images of different modalities to a shared hash coding space, thereby achieving efficient fusion and retrieval of image information. The adaptive recalibration refers to dynamically adjusting the feature space mapping of the image in multimodal image fusion to adapt to the characteristics of different image data. The deformable convolutional network is a method for extending the traditional convolutional neural network. By introducing the deformable convolution operation, it can dynamically adjust the shape and sampling position of the convolution kernel to better adapt to the shape and changes of the target object. It is particularly suitable for processing image tasks with large geometric deformations, such as target detection and image segmentation. The improved adversarial training method optimizes the training process based on the generative adversarial network, usually by adjusting the training strategies of the generator and the discriminator to enhance the stability and effect of the training.

[0018] In an optional embodiment, A medical sharps feature recognition model is constructed, and visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps are collected and combined into medical sharps images. A multimodal image fusion algorithm based on deep hashing is executed on the medical sharps image to obtain a fused image. The shape, material, texture and internal structure corresponding to the field medical sharps are extracted based on the fused image and combined to generate medical sharps features. A multi-scale feature pyramid is constructed through a hierarchical attention mechanism, and the medical sharps features are adaptively recalibrated by combining the combination mechanism of channel attention and spatial attention. The recalibrated medical sharps features are enhanced through a deformable convolutional network and a residual connection structure. The enhanced medical sharps features are decoupled by an improved adversarial training method to obtain independent features. The importance of the independent features is evaluated and a feature selection strategy is established to obtain the key features of medical sharps, including: Collect visible light images, infrared thermal imaging images and ultrasonic scanning images of medical sharp objects, and construct feature extraction branches for the visible light images, infrared thermal imaging images and ultrasonic scanning images respectively, wherein the feature extraction branches include multiple convolutional layers, each of which is connected to a batch normalization layer and a rectified linear unit activation function, and the extracted features are mapped into feature vectors through a fully connected layer; Constructing a hash coding layer, mapping the feature vector into a binary hash code through a fully connected layer and a hyperbolic tangent activation function, calculating the Hamming distance between the binary hash codes corresponding to the visible light image, the infrared thermal imaging image, and the ultrasonic scanning image, screening feature-similar images based on a preset distance threshold, using a softmax function to calculate the fusion weight coefficients of the feature-similar images, and linearly combining the weighted features to obtain fused features; Construct a multi-layer feature pyramid network, adjust the number of channels of the fused features through convolution, use the maximum pooling operation to downsample layer by layer to obtain feature maps of different scales, calculate the channel attention weight and the spatial attention weight for the feature map, the channel attention weight is calculated by global average pooling and multi-layer full connection, the spatial attention weight is calculated by convolution operation and sigmoid activation function, and the channel attention weight and the spatial attention weight are multiplied to obtain a comprehensive attention map; Construct an offset prediction branch including multiple convolution layers, predict an offset value for each position of the fused feature, limit the offset value to a preset range through a hyperbolic tangent function, adjust the sampling position of the standard convolution according to the offset value, calculate the feature value at the sampling point by bilinear interpolation and perform convolution operation to obtain enhanced features, and perform residual connection between the enhanced features and the input features to obtain optimized features; Constructing a generator network and a discriminator network, wherein the generator network adopts an encoder-decoder structure, the encoder comprises a plurality of downsampling blocks, each of which comprises convolution, instance normalization, and rectified linear unit activation, the decoder comprises a plurality of upsampling blocks, each of which comprises transposed convolution, instance normalization, and rectified linear unit activation, the discriminator network comprises a plurality of convolutional layers and adopts a rectified linear unit activation function with leakage, and the optimized features are adversarially trained through the generator network and the discriminator network to obtain independent features; Calculate the mutual information value between the independent feature and the medical sharps category label, normalize the independent feature, count the probability distribution of the independent feature and the category label, calculate the mutual information value based on the probability distribution, and select the feature with the highest mutual information value as the key feature.

[0019] The binary hash code is a special encoding method in the hash algorithm, which performs fast similarity search and retrieval by mapping data to binary codes. The Hamming distance is an indicator for measuring the similarity between two binary strings, indicating the number of bits in the corresponding positions of the two binary strings. The hyperbolic tangent function is a commonly used activation function, and its output value is between -1 and 1. By compressing the output range, it helps the neural network to perform nonlinear transformation during the training process and effectively solves the gradient disappearance problem. The rectified linear unit activation function with leakage is an improved ReLU activation function, which allows negative values ​​to pass at a certain small slope, thereby avoiding the "dead neuron" problem that may occur in ReLU during training and enhancing the learning ability and stability of the model. The mutual information value is an indicator for measuring the degree of mutual dependence between two variables, reflecting the contribution of the information amount of one variable to the other variable.

[0020] Collect visible light images, infrared thermal imaging images, and ultrasonic scanning images of the target medical sharps. For example, use a digital camera to obtain visible light images, use an infrared thermal imager to obtain infrared thermal imaging images, and use a portable ultrasonic scanner to obtain ultrasonic scanning images. Each image reflects the different characteristics of the medical sharps, providing rich original information for subsequent feature extraction and fusion.

[0021] Feature extraction branches are constructed for the three types of images collected. Each branch contains multiple convolutional layers, and each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function. The convolutional layer is used to extract local features of the image. The batch normalization layer can accelerate model training and improve the generalization ability of the model. The rectified linear unit activation function introduces nonlinearity and enhances the expression ability of the model. Taking the visible light image as an example, assuming that the input image size is 256x256 pixels, the first convolutional layer uses 32 convolution kernels of size 3x3, the stride is 1, and the padding is 1, then the output feature map size is still 256x256, and the number of channels becomes 32. Subsequent convolutional layers can adjust parameters such as the number, size, stride, and padding of convolution kernels as needed. Finally, the extracted features are mapped to feature vectors through a fully connected layer, for example, a 256x256x32 feature map is converted to a 1024-dimensional feature vector.

[0022] The feature vectors of the three images are hashed and fused. The feature vectors are mapped to binary hash codes, such as 128-bit hash codes, through a fully connected layer and a hyperbolic tangent activation function. The Hamming distance between the binary hash codes corresponding to the three images is calculated. The smaller the Hamming distance, the more similar the image features are. A distance threshold is set, such as 32, to filter out images with similar features. Then, the softmax function is used to calculate the fusion weight coefficients of images with similar features. For example, assuming that the Hamming distance between the visible light image and the infrared thermal imaging image is 20, and the Hamming distance between the visible light image and the ultrasonic scanning image is 40, the visible light image has a higher similarity with the infrared thermal imaging image, and its corresponding weight coefficient is also larger. Finally, the weighted features are linearly combined to obtain the fused features. The fused features integrate the information of the three images and more comprehensively describe the characteristics of medical sharps.

[0023] After obtaining the fused features, a multi-layer feature pyramid network is constructed. The number of channels of the fused features is adjusted through convolution, and then the maximum pooling operation is used to downsample layer by layer to obtain feature maps of different scales. For example, the initial fused feature size is 64x64, and feature maps of different scales such as 32x32 and 16x16 can be obtained through the maximum pooling operation. Feature maps of different scales can capture feature information of different sizes. The channel attention weight and spatial attention weight are calculated for the feature map of each scale. The channel attention weight is calculated by global average pooling and multi-layer full connection, and the spatial attention weight is calculated by convolution operation and sigmoid activation function. The comprehensive attention map is obtained by multiplying the channel attention weight and the spatial attention weight, which is used to adaptively recalibrate the features and highlight important feature information.

[0024] Enhance the recalibrated features. Construct an offset prediction branch containing multiple convolutional layers to predict the offset value for each position of the fused feature. The offset value is limited to a preset range, such as [-1, 1], by the hyperbolic tangent function. Adjust the sampling position of the standard convolution according to the offset value, use bilinear interpolation to calculate the feature value at the sampling point and perform convolution operation to obtain enhanced features. Perform residual connection between the enhanced features and the input features to obtain optimized features, further improving the expressiveness of the features.

[0025] Decouple the optimized features and select features. Construct the generator network and the discriminator network, decouple the optimized features through adversarial training, and obtain independent features. Calculate the mutual information value between the independent features and the medical sharps category label. Normalize the independent features, count the probability distribution of the independent features and the category label, and calculate the mutual information value based on the probability distribution. Select the features with higher mutual information values ​​as key features for the final medical sharps identification.

[0026] In this embodiment, through multimodal image fusion and deep learning technology, more comprehensive and discriminative features can be extracted, thereby improving the accuracy of medical sharps identification. It has strong robustness to environmental factors such as lighting, angle, occlusion, etc., and can adapt to complex field environments. Through feature selection strategies, redundant features can be removed, computational complexity can be reduced, and recognition efficiency can be improved.

[0027] In an optional embodiment, The Hamming distance between binary hash codes is calculated as follows: in, D ( H 1, H 2) Represents the first hash code H 1 and the second hash code H The Hamming distance between 2, n Indicates the number of sub-blocks the hash code is divided into, w (i) Indicates i The local sensitivity weight of each sub-block, m Represents the number of bits of the hash code in each sub-block, C [ i , j ] indicates the i The hash code position and j The correlation between the hash code positions, Represents the first hash code H 1 in i The first j The value of a binary bit, Represents the second hash code H 2 in i The first j The value of a binary digit.

[0028] In this embodiment, by using the weighted Hamming distance, similar and dissimilar data can be distinguished more effectively, thereby improving retrieval efficiency and finding target data more quickly. Locally sensitive hashing and weighting mechanisms can reduce the impact of noise and errors and improve the robustness of retrieval. Even when there is a certain amount of noise or error in the data, better retrieval results can be obtained. By assigning different weights to different sub-blocks, the similarity between data can be more accurately reflected, thereby improving retrieval accuracy and finding more relevant results.

[0029] S2. Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain the category information and hazard level information corresponding to the medical sharps; The improved dense prediction network is a deep learning model for image segmentation and target detection tasks. It improves the performance of traditional dense prediction networks by enhancing the network's feature extraction capabilities and fine-grained prediction. The hierarchical attention-guided segmentation algorithm is an image segmentation method based on an attention mechanism. By applying hierarchical attention guidance to feature maps at different levels, the model can better focus on important areas in the image, thereby improving segmentation accuracy. The causal reasoning network is a deep learning model based on causal reasoning, which is used to infer the interaction between variables in complex causal relationships. The improved probabilistic graphical model is an enhanced graphical model that combines the ideas of probability theory and graph theory, and is usually used to describe the dependency relationship between random variables.

[0030] In an optional embodiment, Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain the category information and hazard level information corresponding to the medical sharps, including: Constructing an improved dense prediction network to locate the medical sharp instrument based on the key features, the dense prediction network includes a main branch and an auxiliary branch, the main branch includes a plurality of densely connected blocks, each densely connected block includes a plurality of convolutional units, each convolutional unit consists of a batch normalization layer, a rectified linear unit activation function and a convolutional layer, the auxiliary branch includes a spatial pyramid pooling module and a feature pyramid attention module, and the position coordinates and confidence scores of the medical sharp instrument are output through the dense prediction network to generate positioning features; Determine a region of interest based on the position coordinates of the positioning feature, and construct a hierarchical attention-guided segmentation network in the region of interest, wherein the segmentation network includes an encoder and a decoder, wherein the encoder extracts multi-level features using an improved residual network structure according to the confidence score of the positioning feature, and the decoder performs feature fusion through jump connections according to the multi-level features, and introduces a channel attention module and a spatial attention module in the feature fusion process to enhance the features, wherein the weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence of the positioning feature, and an edge perception module is introduced to optimize the edge details of the enhanced features to generate segmentation features; Input the segmentation features into a graph convolutional network to analyze the component composition relationship of the medical sharps, and construct a graph structure based on the edge information of the segmentation features, in which nodes represent different component regions, and edges represent the connection relationship between adjacent components. The graph convolutional network includes multiple graph convolutional layers, and the feature conversion parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation; According to the spatial distribution of the positioning feature, the regional segmentation result of the segmentation feature and the component relationship of the semantic feature, an attention-guided adaptive fusion strategy is used to calculate the feature weight, and the weighted feature is input into the causal reasoning network. The temporal flow of the causal reasoning network models the temporal dependency relationship based on the temporal evolution law of the feature, and the spatial flow models the spatial dependency relationship based on the regional distribution characteristics of the feature. A probabilistic graph model is constructed based on the dependency relationship output by the causal reasoning network, and the dependency relationship is mapped into a node connection relationship of a directed acyclic graph. The graph structure is optimized and the conditional probability distribution is determined in combination with prior knowledge. Based on the positioning features, the segmentation features and the semantic features, the overall contour features, the surface color distribution features and the texture features of the medical sharp instrument are respectively extracted, and the global features corresponding to the medical sharp instrument are obtained by combining them, and the edge detail features, the hand-held part features and the functional features of the medical sharp instrument are extracted to obtain the local features corresponding to the medical sharp instrument; A multi-level classifier is constructed according to the output of the probabilistic graphical model, including a coarse classifier for category division and a fine classifier for fine classification, wherein the coarse classifier adopts a random forest structure to classify the global features, and the fine classifier adopts a gradient boosting tree structure to finely classify the local features. The classification results are fused by voting, and the multi-level classifier is used as a base classifier. An integrated learning strategy is used to construct a classifier set, and model parameters are optimized through cross-validation to output category information and hazard level information of medical sharps.

[0031] The edge perception module is a component in a deep neural network, which is specifically used to improve the sensitivity of the model when processing image edge information. The attention-guided adaptive fusion strategy is a strategy for weighted fusion of features of different modalities or different levels through an attention mechanism. The time series flow refers to the propagation or change process of data in a time series. The gradient boosting tree structure is an integrated learning method that gradually improves the prediction performance of the model through the combination of multiple decision trees.

[0032] An improved dense prediction network is constructed to accurately locate medical sharps, including a main branch and an auxiliary branch. The main branch consists of multiple densely connected blocks, each of which contains several convolutional units, and each convolutional unit contains a batch normalization layer, a ReLU activation function layer, and a convolutional layer in sequence. The auxiliary branch contains a spatial pyramid pooling module and a feature pyramid attention module. After inputting an image of a medical sharp, the main branch extracts the multi-scale features of the image, while the auxiliary branch captures the global contextual information and local salient features of the image. The output features of the main branch and the auxiliary branch are fused to obtain the position coordinates and confidence scores of the sharp, and generate positioning features. For example, when an image containing a scalpel is input, the network outputs the coordinates of the scalpel tip (x=100, y=200) and a confidence score of 0.95.

[0033] The region of interest is determined based on the position coordinates of the positioning feature, and a hierarchical attention-guided segmentation network is constructed within the region of interest, including an encoder and a decoder. The encoder uses an improved residual network structure to extract multi-level features based on the confidence score of the positioning feature. The decoder fuses the multi-level features extracted by the encoder through jump connections. In the feature fusion process, the channel attention module and the spatial attention module are introduced to enhance the features, and the weights of these two modules are adaptively adjusted according to the confidence of the positioning feature. At the same time, the edge perception module is introduced to optimize the edge details of the enhanced features, and finally generate segmentation features. For example, based on the coordinates and confidence of the scalpel tip, the region of interest containing the entire scalpel is determined, and the segmentation network outputs the pixel-level mask of the scalpel.

[0034] The segmentation features are input into the graph convolutional network to analyze the component composition relationship of medical sharps. A graph structure is constructed based on the edge information of the segmentation features. The nodes in the graph represent different component regions, and the edges represent the connection relationship between adjacent components. The graph convolutional network contains multiple graph convolutional layers. The feature conversion parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation. For example, the segmentation result of a scalpel can construct a graph with the handle and blade as nodes and the connection relationship between them as edges. The graph convolutional network analysis obtains the semantic features of the handle and blade.

[0035] According to the spatial distribution of the positioning features, the regional segmentation results of the segmentation features, and the component relationships of the semantic features, the feature weights are calculated using an attention-guided adaptive fusion strategy, and the weighted features are input into the causal reasoning network. The temporal flow of the causal reasoning network models the temporal dependency based on the temporal evolution law of the features, and the spatial flow models the spatial dependency based on the regional distribution characteristics of the features. A probabilistic graph model is constructed based on the dependency output by the causal reasoning network, and the dependency is mapped into the node connection relationship of a directed acyclic graph. The graph structure is optimized and the conditional probability distribution is determined by combining prior knowledge. For example, the dependency between the tip position, blade shape, and handle material of a scalpel is modeled as a probabilistic graph model.

[0036] Based on positioning features, segmentation features and semantic features, the overall contour features, surface color distribution features and texture features of medical sharps are extracted respectively, and the global features are obtained by combining them. The edge detail features, hand-held part features and functional features of medical sharps are then extracted to obtain local features. For example, the overall shape, color and texture of a scalpel are extracted as global features, and the sharpness of the blade, the way the handle is held and the cutting function of the scalpel are extracted as local features.

[0037] A multi-level classifier is constructed based on the output of the probabilistic graphical model, including a coarse classifier for category division and a fine classifier for fine classification. The coarse classifier uses a random forest structure to classify global features, and the fine classifier uses a gradient boosting tree structure to finely classify local features, and the classification results are fused by voting. The multi-level classifier is used as the base classifier, and an ensemble learning strategy is used to construct a classifier set. The model parameters are optimized through cross-validation, and the category information and danger level information of medical sharps are output. For example, the coarse classifier classifies a scalpel into a knife class, and the fine classifier further determines that it is a scalpel, and evaluates its danger level as high based on the sharpness of the blade.

[0038] In this embodiment, through multi-feature fusion and causal reasoning, the category and danger level of medical sharps can be identified more accurately, the misidentification rate can be reduced, and potential medical accidents can be effectively avoided. It combines technologies such as dense prediction networks, hierarchical attention mechanisms, and graph convolutional networks, and can effectively cope with complex situations such as image noise, lighting changes, and occlusions, improve the robustness of recognition, and can finely segment medical sharps, analyze the composition relationship of their components, and extract global and local features, thereby achieving more refined identification and danger level assessment, providing more reliable protection for medical safety.

[0039] In an optional embodiment, The coarse classifier uses a random forest structure to classify the global features, the fine classifier uses a gradient boosting tree structure to finely classify the local features, and the voting method is used to fuse the classification results, including: Acquire the global features corresponding to the medical sharps, perform feature normalization processing, construct an improved random forest structure including multi-level decision units as a coarse classifier, set a feature selection module and a classification module in the improved random forest structure, perform adaptive feature sampling based on feature importance scores through the feature selection module, use the classification module to construct a decision tree based on information gain and apply regularization constraints, and output the basic category of the medical sharps; Acquire local features corresponding to the medical sharps, establish a cascaded gradient boosting tree structure as a detailed classifier, divide the cascaded gradient boosting tree structure into a feature extraction layer and a classification layer, use the feature extraction layer to perform attention weighting on the local features, perform residual learning through the classification layer and iteratively optimize the tree structure parameters, implement sample balancing and feature interaction operations, and output a detailed category of the medical sharps; An adaptive multimodal fusion mechanism is established, and a dynamic weight calculation module and a decision fusion module are configured. The dynamic weight calculation module is used to adjust the weight coefficient according to the classification effect. Multi-level voting is performed through the decision fusion module, and expert knowledge base arbitration is initiated when the confidence levels are inconsistent. The final classification results are output in combination with usage scenarios and security risk assessments.

[0040] The expert knowledge base arbitration is a decision-making mechanism that combines artificial intelligence and expert systems, and is usually used to resolve different opinions or information conflicts from multiple sources. In the system, the expert knowledge base contains the knowledge and rules of experts in the field, and the arbitration process makes decisions based on the knowledge of the expert base to help the system make the final choice or judgment. It is often used for intelligent decision-making in fields such as medicine, law and finance.

[0041] The global features of medical sharps are obtained. The global features describe the overall morphology and structural information of medical sharps, such as the length, width, thickness, volume, surface area, etc. of the sharps. Taking a disposable syringe needle as an example, its global features can be expressed as: length 38mm, width 0.5mm, thickness 0.3mm, volume 0.057 cubic millimeters, and surface area 38.5 square millimeters. The obtained global features are normalized, and each feature value is scaled to the same numerical range, such as between 0 and 1, to eliminate the impact of dimensions between different features.

[0042] An improved random forest structure is constructed as a coarse classifier. The structure contains multi-level decision units, each of which consists of a feature selection module and a classification module. The feature selection module performs adaptive feature sampling based on feature importance scores. For example, in the classification of needles, the importance score of the "length" feature is higher than that of the "surface area" feature, so the feature selection module will select the "length" feature with a higher probability to participate in the construction of the decision tree. The classification module constructs a decision tree based on information gain and imposes regularization constraints, such as limiting the maximum depth of the decision tree or the minimum number of leaf node samples to prevent overfitting. The normalized global features are classified by the improved random forest structure, and the basic categories of medical sharps, such as needles, scalpels, suture needles, etc., are output. Assume that based on the global features, the needle is preliminarily classified into the "syringe needle" category.

[0043] Get the local features of medical sharps. Local features describe the details of specific parts of medical sharps, such as the shape of the needle tip, the sharpness of the blade, the curvature of the suture needle, etc. The local features of the needle can include: needle tip angle, needle tip curvature, needle body texture, etc. Take the needle tip angle as an example, its value is 30 degrees.

[0044] A cascaded gradient boosting tree structure is established as a detailed classifier. The structure is divided into a feature extraction layer and a classification layer. The feature extraction layer performs attention weighting on local features. For example, the needle tip angle has a higher importance in distinguishing different types of syringe needles, so the feature extraction layer gives this feature a higher weight. The classification layer performs residual learning and iteratively optimizes the tree structure parameters, such as the learning rate, the number and depth of trees, etc. At the same time, sample balancing and feature interaction operations are implemented to improve the robustness and generalization ability of the classifier. Assuming that the local features of the needle indicate that its needle tip angle is small and suitable for subcutaneous injection, the detailed classifier further classifies it as a "hypodermic needle".

[0045] An adaptive multimodal fusion mechanism is established, including a dynamic weight calculation module and a decision fusion module. The dynamic weight calculation module adjusts the weight coefficient according to the classification effect of the coarse classifier and the fine classifier. For example, if the accuracy of the fine classifier is higher, it is given a higher weight. The decision fusion module performs multi-level voting, such as voting according to the coarse classification results and the fine classification results respectively, and determines the final classification result according to the voting results. When the confidence is inconsistent, for example, the coarse classification result is "syringe needle" and the fine classification result is "suture needle", the expert knowledge base arbitration is initiated, and the final classification result is output in combination with the usage scenario and safety risk assessment. Assume that the medical sharp is finally determined to be a "hypodermic injection needle".

[0046] In this embodiment, by combining global and local features, as well as coarse classification and fine classification strategies, the characteristic information of medical sharps can be captured more comprehensively, thereby improving the classification accuracy. The improved random forest and cascade gradient boosting tree structures can quickly process a large amount of characteristic data, thereby improving the efficiency of classification. Through accurate classification, the classification and treatment of medical waste can be effectively guided, the risk of medical waste treatment can be reduced, and the safety of the medical environment can be ensured. S3. Construct a classification scheduling system based on the category information and the hazard level information, model the dynamic state representation through the improved spatiotemporal graph neural network and extract the state representation at different spatiotemporal scales, add the state representation to the time series prediction model and determine the state evolution trend, generate an initial scheduling strategy in combination with the hierarchical reinforcement learning algorithm, optimize the initial scheduling strategy through the improved policy gradient algorithm in the pre-set two-stage optimization network, dynamically allocate in combination with the improved auction mechanism, obtain the dynamic allocation result and construct a safety constraint model based on the geometric parameters and material parameters of the anti-permeability structure, adjust the dynamic allocation result through the improved optimization algorithm and add the adjusted dynamic allocation result to the path planning module, generate a transportation path set in combination with the improved ant colony algorithm and the heuristic search strategy, and obtain the optimal recovery route through the Pareto optimization algorithm.

[0047] The improved spatiotemporal graph neural network is an enhanced model that combines graph neural networks with time series modeling, and is used to capture the spatiotemporal dependency between nodes. The state evolution trend refers to the dynamic change law of a system or variable in the time dimension, and is usually used to analyze time series data or predict the change trend of future states. The two-stage optimization network is a deep learning model trained in stages, usually including a coarse-grained optimization stage and a fine-grained optimization stage. The improved auction mechanism is an optimized resource allocation method, which improves the fairness and profit maximization effect of the auction process by introducing more efficient bidding rules or more reasonable allocation strategies, and is widely used in e-commerce and resource scheduling and other fields. The security constraint model is a mathematical model that introduces security constraints into optimization problems, and helps decision makers optimize the objective function while meeting basic security requirements by ensuring that the operation of the system does not violate security conditions. The ant colony algorithm is an optimization algorithm that simulates the foraging behavior of ants. Through the transmission of pheromones between individuals, the global optimal path search is collaboratively completed, and it is often used to solve complex combinatorial optimization problems. The heuristic search strategy is an algorithm strategy that uses heuristic information to guide the search direction. By selecting paths or nodes that are more likely to reach the target, the search efficiency is significantly improved, and it is commonly used in the fields of artificial intelligence and operations optimization.

[0048] In an optional embodiment, A classification scheduling system is constructed based on the category information and the hazard level information. The dynamic state representation is modeled by an improved spatiotemporal graph neural network and the state representation at different spatiotemporal scales is extracted. The state representation is added to the time series prediction model and the state evolution trend is determined. The initial scheduling strategy is generated by combining the hierarchical reinforcement learning algorithm. The initial scheduling strategy is optimized by the improved policy gradient algorithm in the pre-set two-stage optimization network. The improved auction mechanism is combined for dynamic allocation to obtain the dynamic allocation result and a safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeability structure. The dynamic allocation result is adjusted by the improved optimization algorithm and the adjusted dynamic allocation result is added to the path planning module. The transportation path set is generated by combining the improved ant colony algorithm and the heuristic search strategy. The optimal recycling route obtained by solving the Pareto optimization algorithm includes: Constructing a spatiotemporal relationship graph of medical sharps, taking the medical sharps as nodes in the spatiotemporal relationship graph, wherein the nodes contain location information, category information, and hazard level information of the medical sharps, analyzing the spatial distribution characteristics of the medical sharps in different regions by setting multiple spatial convolution layers, analyzing the temporal distribution law of the medical sharps by setting temporal convolution layers, and inputting the spatial distribution characteristics and the temporal distribution law into a feature fusion module to generate a state representation of the medical sharps; The state representation is input into a time series prediction model, the time series prediction model includes a time series encoder and a time series decoder, the time series encoder extracts features from historical distribution data and historical recovery data of medical sharps according to the time sequence to obtain periodic features and burst features, the time series decoder predicts the distribution trend of medical sharps based on the periodic features and the burst features, and performs weighted processing on the historical distribution data and the historical recovery data through an attention mechanism to generate a state evolution trend; Based on the state evolution trend, a hierarchical reinforcement learning network is constructed, wherein the hierarchical reinforcement learning network includes a high-level strategy network and a low-level execution network. The high-level strategy network prioritizes the recovery tasks of medical sharps according to the region and type according to the state evolution trend to obtain a subtask set. The low-level execution network determines the recovery time and recovery route based on the subtask set to obtain an initial scheduling strategy, and the initial scheduling strategy is stored in an experience replay pool and the scheduling experience is selected through priority sampling. The initial scheduling strategy is input into a two-stage optimization network, the first stage of the two-stage optimization network evaluates the value of the scheduling action through a policy gradient algorithm and introduces an entropy regularization term to adjust the policy parameters, and the second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates a dynamic allocation result according to the bidding results and constraints of the recycling point; A safety constraint model is constructed based on the dynamic allocation result. The safety constraint model selects an impermeable container according to the physical characteristics and hazard level of the medical sharps, monitors the loading status of the impermeable container in real time, and automatically adjusts the transportation plan when a safety hazard is detected. The transportation plan is input into a path planning module. The path planning module uses an ant colony algorithm to construct a path map between recycling points, performs path search based on transportation distance, time window and safety factors, optimizes the path sequence through pheromone update and local search, and generates an optimal path set that meets multi-objective constraints.

[0049] The high-level strategy network is a deep learning network used to make global decisions. It provides guiding strategies and target directions for the low-level network by comprehensively analyzing global state information. The low-level execution network is a deep learning network that focuses on the execution of specific tasks. It is responsible for completing fine-grained operations and feedback according to the instructions of the high-level strategy, realizing the separation of strategy and execution. The pheromone is the core parameter for simulating the behavior of ant colonies in the ant colony algorithm, which is used to indicate the quality of the path. The pheromone concentration on the path increases as more ants choose the path, thereby guiding the subsequent search direction.

[0050] Construct a spatiotemporal relationship graph of medical sharps. Each medical sharp is regarded as a node in the graph. Each node contains the geographical location (such as longitude and latitude), category (such as needle, scalpel) and danger level (such as high risk, low risk) of the medical sharp. In order to analyze the spatial distribution characteristics of medical sharps in different regions, multi-layer spatial convolution layers are used to process node information. Each spatial convolution layer considers the information of the node and its neighboring nodes, and aggregates this information through learnable weight parameters to extract spatial features of different scales. At the same time, in order to analyze the temporal distribution law of medical sharps, the temporal convolution layer is used to process the node information. The temporal convolution layer considers the information of the node at different time steps and captures the temporal change trend through learnable weight parameters. Finally, the spatial distribution features extracted by the spatial convolution layer and the temporal distribution law extracted by the temporal convolution layer are input into a feature fusion module, which integrates the two features to generate the state representation of the medical sharps. For example, a node represents a high-risk needle, which is located in area A on October 1. Through the spatiotemporal graph neural network, the spatiotemporal distribution characteristics of the needle among the high-risk sharps in area A can be obtained.

[0051] The state representation of medical sharps is input into the time series prediction model, which consists of a time series encoder and a time series decoder. The time series encoder receives the historical distribution data of medical sharps (such as the number of sharps in each area every day in the past year) and the historical recycling data (such as the number of recycling in each area every day in the past year), and extracts features from these data in chronological order. The encoder learns the periodic features (such as the large number of sharps on Mondays) and burst features (such as the abnormal number of sharps on a certain holiday) in the data. The time series decoder predicts the distribution trend of medical sharps in the future based on the periodic features and burst features extracted by the encoder. In order to better capture the important information in the data, the attention mechanism is used in the model to weight the historical distribution data and historical recycling data, so as to generate a more accurate state evolution trend. For example, based on the data of the past year, the model predicts that the number of high-risk needles in area A will increase by 20% in the next week.

[0052] A hierarchical reinforcement learning network is constructed based on the state evolution trend, which consists of a high-level policy network and a low-level execution network. The high-level policy network prioritizes the recovery task of medical sharps by region and type according to the state evolution trend (for example, the number of high-risk needles in area A will increase by 20% in the next week), and decomposes it into multiple subtasks. For example, "recovering high-risk needles in area A" is a high-priority subtask. The low-level execution network determines the recovery time and recovery route of each subtask based on the subtask set, thereby generating an initial scheduling strategy. For example, the low-level network decides to recover high-risk needles in area A between 9 and 10 a.m. on Tuesday, and plans a specific recovery route. The generated initial scheduling strategy is stored in the experience replay pool, and the scheduling experience is selected through priority sampling for subsequent optimization.

[0053] The initial scheduling strategy is input into the two-stage optimization network. In the first stage, the policy gradient algorithm is used to evaluate the value of the scheduling action, and the entropy regularization term is introduced to adjust the policy parameters to make the strategy more stable and robust. In the second stage, the resources of medical sharps are allocated based on the auction mechanism. Each recycling point submits a bid based on its own situation and the current scheduling strategy. The network generates a dynamic allocation result based on the bidding results and constraints of the recycling point (such as vehicle capacity and time window). For example, recycling point B successfully bid with the lowest price and won the task of recycling high-risk needles in area A.

[0054] A safety constraint model is constructed based on the dynamic allocation results. The model selects appropriate impermeable containers based on the physical characteristics (e.g., size, weight) and hazard level of medical sharps. During transportation, the loading status of the impermeable container is monitored in real time, such as by monitoring the temperature, pressure and other parameters in the container through sensors. When safety hazards are detected (e.g., container damage, high temperature), the transportation plan is automatically adjusted. The adjusted transportation plan is input into the path planning module. This module uses the ant colony algorithm to construct a path map between recycling points and searches for paths based on transportation distance, time window and safety factors. The path sequence is optimized through techniques such as pheromone updating and local search, and finally an optimal path set that meets multi-objective constraints is generated. For example, if the temperature of the container is detected to be too high, the system will automatically adjust the transportation plan, select a shorter and safer path, and reduce the transportation speed.

[0055] In this embodiment, the distribution trend and state evolution of medical sharps can be accurately predicted through the spatiotemporal graph neural network and time series prediction model, so as to plan the recycling tasks in advance, optimize resource allocation, and improve recycling efficiency. The safety constraint model and impermeable containers can effectively reduce the safety risks in the transportation of medical sharps and prevent accidents. Through hierarchical reinforcement learning and auction mechanisms, the recycling routes and resource allocation can be optimized, transportation costs can be reduced, and economic benefits can be improved.

[0056] In an optional embodiment, The second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates dynamic allocation results based on the bidding results and constraints of the recycling points, including: Construct a bidding entity, set the medical sharps recycling point as the buyer entity, and set the medical sharps to be recycled as the seller entity. The medical sharps recycling point has a preset bidding budget and processing capacity constraint. The processing capacity constraint includes the number of medical sharps that can be processed per unit time and the processing capacity limit of different types of medical sharps. The bidding floor price is set based on the category information, danger level and urgency of the medical sharps to be recycled; The bidding rules are constructed by adopting a multi-round incremental bidding method. The medical sharps recycling point submits a bid that meets the maximum bid limit, budget limit and processing capacity requirements in each round of bidding. When the medical sharps recycling point submits a new bid, the price update mechanism is activated. The remaining medical sharps recycling points follow up with bids within the specified time, and the bidding process is promoted based on the minimum price increase limit. A bidding strategy optimization mechanism is built based on historical bidding data and current market conditions. The optimal bidding range is predicted by analyzing historical transaction prices, competitor bidding patterns, and market supply and demand relationships. Dynamic allocation constraints including time window constraints, path constraints, and capacity constraints are set. The time window constraints are used to ensure the time limit for completing the recycling of medical sharps, the path constraints are used to ensure the feasibility of the transportation routes, and the capacity constraints are used to limit the maximum processing capacity of the medical sharps recycling point. Construct a multi-objective evaluation system including the dimensions of allocation efficiency, cost control and risk management. The allocation efficiency dimension includes the timeliness index and balance index of medical sharps recovery, the cost control dimension includes the transportation cost index and the processing cost index, and the risk management dimension includes the safety hazard index and the emergency response capability index. A preliminary allocation plan is generated based on the bidding results and the dynamic allocation constraints, and the feasibility of the preliminary allocation plan is verified. When a constraint conflict occurs, a dynamic allocation result that meets actual needs is generated by adjusting the allocation quantity or replacing the medical sharps recycling point.

[0057] The bidding entity is a participant in the auction mechanism, responsible for bidding for the target object, usually representing different stakeholders. The bidding reserve price is the lowest acceptable price set in the auction mechanism to ensure that the target object will not be sold at an excessively low price. The multi-round incremental bidding method is an auction rule that ensures that the auction process is fair and transparent by gradually increasing the bid in multiple rounds of bidding, while helping to maximize the bidding benefits.

[0058] Construct a bidding entity. Set the medical sharps recycling point as the buyer and the medical sharps to be recycled as the seller. Each recycling point has a preset bidding budget and processing capacity constraints. The processing capacity constraints include the number of medical sharps that can be processed per unit time and the processing capacity limits of different types of medical sharps. For example, the daily processing capacity of recycling point A is 1,000 pieces, of which the processing capacity of needle-type medical sharps is 800 pieces, the processing capacity of blade-type medical sharps is 200 pieces, and the total budget is 5,000 yuan. The daily processing capacity of recycling point B is 1,500 pieces, of which the processing capacity of needle-type medical sharps is 500 pieces, the processing capacity of blade-type medical sharps is 1,000 pieces, and the total budget is 7,000 yuan. Set the bidding floor price according to the category information, hazard level and urgency of the medical sharps to be recycled. For example, the floor price of needle-type medical sharps is 1 yuan / piece, and the floor price of blade-type medical sharps is 2 yuan / piece. If the urgency is high, the floor price will be increased by 20%.

[0059] The bidding rules are constructed by using multiple rounds of incremental bidding. Each recycling point submits a bid that meets the maximum bid limit, budget limit and processing capacity requirements in each round of bidding. For example, recycling point A quotes 1.2 yuan / piece for a batch of 100 needle-type medical sharps, with a total price of 120 yuan, which does not exceed its budget and processing capacity. When the recycling point submits a new bid, the price update mechanism is activated, and the remaining recycling points follow up with bids within a specified time, such as within 5 minutes. The system sets a minimum price increase, such as 0.1 yuan / piece, to advance the bidding process.

[0060] Build a bidding strategy optimization mechanism based on historical bidding data and current market conditions. Predict the optimal quotation range by analyzing historical transaction prices, competitor quotation patterns, and market supply and demand relationships. For example, the system analyzes that the average transaction price of needle-type medical sharps in the past month was 1.1 yuan / piece. The current market demand is high, and it is recommended that the quotation range of recycling point A be 1.15-1.25 yuan / piece. At the same time, set dynamic allocation constraints, including time window constraints, path constraints, and capacity constraints. For example, medical sharps are required to be recycled within 24 hours, transportation routes must avoid sensitive areas such as schools, and the maximum processing volume of each recycling point cannot exceed its daily processing capacity.

[0061] Construct a multi-objective evaluation system, including the dimensions of allocation efficiency, cost control, and risk management. The dimension of allocation efficiency includes indicators of timeliness and balance of medical sharps recycling. For example, all medical sharps are recycled within the specified time, and the processing volume of each recycling point is relatively balanced. The cost control dimension includes transportation cost indicators and processing cost indicators. For example, the shortest transportation route is selected to reduce transportation costs, and the processing cost of each recycling point is controlled within its budget. The risk management dimension includes safety hazard indicators and emergency response capability indicators. For example, necessary safety measures are taken during transportation to ensure the safe transportation of medical sharps, and emergency plans are formulated to deal with emergencies.

[0062] Generate a preliminary allocation plan based on the bidding results and dynamic allocation constraints. Verify the feasibility of the preliminary allocation plan. When there is a constraint conflict, generate a dynamic allocation result that meets actual needs by adjusting the allocation quantity or changing the recycling point. For example, the preliminary plan allocates 500 needle-like medical sharps to recycling point A, but recycling point A only has 400 needle-like medical sharps processing capacity on that day. In this case, the remaining 100 pieces need to be allocated to other recycling points or the allocation quantity of recycling point A needs to be adjusted.

[0063] In this embodiment, the demand for medical sharps recycling and the processing capacity of the recycling points can be quickly matched, the timeliness and balance of medical sharps recycling can be achieved, and the recycling efficiency can be improved. Through the bidding mechanism and dynamic allocation constraints, the transportation cost and processing cost can be effectively controlled to achieve cost optimization. Risk factors such as safety hazards and emergency response capabilities are taken into consideration, and risks can be effectively managed to ensure the safety and reliability of the medical sharps recycling process.

[0064] Figure 2 FIG. 1 is a schematic diagram of a field medical sharps recovery management system based on a modular anti-permeability structure according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to build a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, build a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features by an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; The second unit is used to accurately locate the medical sharps based on the key features of the medical sharps through an improved dense prediction network to obtain positioning features, use a hierarchical attention-guided segmentation algorithm to construct a refined mask for the medical sharps to obtain segmentation features, use a graph convolutional network to analyze the component composition relationship of the medical sharps to obtain semantic features, fuse the positioning features, segmentation features and semantic features and input them into a causal reasoning network, establish a dependency relationship between features through an improved probabilistic graph model, build a multi-level classifier based on the dependency relationship, and classify the medical sharps in combination with an integrated learning strategy to obtain category information and hazard level information corresponding to the medical sharps; The third unit is used to build a classified scheduling system based on the category information and the hazard level information, model dynamic state representation through an improved spatiotemporal graph neural network and extract state representations at different spatiotemporal scales, add the state representation to a time series prediction model and determine the state evolution trend, generate an initial scheduling strategy in combination with a hierarchical reinforcement learning algorithm, optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network, dynamically allocate in combination with an improved auction mechanism, obtain dynamic allocation results and build a safety constraint model based on the geometric parameters and material parameters of the anti-permeability structure, adjust the dynamic allocation results through an improved optimization algorithm and add the adjusted dynamic allocation results to a path planning module, generate a transportation path set in combination with an improved ant colony algorithm and a heuristic search strategy, and obtain the optimal recovery route through a Pareto optimization algorithm.

[0065] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0067] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for managing the recovery of field medical sharps based on a modular anti-penetration structure, characterized in that: include: Construct a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features using an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain category information and hazard level information corresponding to the medical sharps; A classification scheduling system is constructed based on the category information and the hazard level information, a dynamic state representation is modeled through an improved spatiotemporal graph neural network and state representations at different spatiotemporal scales are extracted, the state representation is added to a time series prediction model and the state evolution trend is determined, an initial scheduling strategy is generated in combination with a hierarchical reinforcement learning algorithm, the initial scheduling strategy is optimized by an improved policy gradient algorithm in a pre-set two-stage optimization network, a dynamic allocation is performed in combination with an improved auction mechanism, a dynamic allocation result is obtained, and a safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeability structure, the dynamic allocation result is adjusted by an improved optimization algorithm and the adjusted dynamic allocation result is added to a path planning module, a transportation path set is generated in combination with an improved ant colony algorithm and a heuristic search strategy, and the optimal recovery route is obtained by solving the Pareto optimization algorithm.

2. The method according to claim 1, characterized in that A medical sharps feature recognition model is constructed, and visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps are collected and combined into medical sharps images. A multimodal image fusion algorithm based on deep hashing is executed on the medical sharps image to obtain a fused image. The shape, material, texture and internal structure corresponding to the field medical sharps are extracted based on the fused image and combined to generate medical sharps features. A multi-scale feature pyramid is constructed through a hierarchical attention mechanism, and the medical sharps features are adaptively recalibrated by combining the combination mechanism of channel attention and spatial attention. The recalibrated medical sharps features are enhanced through a deformable convolutional network and a residual connection structure. The enhanced medical sharps features are decoupled by an improved adversarial training method to obtain independent features. The importance of the independent features is evaluated and a feature selection strategy is established to obtain the key features of medical sharps, including: Collect visible light images, infrared thermal imaging images and ultrasonic scanning images of medical sharp objects, and construct feature extraction branches for the visible light images, infrared thermal imaging images and ultrasonic scanning images respectively, wherein the feature extraction branches include multiple convolutional layers, each of which is connected to a batch normalization layer and a rectified linear unit activation function, and the extracted features are mapped into feature vectors through a fully connected layer; Constructing a hash coding layer, mapping the feature vector into a binary hash code through a fully connected layer and a hyperbolic tangent activation function, calculating the Hamming distance between the binary hash codes corresponding to the visible light image, the infrared thermal imaging image, and the ultrasonic scanning image, screening feature-similar images based on a preset distance threshold, using a softmax function to calculate the fusion weight coefficients of the feature-similar images, and linearly combining the weighted features to obtain fused features; Construct a multi-layer feature pyramid network, adjust the number of channels of the fused features through convolution, use the maximum pooling operation to downsample layer by layer to obtain feature maps of different scales, calculate the channel attention weight and the spatial attention weight for the feature map, the channel attention weight is calculated by global average pooling and multi-layer full connection, the spatial attention weight is calculated by convolution operation and sigmoid activation function, and the channel attention weight and the spatial attention weight are multiplied to obtain a comprehensive attention map; Construct an offset prediction branch including multiple convolution layers, predict an offset value for each position of the fused feature, limit the offset value to a preset range through a hyperbolic tangent function, adjust the sampling position of the standard convolution according to the offset value, calculate the feature value at the sampling point by bilinear interpolation and perform convolution operation to obtain enhanced features, and perform residual connection between the enhanced features and the input features to obtain optimized features; Constructing a generator network and a discriminator network, wherein the generator network adopts an encoder-decoder structure, the encoder comprises a plurality of downsampling blocks, each of which comprises convolution, instance normalization, and rectified linear unit activation, the decoder comprises a plurality of upsampling blocks, each of which comprises transposed convolution, instance normalization, and rectified linear unit activation, the discriminator network comprises a plurality of convolutional layers and adopts a rectified linear unit activation function with leakage, and the optimized features are adversarially trained through the generator network and the discriminator network to obtain independent features; Calculate the mutual information value between the independent feature and the medical sharps category label, normalize the independent feature, count the probability distribution of the independent feature and the category label, calculate the mutual information value based on the probability distribution, and select the feature with the highest mutual information value as the key feature.

3. The method according to claim 2, characterized in that The Hamming distance between binary hash codes is calculated as follows: in, D ( H 1, H 2) Represents the first hash code H 1 and the second hash code H The Hamming distance between 2, n represents the number of sub-blocks the hash code is divided into, w (i) Indicates i The local sensitivity weight of each sub-block, m Represents the number of bits of the hash code in each sub-block, C [ i , j ] indicates the i The hash code position and j The correlation between the hash code positions, Represents the first hash code H 1 in i The first j The value of a binary bit, Represents the second hash code H 2 in i The first j The value of a binary digit.

4. The method according to claim 1, characterized in that: Based on the key features, the medical sharps are accurately located through an improved dense prediction network to obtain positioning features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for the medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharps to obtain semantic features, the positioning features, segmentation features and semantic features are fused and input into a causal reasoning network, a dependency relationship between features is established through an improved probabilistic graph model, a multi-level classifier is constructed based on the dependency relationship, and the medical sharps are classified in combination with an integrated learning strategy to obtain the category information and hazard level information corresponding to the medical sharps, including: Constructing an improved dense prediction network to locate the medical sharp instrument based on the key features, the dense prediction network includes a main branch and an auxiliary branch, the main branch includes a plurality of densely connected blocks, each densely connected block includes a plurality of convolutional units, each convolutional unit consists of a batch normalization layer, a rectified linear unit activation function and a convolutional layer, the auxiliary branch includes a spatial pyramid pooling module and a feature pyramid attention module, and the position coordinates and confidence scores of the medical sharp instrument are output through the dense prediction network to generate positioning features; Determine a region of interest based on the position coordinates of the positioning feature, and construct a hierarchical attention-guided segmentation network in the region of interest, wherein the segmentation network includes an encoder and a decoder, wherein the encoder extracts multi-level features using an improved residual network structure according to the confidence score of the positioning feature, and the decoder performs feature fusion through jump connections according to the multi-level features, and introduces a channel attention module and a spatial attention module in the feature fusion process to enhance the features, wherein the weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence of the positioning feature, and an edge perception module is introduced to optimize the edge details of the enhanced features to generate segmentation features; Input the segmentation features into a graph convolutional network to analyze the component composition relationship of the medical sharps, and construct a graph structure based on the edge information of the segmentation features, in which nodes represent different component regions, and edges represent the connection relationship between adjacent components. The graph convolutional network includes multiple graph convolutional layers, and the feature conversion parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation; According to the spatial distribution of the positioning feature, the regional segmentation result of the segmentation feature and the component relationship of the semantic feature, an attention-guided adaptive fusion strategy is used to calculate the feature weight, and the weighted feature is input into the causal reasoning network. The temporal flow of the causal reasoning network models the temporal dependency relationship based on the temporal evolution law of the feature, and the spatial flow models the spatial dependency relationship based on the regional distribution characteristics of the feature. A probabilistic graph model is constructed based on the dependency relationship output by the causal reasoning network, and the dependency relationship is mapped into a node connection relationship of a directed acyclic graph. The graph structure is optimized and the conditional probability distribution is determined in combination with prior knowledge. Based on the positioning features, the segmentation features and the semantic features, the overall contour features, the surface color distribution features and the texture features of the medical sharp instrument are respectively extracted, and the global features corresponding to the medical sharp instrument are obtained by combining them, and the edge detail features, the hand-held part features and the functional features of the medical sharp instrument are extracted to obtain the local features corresponding to the medical sharp instrument; A multi-level classifier is constructed according to the output of the probabilistic graphical model, including a coarse classifier for category division and a fine classifier for fine classification, wherein the coarse classifier adopts a random forest structure to classify the global features, and the fine classifier adopts a gradient boosting tree structure to finely classify the local features. The classification results are fused by voting, and the multi-level classifier is used as a base classifier. An integrated learning strategy is used to construct a classifier set, and model parameters are optimized through cross-validation to output category information and hazard level information of medical sharps.

5. The method according to claim 4, characterized in that The coarse classifier uses a random forest structure to classify the global features, the fine classifier uses a gradient boosting tree structure to finely classify the local features, and the voting method is used to fuse the classification results, including: Acquire the global features corresponding to the medical sharps, perform feature normalization processing, construct an improved random forest structure including multi-level decision units as a coarse classifier, set a feature selection module and a classification module in the improved random forest structure, perform adaptive feature sampling based on feature importance scores through the feature selection module, use the classification module to construct a decision tree based on information gain and apply regularization constraints, and output the basic category of the medical sharps; Acquire local features corresponding to the medical sharps, establish a cascaded gradient boosting tree structure as a detailed classifier, divide the cascaded gradient boosting tree structure into a feature extraction layer and a classification layer, use the feature extraction layer to perform attention weighting on the local features, perform residual learning through the classification layer and iteratively optimize the tree structure parameters, implement sample balancing and feature interaction operations, and output a detailed category of the medical sharps; An adaptive multimodal fusion mechanism is established, and a dynamic weight calculation module and a decision fusion module are configured. The dynamic weight calculation module is used to adjust the weight coefficient according to the classification effect. Multi-level voting is performed through the decision fusion module, and expert knowledge base arbitration is initiated when the confidence levels are inconsistent. The final classification results are output in combination with usage scenarios and security risk assessments.

6. The method according to claim 1, characterized in that A classification scheduling system is constructed based on the category information and the hazard level information. The dynamic state representation is modeled by an improved spatiotemporal graph neural network and the state representation at different spatiotemporal scales is extracted. The state representation is added to the time series prediction model and the state evolution trend is determined. The initial scheduling strategy is generated by combining the hierarchical reinforcement learning algorithm. The initial scheduling strategy is optimized by the improved policy gradient algorithm in the pre-set two-stage optimization network. The improved auction mechanism is combined for dynamic allocation to obtain the dynamic allocation result and a safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeability structure. The dynamic allocation result is adjusted by the improved optimization algorithm and the adjusted dynamic allocation result is added to the path planning module. The transportation path set is generated by combining the improved ant colony algorithm and the heuristic search strategy. The optimal recycling route obtained by solving the Pareto optimization algorithm includes: Constructing a spatiotemporal relationship graph of medical sharps, taking the medical sharps as nodes in the spatiotemporal relationship graph, wherein the nodes contain location information, category information, and hazard level information of the medical sharps, analyzing the spatial distribution characteristics of the medical sharps in different regions by setting multiple spatial convolution layers, analyzing the temporal distribution law of the medical sharps by setting temporal convolution layers, and inputting the spatial distribution characteristics and the temporal distribution law into a feature fusion module to generate a state representation of the medical sharps; The state representation is input into a time series prediction model, the time series prediction model includes a time series encoder and a time series decoder, the time series encoder extracts features from historical distribution data and historical recovery data of medical sharps according to the time sequence to obtain periodic features and burst features, the time series decoder predicts the distribution trend of medical sharps based on the periodic features and the burst features, and performs weighted processing on the historical distribution data and the historical recovery data through an attention mechanism to generate a state evolution trend; Based on the state evolution trend, a hierarchical reinforcement learning network is constructed, wherein the hierarchical reinforcement learning network includes a high-level strategy network and a low-level execution network. The high-level strategy network prioritizes the recovery tasks of medical sharps according to the region and type according to the state evolution trend to obtain a subtask set. The low-level execution network determines the recovery time and recovery route based on the subtask set to obtain an initial scheduling strategy, and the initial scheduling strategy is stored in an experience replay pool and the scheduling experience is selected through priority sampling. The initial scheduling strategy is input into a two-stage optimization network, the first stage of the two-stage optimization network evaluates the value of the scheduling action through a policy gradient algorithm and introduces an entropy regularization term to adjust the policy parameters, and the second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates a dynamic allocation result according to the bidding results and constraints of the recycling point; A safety constraint model is constructed based on the dynamic allocation result. The safety constraint model selects an impermeable container according to the physical characteristics and hazard level of the medical sharps, monitors the loading status of the impermeable container in real time, and automatically adjusts the transportation plan when a safety hazard is detected. The transportation plan is input into a path planning module. The path planning module uses an ant colony algorithm to construct a path map between recycling points, performs path search based on transportation distance, time window and safety factors, optimizes the path sequence through pheromone update and local search, and generates an optimal path set that meets multi-objective constraints.

7. The method according to claim 6, characterized in that The second stage of the two-stage optimization network allocates resources for medical sharps based on an auction mechanism, and generates dynamic allocation results based on the bidding results and constraints of the recycling points, including: Construct a bidding entity, set the medical sharps recycling point as the buyer entity, and set the medical sharps to be recycled as the seller entity. The medical sharps recycling point has a preset bidding budget and processing capacity constraint. The processing capacity constraint includes the number of medical sharps that can be processed per unit time and the processing capacity limit of different types of medical sharps. The bidding floor price is set based on the category information, danger level and urgency of the medical sharps to be recycled; The bidding rules are constructed by adopting a multi-round incremental bidding method. The medical sharps recycling point submits a bid that meets the maximum bid limit, budget limit and processing capacity requirements in each round of bidding. When the medical sharps recycling point submits a new bid, the price update mechanism is activated. The remaining medical sharps recycling points follow up with bids within the specified time, and the bidding process is promoted based on the minimum price increase limit. A bidding strategy optimization mechanism is built based on historical bidding data and current market conditions. The optimal bidding range is predicted by analyzing historical transaction prices, competitor bidding patterns, and market supply and demand relationships. Dynamic allocation constraints including time window constraints, path constraints, and capacity constraints are set. The time window constraints are used to ensure the time limit for completing the recycling of medical sharps, the path constraints are used to ensure the feasibility of the transportation routes, and the capacity constraints are used to limit the maximum processing capacity of the medical sharps recycling point. Construct a multi-objective evaluation system including the dimensions of allocation efficiency, cost control and risk management. The allocation efficiency dimension includes the timeliness index and balance index of medical sharps recovery, the cost control dimension includes the transportation cost index and the processing cost index, and the risk management dimension includes the safety hazard index and the emergency response capability index. A preliminary allocation plan is generated based on the bidding results and the dynamic allocation constraints, and the feasibility of the preliminary allocation plan is verified. When a constraint conflict occurs, a dynamic allocation result that meets actual needs is generated by adjusting the allocation quantity or replacing the medical sharps recycling point.

8. A field medical sharps recovery management system based on a modular anti-permeability structure, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to build a medical sharps feature recognition model, collect visible light images, infrared thermal imaging images and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharps images, execute a multimodal image fusion algorithm based on deep hashing on the medical sharps image to obtain a fused image, extract the shape, material, texture and internal structure corresponding to the field medical sharps based on the fused image and combine them to generate medical sharps features, build a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharps features by combining the combination mechanism of channel attention and spatial attention, enhance the recalibrated medical sharps features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharps features by an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of medical sharps; The second unit is used to accurately locate the medical sharps based on the key features of the medical sharps through an improved dense prediction network to obtain positioning features, use a hierarchical attention-guided segmentation algorithm to construct a refined mask for the medical sharps to obtain segmentation features, use a graph convolutional network to analyze the component composition relationship of the medical sharps to obtain semantic features, fuse the positioning features, segmentation features and semantic features and input them into a causal reasoning network, establish a dependency relationship between features through an improved probabilistic graph model, build a multi-level classifier based on the dependency relationship, and classify the medical sharps in combination with an integrated learning strategy to obtain category information and hazard level information corresponding to the medical sharps; The third unit is used to build a classified scheduling system based on the category information and the hazard level information, model dynamic state representation through an improved spatiotemporal graph neural network and extract state representations at different spatiotemporal scales, add the state representation to a time series prediction model and determine the state evolution trend, generate an initial scheduling strategy in combination with a hierarchical reinforcement learning algorithm, optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network, dynamically allocate in combination with an improved auction mechanism, obtain dynamic allocation results and build a safety constraint model based on the geometric parameters and material parameters of the anti-permeability structure, adjust the dynamic allocation results through an improved optimization algorithm and add the adjusted dynamic allocation results to a path planning module, generate a transportation path set in combination with an improved ant colony algorithm and a heuristic search strategy, and obtain the optimal recovery route through a Pareto optimization algorithm.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images