Block chain traceability-based hazardous waste full-chain intelligent management and control system and method thereof

Abstract

The invention discloses a hazardous waste full-chain intelligent management and control system and method based on block chain traceability, the system comprises a block chain traceability module, an AI identification and classification module, an intelligent matching recommendation module, a time sequence prediction module and a path optimization module, multi-party collaborative verification is realized through an alliance chain architecture, high-precision identification is realized through an improved ResNet network, and the risk of multi-party collaborative verification is reduced. Intelligent matching is realized through a knowledge graph, accurate prediction is realized through a Transform-LSTM hybrid architecture, dynamic path optimization is realized through a deep Q network, the five modules are deeply coupled to form a closed-loop cooperative system, hazardous waste full-chain credible traceability, intelligent management and control and safe transportation are realized, and the intelligent level and safety of hazardous waste management are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of hazardous waste management technology, specifically to a blockchain-based intelligent management system and method for the entire chain of hazardous waste. Background Technology

[0002] With the acceleration of industrialization, the generation of hazardous waste has increased dramatically, making its management a global environmental challenge. Hazardous waste possesses characteristics such as toxicity, corrosivity, and flammability; improper handling can cause serious harm to the ecological environment and human health. Current hazardous waste management mainly faces the following problems: First, regulatory blind spots and information silos are prominent issues. CN110705941A discloses an IoT-based hazardous waste storage monitoring system. This system includes a sensor module, a smart terminal module at the storage site, an information data module, and a platform system module. It monitors the hazardous waste storage stage through storage capacity ratio early warning algorithms, overdue early warning algorithms, and deviation exceeding limit early warning algorithms. However, this system only focuses on the storage stage and lacks traceability capabilities for the entire chain of hazardous waste generation, transfer, and disposal. It cannot solve the information silo problem between waste-generating enterprises, transportation enterprises, and disposal enterprises, nor does it use blockchain technology to ensure the immutability of data.

[0003] Second, the automation level of hazardous waste identification and classification is low. Traditional hazardous waste management relies on manual identification and classification, which is inefficient and prone to errors. Although existing technologies are beginning to adopt image recognition technology, ordinary convolutional neural networks have insufficient accuracy in complex scenarios, especially in recognizing situations such as damaged hazardous waste packaging and blurred labels, which cannot meet the needs of real-time monitoring.

[0004] Third, the matching of disposal facilities is unreasonable. The existing system lacks an intelligent mechanism for recommending disposal facilities, and disposal units are often selected manually based on experience. This fails to comprehensively consider multiple factors such as the characteristics of hazardous waste, the suitability of disposal processes, transportation distance, and disposal costs, resulting in low disposal efficiency, high costs, and potential safety hazards.

[0005] Fourth, there is a lack of ability to predict the amount of hazardous waste generated. Most existing management systems are reactive and cannot predict future hazardous waste generation based on enterprise production plans and historical data. This leads to unreasonable planning of disposal capacity, often resulting in insufficient or excessive disposal capacity, which affects emergency response capabilities.

[0006] Fifth, there is insufficient optimization of transportation routes. Hazardous waste transportation is high-risk and requires comprehensive consideration of multiple factors such as road conditions, environmentally sensitive areas, and emergency response capabilities. Existing systems mostly use fixed routes or simple shortest path algorithms, failing to dynamically adjust routes based on real-time traffic conditions and environmental risks, thus posing safety hazards.

[0007] Therefore, there is an urgent need for a comprehensive solution that can achieve intelligent management and control of the entire hazardous waste chain, reliable data traceability, automatic identification and classification, intelligent matching and recommendation, predictive management and dynamic path optimization. Summary of the Invention

[0008] The purpose of this invention is to provide a blockchain-based intelligent management system and method for the entire chain of hazardous waste. It achieves multi-party collaborative verification and data immutability through blockchain technology, high-precision automatic identification of hazardous waste through an improved ResNet network, intelligent matching of disposal facilities through knowledge graphs, accurate prediction of hazardous waste generation through a Transformer-LSTM hybrid model, and dynamic optimization of transportation routes through deep reinforcement learning. This addresses the problems of regulatory blind spots, information silos, low identification accuracy, unreasonable matching, lack of prediction, and insufficient route optimization in existing technologies.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a blockchain-based intelligent management system for the entire hazardous waste chain, comprising a blockchain traceability module, an AI identification and classification module, an intelligent matching and recommendation module, a time-series prediction module, and a route optimization module. The blockchain traceability module establishes a multi-party collaborative verification network based on a consortium blockchain architecture, ensuring data immutability and end-to-end traceability. The AI ​​identification and classification module employs an improved ResNet network, enhancing hazardous waste identification accuracy through a dual-attention fusion structure, and performs cross-validation with the blockchain traceability module. The intelligent matching and recommendation module automatically matches the optimal disposal facility based on knowledge graphs and multi-objective optimization algorithms. The time-series prediction module uses a Transformer-LSTM hybrid architecture to predict hazardous waste generation and feeds this prediction back to the matching module to adjust disposal capacity planning. The route optimization module, based on a deep Q-network, dynamically plans the safest transportation route and returns the transportation data to the blockchain to complete closed-loop traceability.

[0010] The five core modules form a deeply coupled, closed-loop collaborative system: data output from the blockchain traceability module is cross-validated by the AI ​​recognition and classification module; the recognition results from the AI ​​recognition and classification module are input to the intelligent matching and recommendation module as a matching basis; the matching results from the intelligent matching and recommendation module are passed to the path optimization module as input for path planning; the prediction data from the time-series prediction module is fed back to the intelligent matching and recommendation module to adjust the planning of handling capabilities; and the transportation data from the path optimization module is returned to the blockchain traceability module to complete the closed loop. This coupling mechanism enables mutual promotion and synergistic effects among the modules: cross-validation between blockchain and AI recognition improves data credibility; the combination of knowledge graph and time-series prediction enables predictive matching; and the synergy between deep reinforcement learning and blockchain ensures a balance between security and efficiency.

[0011] Compared with the prior art, the present invention has the following beneficial effects: First, it enables reliable traceability of hazardous waste across the entire supply chain. Through a consortium blockchain architecture and smart contracts, data from each stage of hazardous waste generation, transfer, and disposal is stored on the blockchain, enabling multi-party collaborative verification and ensuring data immutability. This solves the problems of information silos and regulatory blind spots. Compared to the limitation of CN110705941A, which only focuses on the storage stage, this invention achieves full-chain traceability from generation to disposal.

[0012] Second, it improves the accuracy and robustness of hazardous waste identification. The improved ResNet network, through a dual-attention fusion structure, achieves an accuracy rate of over 98% in complex scenarios, which is about 3% higher than that of the ordinary ResNet network. In particular, it significantly enhances the ability to identify damaged packaging and blurred labels, and further improves the credibility of the identification results through a cross-validation mechanism with blockchain data.

[0013] Third, it achieves intelligent and precise matching. The knowledge graph-based matching and recommendation algorithm comprehensively considers multiple factors such as hazardous waste characteristics, disposal processes, transportation distance, and disposal capacity. Through multi-objective optimization, it finds the optimal balance between disposal cost, transportation distance, and disposal capacity. Compared with manual experience-based selection, the matching accuracy is improved by more than 40%, and the disposal cost is reduced by about 25%.

[0014] Fourth, it enables predictive management. The Transformer-LSTM hybrid architecture achieves a prediction accuracy of over 95% for hazardous waste generation, with a prediction error reduced by approximately 30% compared to a single LSTM model. This allows for more rational planning of disposal capacity, significantly improves emergency response capabilities, and avoids situations of insufficient or excessive disposal capacity.

[0015] Fifth, dynamic safety path optimization is achieved. The reinforcement learning algorithm based on deep Q-networks comprehensively considers road conditions, environmentally sensitive areas, and emergency rescue capabilities to dynamically plan the safest transportation route. Compared with fixed routes, transportation time is reduced by about 15%, safety is improved by more than 35%, and it can adaptively adjust according to real-time conditions.

[0016] Sixth, it achieves deep synergy and efficiency enhancement among modules. The five core modules form a complete closed loop through deep coupling at the parameter level, state level, and logic level, achieving a synergistic effect of 1+1>2. The overall system performance is improved by more than 50% compared to the combination of single modules. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of the blockchain-based intelligent management and control system for hazardous waste across the entire supply chain, as presented in this invention. Figure 2 This is a schematic diagram of the blockchain traceability module of the present invention; Figure 3 This is a schematic diagram of the structure of the AI ​​recognition and classification module of the present invention; Figure 4 This is a schematic diagram of the intelligent matching and recommendation module of the present invention; Figure 5 This is a schematic diagram of the time series prediction module of the present invention; Figure 6 This is a schematic diagram of the path optimization module of the present invention; Figure 7 This is a flowchart illustrating the intelligent management and control method for the entire chain of hazardous waste based on blockchain traceability, as proposed in this invention. Detailed Implementation

[0018] Please refer to the attached document. Figures 1-7 To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0019] Reference Figure 1 This embodiment provides a blockchain-based intelligent management and control system for hazardous waste across the entire supply chain, including a blockchain traceability module 1, an AI identification and classification module 2, an intelligent matching and recommendation module 3, a time-series prediction module 4, and a path optimization module 5.

[0020] Reference Figure 2 The blockchain traceability module 1 includes a consortium blockchain network unit, a smart contract execution unit, and a data upload unit.

[0021] The consortium blockchain network unit is built on Hyperledger Fabric 2.5. Hyperledger Fabric is a permissioned blockchain framework that supports multi-organizational participation and modular design. In this embodiment, waste-generating companies, transportation companies, disposal companies, and regulatory departments are set as consortium nodes, with each node deploying peer nodes (Peers) and ordering nodes (Orderers). Waste-generating company nodes are responsible for recording hazardous waste generation information, transportation company nodes are responsible for recording transfer process information, disposal company nodes are responsible for recording disposal completion information, and regulatory department nodes are responsible for reviewing and supervising the entire process. The consortium blockchain uses the Raft consensus algorithm, which has higher efficiency and lower latency compared to the Byzantine fault-tolerant algorithm, making it suitable for consortium blockchain scenarios. The network topology adopts a star structure, with the regulatory department node as the central node and other nodes as edge nodes, ensuring the authority of supervision and the consistency of data.

[0022] The smart contract execution unit defines transfer manifest review rules, disposal process supervision rules, and multi-party signature verification rules. The transfer manifest review rules include: the waste-generating company submits a transfer manifest application; the system automatically checks the completeness and compliance of information such as hazardous waste category, quantity, transportation company qualifications, and disposal company qualifications. If the check passes, a transfer manifest number is generated and uploaded to the blockchain; if the check fails, an error message is returned requiring modification. The disposal process supervision rules include: the transportation company uploads real-time monitoring data such as GPS location, temperature, and humidity during transportation; the system automatically determines whether it deviates from the predetermined route or exceeds environmental parameter thresholds; if abnormalities are detected, an alert is triggered and the regulatory department is notified. After disposal is completed, the disposal company uploads a disposal report; the system automatically checks information such as disposal method, disposal volume, and residue destination, and requires signatures from the waste-generating company, transportation company, disposal company, and regulatory department for confirmation. Only with complete signatures from all four parties can the disposal process be completed and the transfer manifest closed. Multi-party signature verification uses the Elliptic Curve Digital Signature Algorithm (ECDSA) to ensure the security and non-repudiation of signatures.

[0023] The data upload unit hashes and encrypts RFID tag data, smart weighing data, video surveillance data, and AI recognition results before storing them on the blockchain. The RFID tags are UHF tags, operating at 860-960MHz, with a recognition distance of 5-10m. They automatically collect information such as the unique identifier of hazardous waste packaging, hazardous waste category, generation time, and generating unit. The smart weighing equipment uses a high-precision electronic scale with an accuracy of 0.01kg, capable of measuring the weight of hazardous waste in real time and uploading data via an RS485 interface. The video surveillance uses a high-definition camera with a resolution of 1920×1080 and a frame rate of 30fps, capable of monitoring the entire process of hazardous waste generation, transfer, and disposal in real time, and uploading data after compression using H.264 encoding. The AI ​​recognition results include information such as hazardous waste category, confidence level, and recognition time. All data is first hashed using SHA-256 before being uploaded to the blockchain, generating a 256-bit hash value. The hash value and data digest are then uploaded to the blockchain, while the original data is stored in the distributed file system IPFS. The hash value verifies the integrity and consistency of the data. This approach ensures the immutability of data while avoiding performance issues caused by storing large amounts of raw data on the blockchain.

[0024] In a preferred embodiment, the blockchain traceability module 1 also includes a cross-chain bridging mechanism to support data exchange with hazardous waste management blockchains in other provinces and cities, enabling cross-regional traceability of hazardous waste transfers. The cross-chain bridging employs Hash Time Locked Contracts (HTLCs) to ensure the atomicity and security of cross-chain transactions.

[0025] Reference Figure 3 The AI ​​recognition and classification module 2 includes an image acquisition unit, a feature extraction unit, and a classification decision unit.

[0026] The image acquisition unit captures high-resolution images of hazardous waste packaging, labeling, and appearance using industrial cameras. The industrial cameras employ CMOS sensors with a resolution of 2048×2048 and a frame rate of 60fps. Equipped with autofocus lenses and ring LED fill lights, they can acquire clear images of hazardous waste under various lighting conditions. The cameras are installed at key nodes in the generation, transfer, and disposal of hazardous waste, such as the exit of the waste-generating workshop, loading and unloading points for transport vehicles, and entrances to disposal facilities. Images are transmitted in real-time to the image processing server via a gigabit Ethernet interface. Image preprocessing includes noise reduction, contrast enhancement, and color correction. Bilateral filtering is used to remove Gaussian noise, histogram equalization is used to enhance contrast, and a white balance algorithm is used to correct color deviations, ensuring image quality meets recognition requirements.

[0027] The feature extraction unit extracts depth features from hazardous waste images based on an improved ResNet50 network. ResNet50 is a deep residual network containing 50 convolutional layers, which solves the gradient vanishing problem in deep networks through residual connections. This embodiment improves upon ResNet50 by introducing a channel attention submodule and a spatial attention submodule after each residual block, forming a dual attention fusion structure.

[0028] The channel attention submodule is implemented as follows: First, the feature map is processed... Perform global average pooling and global max pooling to generate two sets of channel descriptors. and ,in For the number of channels, For height, The width is [value]. Then, the two sets of descriptors are input into a shared multilayer perceptron (MLP). The MLP contains two fully connected layers, with the first layer having [number] neurons. (in To reduce the ratio, this embodiment takes The activation function is ReLU, and the number of neurons in the second layer is... There is no activation function. Two sets of descriptors are processed by an MLP to obtain two sets of weight vectors. These two weight vectors are then added together and passed through a sigmoid activation function to generate channel attention weights. The channel attention weights are multiplied element-wise with the original feature map to obtain the channel-weighted feature map. .

[0029] The spatial attention submodule is implemented by weighting the feature maps of the channels. Average pooling and max pooling are performed on the channel dimension respectively to generate two sets of spatial descriptors. and The two sets of descriptors are concatenated along the channel dimension and then passed through a convolutional layer to generate spatial attention weights. The kernel size is The activation function is Sigmoid. The spatial attention weights are element-wise multiplied with the channel-weighted feature maps to obtain the final attention-enhanced features. .

[0030] The advantage of the dual-attention fusion structure lies in the following: channel attention focuses on the importance of different feature channels, highlighting feature channels related to hazardous waste categories; spatial attention focuses on the importance of different spatial locations in the feature map, highlighting key areas of hazardous waste packaging and labeling. The combination of these two approaches significantly improves feature representation capabilities and recognition accuracy.

[0031] During the training phase, a transfer learning strategy was employed. First, pre-training was performed on the ImageNet dataset, followed by fine-tuning on a self-built hazardous waste image dataset. This dataset contains 30 categories of hazardous waste, with 1000-3000 images per category, totaling 50,000 images. Image sources included on-site photography, web scraping, and data augmentation. Data augmentation included random cropping, random flipping, random rotation, and color jittering to expand the dataset size and improve the model's generalization ability. Training utilized the SGD optimizer with an initial learning rate of 0.01, a dynamic moment of 0.9, a weight decay of 0.0001, a batch size of 32, and 100 training epochs. The learning rate employed a cosine annealing strategy, decaying to 0.5 times the initial learning rate every 10 epochs. The cross-entropy loss function was used.

[0032] The classification decision unit outputs a probability distribution of hazardous waste categories based on a fully connected layer and a Softmax activation function, and sets a confidence threshold for classification decisions. The fully connected layer maps the feature vector output by the feature extraction unit into a 30-dimensional vector, with each dimension corresponding to a hazardous waste category. The Softmax activation function converts the 30-dimensional vector into a probability distribution with a probability sum of 1. The confidence threshold is set to 0.85. If the maximum probability is greater than 0.85, the identification is considered successful, and the corresponding hazardous waste category is output; if the maximum probability is less than 0.85, the identification is considered uncertain, triggering a manual review process. The identification results include information such as hazardous waste category, confidence level, identification time, and image number, which are fed back to the blockchain traceability module 1 in real time for cross-validation. The cross-validation logic is as follows: compare the AI ​​identification results with the hazardous waste categories declared by the waste-generating enterprises. If they match, it is marked as normal; if they do not match, an alert is triggered and recorded on the blockchain, while simultaneously notifying the regulatory authorities to conduct on-site verification.

[0033] In a preferred embodiment, the AI ​​identification and classification module 2 also supports the identification of hazardous waste packaging damage, blurred labels, and abnormal appearance. By adding an anomaly detection branch, anomaly detection algorithms such as Isolation Forest or One-Class SVM are used to judge the anomalies in the identification results. If an anomaly is judged, a manual review process is triggered.

[0034] Reference Figure 4 The intelligent matching and recommendation module 3 includes a knowledge graph construction unit, a graph reasoning unit, and a multi-objective optimization unit.

[0035] The knowledge graph construction unit builds a knowledge graph containing entities related to hazardous waste characteristics, disposal processes, disposal facilities, and geographical locations, and defines the applicability, qualification, and geographical relationships between entities. The knowledge graph is stored in the Neo4j graph database, which supports efficient graph querying and analysis. Hazardous waste characteristic entities include attributes such as hazardous waste category, physical form, chemical composition, and hazardous characteristics, totaling 30 types of hazardous waste. Disposal process entities include process types such as incineration, landfill, physicochemical treatment, and resource utilization, as well as process parameters such as temperature, pressure, and time, totaling 15 disposal processes. Disposal facility entities include attributes such as facility name, disposal qualification, disposal capacity, geographical location, and historical performance, totaling 200 disposal facilities. Geographical location entities include administrative divisions such as province, city, district / county, and street, as well as latitude and longitude coordinates.

[0036] The applicability relationship defines the matching relationship between hazardous waste characteristics and disposal processes. For example, oily sludge is suitable for incineration and resource utilization processes, waste acid is suitable for neutralization treatment processes, and heavy metal waste is suitable for solidification and stabilization processes. Applicability relationships are obtained by consulting the National Hazardous Waste List, disposal technical specifications, and expert knowledge, resulting in 150 defined applicability relationships. The qualification relationship defines the licensing relationship between disposal institutions and disposal processes. For example, if a disposal institution has the qualifications for incineration and landfill disposal, then that institution can dispose of hazardous waste suitable for incineration and landfill processes. Qualification relationships are obtained by consulting hazardous waste operation licenses issued by the ecological and environmental departments. The geographical relationship defines the distance relationship between the hazardous waste generation location and the disposal institution, obtained by calculating the Euclidean distance of latitude and longitude coordinates or by using a map API to calculate the actual road distance.

[0037] The graph reasoning unit uses a graph neural network (GNN) for knowledge graph embedding representation learning. It calculates the matching score between hazardous waste characteristics and disposal processes through node feature aggregation and relationship modeling. The GNN employs a graph convolutional network (GCN) architecture, containing two graph convolutional layers. For nodes... Its embedded representation It is calculated by aggregating the features of neighboring nodes and edge features. The specific formula is: , in, For nodes In the Layer embedding representation, For nodes The set of neighboring nodes, For the first The weight matrix of the layer, For the first Layer bias vector, The activation function is ReLU (in this example, ReLU is used). This is the aggregation function (average aggregation is used in this example).

[0038] For hazardous waste characteristic nodes and processing nodes Their matching score The cosine similarity between the embeddings of the two nodes is obtained by calculating: , in, This is an embedded representation of hazardous waste characteristic nodes. To handle the embedded representation of process nodes, Let be the Euclidean norm of the vector. Match score. A higher value indicates a higher degree of matching.

[0039] During the training phase, a knowledge graph completion task was used for embedding representation learning, with the goal of predicting missing applicability relations. The loss function used was binary cross-entropy loss, the optimizer was Adam, the learning rate was 0.001, the batch size was 64, and the number of training epochs was 50. The training data included positive samples (known applicability relations) and negative samples (randomly sampled non-applicability relations), with a positive-to-negative sample ratio of 1:5.

[0040] The multi-objective optimization unit uses the non-dominated sorting genetic algorithm NSGA-II to find a Pareto optimal solution set among the disposal cost objective function, the transportation distance objective function, and the disposal capacity objective function. NSGA-II is a classic multi-objective optimization algorithm that maintains solution diversity and convergence through fast non-dominated sorting and congestion distance calculation.

[0041] Disposal cost objective function Defined as: , in, For decision variables (the chosen disposal agency). The transportation cost coefficient (unit: yuan / ton·km, taken in this embodiment) is... ), The transport distance is in kilometers. This is a disposal cost coefficient (unit: yuan / ton, determined according to different disposal processes and disposal institutions, ranging from 500 to 3000 yuan / ton). The value is the mass of hazardous waste (in tons). The objective is to minimize disposal costs.

[0042] Transportation distance objective function Defined as: , in, This represents the actual road distance (in km) between the hazardous waste generation site and the disposal facility, calculated using a map API. The goal is to minimize the transportation distance.

[0043] Disposal capability objective function Defined as: , in, The annual disposal capacity of the disposal facility (in tons / year). This represents the current amount of hazardous waste handled by the disposal agency (in tons). The mass of hazardous waste to be disposed of (in tons). When hour, This indicates that the handling capacity is sufficient; when hour, This indicates insufficient handling capacity, and the greater the excess, the better. The larger the capacity, the better. The goal is to minimize the proportion of insufficient capacity.

[0044] The optimization process of NSGA-II is as follows: First, initialize the population, population size... Each individual corresponds to a candidate disposal institution. The three objective function values ​​for each individual are calculated. Then, a fast non-dominated sort is performed, assigning individuals to different non-dominated levels: the first level contains all individuals on the Pareto front (not dominated by any other individuals), the second level contains individuals on the Pareto front after removing the first level, and so on. Next, the crowding distance of each individual is calculated. Crowding distance represents the density of neighboring individuals in the target space; a larger crowding distance indicates a sparser solution around that individual, and retaining individuals with large crowding distances helps maintain solution diversity. Crowding distance The calculation formula is: , in, and Individuals In the target The objective function values ​​of adjacent individuals on the graph. and The target The maximum and minimum values. The crowding distance for boundary individuals is set to infinity.

[0045] Then, a new population is generated through tournament selection, simulated binary crossover, and polynomial mutation. Tournament selection is performed randomly each time. For each individual, individuals with lower non-dominated levels are selected; if the non-dominated levels are the same, individuals with larger crowding distances are selected. The crossover probability of binary crossover is simulated. Distribution index The probability of polynomial mutation. Distribution index Iterative updates The process eventually yields a Pareto optimal solution set, which contains 20-50 non-dominated solutions, covering different trade-offs between disposal cost, transportation distance, and disposal capacity.

[0046] In practical applications, a solution from the Pareto optimal set is selected as the final recommended solution based on the decision-maker's preferences. Preferably, a weighted summation of the three objective functions is used, with weights of [weights to be specified]. (Disposal costs) (Transportation distance) (Disposal capacity) Select the solution with the smallest weighted sum as the recommended solution.

[0047] In a preferred embodiment, the intelligent matching and recommendation module 3 also supports dynamic evaluation of disposal agencies, adjusting the matching score based on the historical performance of the disposal agencies (such as on-time completion rate, compliance disposal rate, customer satisfaction, etc.) and prioritizing the recommendation of disposal agencies with good historical performance.

[0048] Reference Figure 5 The time-series prediction module 4 includes a feature encoding unit, a sequence decoding unit, and an error correction unit.

[0049] The feature encoding unit uses a Transformer encoder to perform multi-head self-attention encoding on enterprise production plans, raw material procurement quantities, and process parameters. It extracts global dependencies of multi-dimensional temporal features through positional encoding and a feedforward neural network. The Transformer encoder is a sequence modeling architecture based on a self-attention mechanism, which has stronger parallel computing capabilities and longer dependency capture capabilities compared to traditional recurrent neural networks.

[0050] Input features include enterprise production plans (such as monthly production volume and production batches), raw material procurement quantities (such as monthly procurement quantities and raw material types), and process parameters (such as reaction temperature, reaction time, and catalyst dosage), totaling... Each feature dimension. The length of the input sequence is... (Data from the past 24 months). Input features are first mapped to... A 3D vector is used, and then positional encoding is added to obtain the encoder input. Position encoding uses sine and cosine functions: , , in, For location index, For dimensional indexing, This is the position encoding matrix.

[0051] Transformer encoder includes The system consists of layers, each containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention mechanism captures dependencies between different positions in a sequence by calculating the similarity between the query, key, and value. The specific formula is: , in, , , , The weight matrix is ​​a learnable matrix. For each head dimension, The number of heads. Multi-head attention divides the input into... Each head independently calculates its attention, and then the results are concatenated and transformed linearly to obtain the output.

[0052] The feedforward neural network sublayer contains two fully connected layers, with the first layer having the following number of neurons: The activation function is ReLU, and the number of neurons in the second layer is... Add residual connections and layer normalization after each sub-layer.

[0053] The sequence decoding unit performs sequence decoding on the encoded features based on the LSTM decoder, capturing long-term dependencies in hazardous waste generation through forget gates, input gates, and output gates. LSTM is an improved recurrent neural network that solves the gradient vanishing problem of traditional RNNs through gating mechanisms, enabling it to capture long-term dependencies.

[0054] LSTM decoder includes Layers, each layer contains One hidden unit. For time step The calculation formula for an LSTM cell is: , , , , , , in, For the Gate of Oblivion For input gate, For output gate, In cellular state, In hidden state, Candidate cell state, For input, This is the weight matrix. For bias vectors, It is the Sigmoid activation function. This is for element-wise multiplication. The forget gate controls how much historical information is retained, the input gate controls how much new information is received, and the output gate controls how much information is output.

[0055] Decoder input The output features of the encoder are mapped to the following values ​​through a linear transformation: A 3D vector. The decoder's output. The predicted hazardous waste generation is mapped to a 1D vector through a fully connected layer. The prediction timeframe is future. Months.

[0056] The error correction unit adds the encoded features to the decoded output through residual connections and adaptively corrects the prediction error based on an attention weight mechanism. The formula for the residual connections is: , in, For the decoder at time step The output, The output characteristics of the encoder (averaged over all time steps). This is a linear transformation. Residual connections can alleviate the vanishing gradient problem in deep networks and provide a direct information transfer path from the encoder to the decoder.

[0057] The attention weighting mechanism assigns different weights to different time steps of the encoder by calculating the similarity between the decoder's hidden state and the encoder's output features, thus highlighting important historical information. The calculation formula is: , , , in, For the encoder at time step The output, This is the weight matrix. For the weight vector, This is the weighted context vector. (The context vector is then...) With decoder hidden state The final predicted value is obtained by splicing the data and passing it through a fully connected layer.

[0058] During the training phase, a supervised learning approach was employed. The input consisted of the company's production plans, raw material procurement volumes, and process parameters for the past 24 months, while the output was the hazardous waste generation for the next 12 months. Training data was derived from historical records, encompassing 500 companies, with 100 months of data per company, totaling 50,000 samples. The loss function was Mean Squared Error (MSE), the optimizer was Adam, the learning rate was 0.001, the batch size was 32, and the number of training epochs was 50. The model achieved a prediction accuracy of 95.3% on the test set (evaluated using Mean Absolute Percentage Error (MAPE), representing a reduction in prediction error of approximately 30% compared to a single LSTM model.

[0059] In a preferred embodiment, the time series prediction module 4 also supports multi-task learning, simultaneously predicting the amount of hazardous waste generated, the distribution of hazardous waste categories, and the trend of hazardous waste generation (rising, stable, or declining). Multi-task prediction is achieved by sharing an encoder and setting up multiple decoders, thereby improving the model's generalization ability and prediction accuracy.

[0060] Reference Figure 6 The path optimization module 5 includes a state space construction unit, a DQN network unit, and an experience playback unit.

[0061] The state space construction unit constructs a state vector containing the current location, road congestion index, distance to environmentally sensitive areas, distribution of emergency rescue stations, and hazardous waste type. The definition of is: , in, These are the latitude and longitude coordinates of the current location. The current location's road congestion index (range 0-10, where 0 indicates smooth traffic and 10 indicates severe congestion, obtained by calling the real-time traffic API). This is the distance (in km) from the current location to the nearest environmentally sensitive area (such as a school, hospital, residential area, or water source). This is the distance (in km) from the current location to the nearest emergency rescue station (such as a fire station or medical emergency station). Hazardous waste types are encoded (using one-hot encoding, with a dimension of 30). The total dimension of the state vector is... .

[0062] The road congestion index is obtained by calling the real-time traffic API of Baidu Maps or Gaode Maps. The API returns the congestion status of roads around the current location, including levels such as smooth traffic, slow traffic, congested, and severe congestion. The system maps the levels to values ​​from 0 to 10. The location information of environmentally sensitive areas is pre-stored in a geographic information database. By calculating the distance between the current location and environmentally sensitive areas, the minimum distance is selected as the index. The location information of emergency rescue stations is also pre-stored. By calculating the distance between the current location and the emergency rescue station, the minimum distance is selected as the location. Hazardous waste type coding uses one-hot encoding, such as HW01 coded as follows: HW02 encoding is There are 30 categories in total.

[0063] DQN network units construct a deep Q-network, extracting state features through convolutional layers and outputting Q-value estimates for each action through fully connected layers. DQN is a deep learning-based reinforcement learning algorithm that approximates the Q-function through a neural network, enabling it to handle high-dimensional state spaces.

[0064] The structure of the DQN network is as follows: Input layer receives state vector Feature extraction is performed through two fully connected layers. The first layer has 128 neurons, and the second layer has 64 neurons, both using ReLU activation. The output layer is a fully connected layer with [number of neurons missing]. (Action space dimension), no activation function, output Q-value estimates for each action. ,in These are network parameters.

[0065] Action space The definition is: starting from the current location, select the next path node. Path nodes include road intersections and the destination (disposal agency), totaling... There are 10 candidate nodes. The action space is represented by discrete actions, and each action corresponds to selecting one candidate node.

[0066] reward function The design takes into account transportation time, safety, and cost: , in, This is the travel time from the current location to the next node (in minutes, calculated by calling the map API). Environmental penalties (when passing through environmentally sensitive areas) ,otherwise ), For safety bonus items (when approaching an emergency rescue station) ,otherwise ), , , These are the weighting coefficients. The goal is to maximize the cumulative reward.

[0067] During the training phase, the following methods were adopted: - Greedy strategy selects actions based on probability. Randomly select an action (exploration) with probability. Choose the action with the highest Q value (exploit it). The initial value is 1.0, which gradually decays to 0.1 during training, with a decay rate of 0.995. After each action is selected, the action is executed and the next state is observed. and instant rewards Transition samples Store in the experience replay pool.

[0068] The experience replay unit stores state transition samples, breaks sample correlation through mini-batch random sampling, and improves training stability by employing a periodic update mechanism for the target network. The capacity of the experience replay pool is [missing information]. When the experience replay pool is full, the oldest sample is deleted. During training, mini-batches of samples are randomly sampled from the experience replay pool, with the batch size... The network parameters are updated by minimizing the mean squared error loss function between the current Q-network output and the target Q-value calculated by the target Q-network. The loss function is: , in, For experience replay pool, As a discount factor, These are the parameters of the target network. Each parameter of the target network... Each step synchronizes the parameters of the current network once, that is... The introduction of the target network can stabilize the training process and avoid oscillations in Q-value estimation.

[0069] The optimizer used was Adam with a learning rate of 0.0005 and 500 training epochs. Training data was derived from a simulation environment built on real road networks and traffic data, including 50 waste-generating enterprises and 20 disposal facilities, totaling 1000 transportation tasks. The model's performance on the test set showed: average transportation time reduced by 15.3% compared to fixed routes; the number of times it passed through environmentally sensitive areas decreased by 42.6%; and the coverage of emergency rescue stations increased by 31.8%.

[0070] In a preferred embodiment, the path optimization module 5 also supports multi-vehicle collaborative path planning. By constructing a multi-agent reinforcement learning framework, such as the multi-agent deep deterministic policy gradient (MADDPG) algorithm, it can achieve collaborative scheduling and path optimization of multiple transport vehicles, thereby avoiding conflicts and congestion between vehicles.

[0071] The five core modules of this system form a complete closed-loop collaborative system through deep coupling at the parameter level, state level and logic level, achieving a synergistic effect of 1+1>2.

[0072] The coupling between blockchain traceability module 1 and AI identification and classification module 2: Blockchain traceability module 1 transmits hazardous waste declaration information (including hazardous waste category, quantity, etc.) to AI identification and classification module 2 as prior information. AI identification and classification module 2 then feeds back the identification results to blockchain traceability module 1 for cross-verification. When the identification result matches the declaration information, the system automatically marks it as normal and continues the subsequent process; when the identification result does not match the declaration information, the system triggers an early warning and records the warning information to the blockchain, while simultaneously notifying the regulatory authorities to conduct on-site verification. This cross-verification mechanism significantly improves data credibility, with a false declaration detection rate exceeding 92%.

[0073] The coupling between AI identification and classification module 2 and intelligent matching and recommendation module 3: The identification results (hazardous waste category, confidence level) of AI identification and classification module 2 are input into intelligent matching and recommendation module 3 as the matching basis. Intelligent matching and recommendation module 3 searches for applicable disposal processes in the knowledge graph based on the hazardous waste category, then searches for qualified disposal facilities based on the disposal processes, and finally selects the optimal disposal facility through multi-objective optimization. The accuracy of the identification results directly affects the accuracy of the matching recommendation, forming a parameter-level coupling between the two. When the identification confidence level is lower than the threshold, intelligent matching and recommendation module 3 triggers a manual review process; matching and recommendation are only performed after manual confirmation.

[0074] The coupling between the intelligent matching and recommendation module 3 and the time-series prediction module 4: The time-series prediction module 4 predicts the amount of hazardous waste generated in the next 12 months and feeds the prediction results back to the intelligent matching and recommendation module 3 to adjust the disposal capacity planning. Based on the predicted hazardous waste generation, the intelligent matching and recommendation module 3 assesses in advance whether the disposal capacity of existing facilities is sufficient. If the predicted generation exceeds the remaining disposal capacity of existing facilities, it prioritizes recommending other facilities with sufficient capacity or suggests that waste-generating companies dispose of waste in batches. This predictive matching mechanism significantly improves the utilization rate of disposal capacity and emergency response capabilities, reducing insufficient disposal capacity by 67%.

[0075] The coupling between the intelligent matching recommendation module 3 and the route optimization module 5: The matching results (selected disposal facilities and their geographical locations) from the intelligent matching recommendation module 3 are transmitted to the route optimization module 5 as the destination for route planning. The route optimization module 5 dynamically plans the safest transportation route based on the location of the waste-generating enterprise (starting point) and the location of the disposal facility (ending point), combined with road conditions, environmentally sensitive areas, and emergency rescue capabilities. The route planning results (route node sequence, estimated travel time) are fed back to the intelligent matching recommendation module 3 as accurate estimates of disposal costs and transportation distances, used for recalculation of multi-objective optimization. The two modules form a state-level coupling, achieving global optimization of matching recommendation and route planning through iterative optimization.

[0076] The coupling between the route optimization module 5 and the blockchain traceability module 1: During transportation, the route optimization module 5 acquires real-time monitoring data such as the GPS location, speed, temperature, and humidity of the transport vehicles, and returns this data to the blockchain traceability module 1 for on-chain storage, forming a complete transportation trajectory record. The blockchain traceability module 1 compares the predetermined route with the actual route. If the transport vehicle deviates from the predetermined route by more than a threshold (e.g., a deviation distance exceeding 5km or a deviation time exceeding 30 minutes), an alert is triggered and the regulatory authorities are notified. This closed-loop traceability mechanism ensures the transparency and traceability of the transportation process, with a detection rate of over 88% for transportation violations.

[0077] The five modules form a complete closed loop through the aforementioned coupling mechanism: blockchain data → AI identification → intelligent matching → route optimization → transportation monitoring → blockchain recording, creating a positive data flow. The accuracy of AI identification is verified through blockchain data, the rationality of intelligent matching is adjusted through time-series prediction, and the safety of route optimization is evaluated through transportation monitoring, forming a reverse feedback. Each module dynamically adjusts its parameters based on the feedback results: the AI ​​identification module adjusts the confidence threshold, the intelligent matching module adjusts the objective function weights, and the route optimization module adjusts the reward function coefficients, achieving adaptive optimization of the system.

[0078] Reference Figure 7 This embodiment provides a blockchain-based intelligent management and control method for hazardous waste across the entire supply chain, including the following steps: Step S1: Establish a multi-party collaborative verification network based on the consortium blockchain architecture, involving hazardous waste generating enterprises, transportation enterprises, disposal enterprises, and regulatory authorities, and automatically execute transfer manifest review and disposal process supervision through smart contracts.

[0079] Specifically, a Hyperledger Fabric consortium blockchain network is deployed, with nodes representing waste-generating enterprises, transportation enterprises, disposal enterprises, and regulatory authorities as consensus nodes. Smart contracts are defined, including rules for reviewing transfer manifests, rules for supervising the disposal process, and rules for multi-party signature verification. When hazardous waste is generated, the waste-generating enterprise creates a transfer manifest application, filling in information such as the type and quantity of hazardous waste, the time of generation, the transportation enterprise, and the disposal enterprise, and submits it to the consortium blockchain. The smart contract automatically checks the completeness and compliance of the information, such as whether the transportation and disposal enterprises have the corresponding qualifications and whether the disposal enterprise has remaining disposal capacity. If the check passes, a transfer manifest number is generated and uploaded to the blockchain; if the check fails, an error message is returned.

[0080] Step S2: Based on the improved ResNet network, the packaging, labeling and appearance of hazardous waste are identified and classified in real time. The improved ResNet network adopts a dual attention fusion structure of channel attention mechanism and spatial attention mechanism, and feeds the identification results back to the consortium blockchain for cross-validation.

[0081] Specifically, high-resolution images of hazardous waste packaging, labeling, and appearance are captured using industrial cameras and input into an improved ResNet50 network for identification. The network extracts depth features from the images, highlights key features through a dual-attention fusion structure, and outputs the hazardous waste category and confidence score. The identification result is compared with the hazardous waste category declared by the waste-generating enterprise. If they match, it is marked as normal; if they do not match, an alert is triggered and recorded on the blockchain, while simultaneously notifying regulatory authorities to conduct on-site verification.

[0082] Step S3: The knowledge graph-based hazardous waste matching and recommendation algorithm automatically matches the optimal disposal facility. The matching degree score between hazardous waste characteristics and disposal process is calculated through graph reasoning. The optimal balance between disposal cost, transportation distance and disposal capacity is found based on a multi-objective optimization algorithm.

[0083] Specifically, based on the hazardous waste category, suitable treatment processes are searched in a knowledge graph. A graph neural network is used to calculate the matching score between hazardous waste characteristics and treatment processes, and processes with high matching scores are selected. Based on the treatment processes, qualified treatment facilities are identified, forming a candidate set. The non-dominated sorting genetic algorithm NSGA-II is applied to this candidate set to optimize three objective functions: treatment cost, transportation distance, and treatment capacity, yielding a Pareto optimal solution set. Based on the decision-maker's preferences, one solution is selected as the recommended solution, and the recommended treatment facility information is fed back to the waste-generating company and the transportation company.

[0084] Step S4: Based on the Transformer encoder-LSTM decoder hybrid architecture, predict the amount of hazardous waste generated, and feed the prediction results back to the intelligent matching recommendation module to adjust the disposal capacity plan.

[0085] Specifically, the system collects the enterprise's production plans, raw material procurement volumes, and process parameters from the past 24 months. These data are then input into a Transformer encoder for feature encoding, extracting global dependencies among multi-dimensional temporal features. The encoded features are then input into an LSTM decoder for sequence decoding to predict the hazardous waste generation volume for the next 12 months. Residual connections and attention weighting mechanisms are used to adaptively correct prediction errors. The prediction results are fed back to an intelligent matching and recommendation module to assess the adequacy of disposal facilities' capacity. If the predicted generation volume exceeds the remaining disposal capacity, the disposal plan is adjusted in advance or other disposal facilities are recommended.

[0086] Step S5: Dynamically plan the hazardous waste transportation route based on the reinforcement learning algorithm of Deep Q-Network (DQN), and return the optimized route and transportation data to the blockchain to complete the closed-loop traceability of the entire chain.

[0087] Specifically, a state space is constructed, including the current location, road congestion index, distance to environmentally sensitive areas, distribution of emergency rescue stations, and type of hazardous waste. A deep Q-network is used to select the next path node based on the current state, comprehensively considering transportation time, safety, and cost. This process is repeated until the destination (disposal facility) is reached, generating a complete transportation route. During transportation, real-time monitoring data such as GPS location, speed, temperature, and humidity of the transport vehicles is acquired and stored on the blockchain, forming a complete transportation trajectory record. The blockchain compares the predetermined route with the actual route, triggering an alert if a deviation is detected. After disposal, the disposal company uploads a disposal report, requiring signatures from the waste-generating company, transportation company, disposal company, and regulatory department for full-chain closed-loop traceability.

[0088] The method of this invention achieves intelligent management and control of hazardous waste from generation to disposal through five steps. Each step corresponds to the function of a core module, and the steps are deeply coupled through data transmission and feedback mechanisms to achieve system-level collaborative optimization, significantly improving the intelligence level, data reliability and safety of hazardous waste management.

[0089] In summary, the intelligent management system and method for the entire chain of hazardous waste based on blockchain traceability provided by this invention achieves multi-party collaborative verification and data immutability through blockchain technology, high-precision automatic identification of hazardous waste through an improved ResNet network, intelligent matching of disposal facilities through knowledge graphs, accurate prediction of hazardous waste generation through a Transformer-LSTM hybrid model, and dynamic optimization of transportation routes through deep reinforcement learning. These five core modules form a closed-loop collaborative system through deep coupling, solving problems such as regulatory blind spots, information silos, low identification accuracy, unreasonable matching, lack of prediction, and insufficient route optimization in existing technologies. This significantly improves the intelligence level, data credibility, disposal efficiency, and safety of hazardous waste management, and has significant application value and promising prospects for promotion.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A blockchain-based intelligent management and control system for the entire chain of hazardous waste, characterized in that: include: The blockchain traceability module is used to establish a multi-party collaborative verification network based on the consortium blockchain architecture, involving hazardous waste generating enterprises, transportation enterprises, disposal enterprises and regulatory departments. Through smart contracts, it automatically executes the review of transfer manifests and the supervision of disposal processes. It stores IoT-collected data and AI recognition results from each stage of hazardous waste generation, transfer and disposal on the blockchain to achieve data immutability and full traceability. The AI ​​identification and classification module is connected to the blockchain traceability module and is used to identify and classify hazardous waste packaging, labels and appearance in real time based on an improved ResNet network. The improved ResNet network adopts a dual attention fusion structure of channel attention mechanism and spatial attention mechanism, and feeds the identification results back to the blockchain traceability module for cross-validation. When the identification result is inconsistent with the declaration information, an early warning is triggered and recorded to the blockchain. The intelligent matching and recommendation module is connected to the blockchain traceability module and the AI ​​identification and classification module. It is used to automatically match the optimal disposal facility based on the knowledge graph-based hazardous waste matching and recommendation algorithm. It calculates the matching degree score between hazardous waste characteristics and disposal process through graph reasoning, and finds the optimal balance between disposal cost, transportation distance and disposal capacity based on multi-objective optimization algorithm. The matching result is output to the blockchain traceability module for recording and transmitted to the path optimization module. The time-series prediction module, connected to the intelligent matching and recommendation module, is used to predict the amount of hazardous waste generated based on the Transformer encoder-LSTM decoder hybrid architecture, feed the prediction results back to the intelligent matching and recommendation module to adjust the disposal capacity planning, and achieve adaptive correction of prediction errors through residual connection and attention weight mechanism. The path optimization module, connected to the intelligent matching and recommendation module and the blockchain traceability module, is used to dynamically plan hazardous waste transportation routes based on the reinforcement learning algorithm of deep Q-network (DQN). By constructing a state space, action space, and reward function, it improves the stability of path planning by using experience replay and target network mechanisms, and returns the optimized path and transportation data to the blockchain traceability module to complete the closed-loop traceability of the entire chain.

2. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The blockchain traceability module includes: The consortium blockchain network unit is used to build a consortium blockchain network architecture based on Hyperledger Fabric, and sets up nodes of waste-generating enterprises, transportation enterprises, disposal enterprises, and regulatory departments as consensus nodes. The smart contract execution unit, connected to the consortium blockchain network unit, is used to define rules for transferring transfer slips, rules for monitoring the processing flow, and rules for multi-party signature verification. The data upload unit, connected to the smart contract execution unit, is used to upload and store RFID tag data, smart weighing data, video surveillance data, and AI recognition results on the blockchain after hash encryption.

3. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The AI ​​recognition and classification module includes: Image acquisition unit, used to acquire high-resolution images of hazardous waste packaging, labeling and appearance using an industrial camera; The feature extraction unit, connected to the image acquisition unit, is used to extract the depth features of the hazardous waste image based on the improved ResNet50 network. The improved ResNet50 network introduces a channel attention submodule and a spatial attention submodule after the residual block, and improves the attention to the area where the hazardous waste packaging and label are located through adaptive feature recalibration. The classification decision unit, connected to the feature extraction unit, is used to output the probability distribution of hazardous waste categories based on the fully connected layer and the Softmax activation function, and to set a confidence threshold for classification decision.

4. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The intelligent matching and recommendation module includes: The knowledge graph construction unit is used to construct a knowledge graph containing entities of hazardous waste characteristics, disposal processes, disposal facilities, and geographical locations, and to define the applicable relationships, qualification relationships, and geographical relationships between entities. The graph reasoning unit, connected to the knowledge graph construction unit, is used to perform knowledge graph embedding representation learning based on graph neural network (GNN), and to calculate the matching degree score between hazardous waste characteristics and disposal process through node feature aggregation and relationship modeling. The multi-objective optimization unit, connected to the graph reasoning unit, is used to find the Pareto optimal solution set among the disposal cost objective function, the transportation distance objective function, and the disposal capacity objective function based on the non-dominated sorting genetic algorithm NSGA-II.

5. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The time series prediction module includes: The feature encoding unit is used to perform multi-head self-attention encoding of enterprise production plans, raw material procurement quantities and process parameters based on the Transformer encoder, and to extract the global dependencies of multi-dimensional temporal features through position encoding and feedforward neural networks. A sequence decoding unit, connected to the feature encoding unit, is used to perform sequence decoding on the encoded features output by the feature encoding unit based on the LSTM decoder, and to capture the long-term dependency relationship of hazardous waste generation through forget gate, input gate and output gate mechanism; An error correction unit, connected to the sequence decoding unit, is used to add the encoded features output by the feature encoding unit to the decoded output of the sequence decoding unit through residual connection, and to adaptively correct the prediction error based on an attention weight mechanism.

6. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The path optimization module includes: State space construction unit, used to construct a state vector containing current location, road congestion index, distance to environmentally sensitive areas, distribution of emergency rescue stations and hazardous waste type; The DQN network unit, connected to the state space construction unit, is used to construct a deep Q-network, extract state features through convolutional layers, and output Q-value estimates for each action through fully connected layers. The experience replay unit, connected to the DQN network unit, is used to store state transition samples consisting of the current state, the action performed, the immediate reward, and the next state. It breaks sample correlation through mini-batch random sampling and improves training stability by adopting a target network periodic update mechanism.

7. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The dual attention fusion structure is implemented as follows: First, the feature map output by the ResNet network in the AI ​​recognition and classification module is subjected to global average pooling and global max pooling to generate two sets of channel descriptors. Then, the two sets of descriptors are input into a shared multilayer perceptron to generate channel attention weights. Next, the weighted feature map is subjected to average pooling and max pooling in the channel dimension to generate spatial descriptors. Finally, spatial attention weights are generated through convolutional layers and multiplied with the channel-weighted feature map to obtain the final attention-enhanced features.

8. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The multi-objective optimization algorithm is implemented as follows: First, the population is initialized and the three objective function values ​​of disposal cost, transportation distance and disposal capacity for each individual are calculated. Then, individuals are assigned to different non-dominated levels based on fast non-dominated sorting. Next, the crowding distance of each individual is calculated to maintain the diversity of solutions. Finally, a new population is generated through tournament selection, simulated binary crossover and polynomial mutation. The population is iteratively updated until convergence to obtain the Pareto optimal solution set.

9. The intelligent management and control system for hazardous waste based on blockchain traceability according to claim 1, characterized in that, The Deep Q-Network (DQN) achieves path optimization in the following way: First, it selects an action based on the current state using an ε-greedy strategy, then executes the action and observes the next state and immediate reward. Next, it stores the state transition samples in the experience replay pool. Finally, it randomly samples a small batch of samples from the experience replay pool and updates the network parameters by minimizing the mean squared error loss function between the current Q-Network output and the target Q-value calculated by the target Q-Network. The target Q-Network synchronizes the parameters of the current Q-Network every fixed number of steps.

10. A method for intelligent management and control of hazardous waste across the entire chain based on blockchain traceability, characterized in that: Includes the following steps: Step S1: Establish a multi-party collaborative verification network based on the consortium blockchain architecture, involving hazardous waste generating enterprises, transportation enterprises, disposal enterprises, and regulatory authorities, and automatically execute transfer manifest review and disposal process supervision through smart contracts; Step S2: Based on the improved ResNet network, the packaging, labeling and appearance of hazardous waste are identified and classified in real time. The improved ResNet network adopts a dual attention fusion structure of channel attention mechanism and spatial attention mechanism, and feeds the identification results back to the consortium blockchain for cross-validation. Step S3: The knowledge graph-based hazardous waste matching and recommendation algorithm automatically matches the optimal disposal facility, calculates the matching degree score between hazardous waste characteristics and disposal process through graph reasoning, and finds the optimal balance between disposal cost, transportation distance and disposal capacity based on a multi-objective optimization algorithm. Step S4: Based on the Transformer encoder-LSTM decoder hybrid architecture, predict the amount of hazardous waste generated, and feed the prediction results back to step S3 to adjust the disposal capacity plan. Step S5: Dynamically plan the hazardous waste transportation route based on the reinforcement learning algorithm of Deep Q-Network (DQN), and return the optimized route and transportation data to the blockchain to complete the closed-loop traceability of the entire chain.

Images