A wireless communication-based remote control method for waste sorting equipment

Abstract

The application discloses a kind of waste sorting equipment remote control methods based on wireless communication, it is related to industrial internet of things intelligent control technical field, including, remote control center receives sorting task work order, obtains optimal resource allocation strategy by multi-agent reinforcement learning framework, and generates task resource demand list using stochastic optimization algorithm;Based on task resource demand list, remote control center generates network slice instantiation request and sends to mobile network operator, obtains the access identifier and configuration parameter of special network slice matched with sorting task;Remote control center issues the access identifier and configuration parameter of special network slice to waste sorting equipment to execute sorting task, obtains the network environment access state customized for sorting task.The application realizes the accurate matching of sorting task demand and network computing resources, solves the cross-domain resource collaborative optimization problem existing in traditional static resource configuration, and reduces resource allocation error rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial internet of things intelligent control, and in particular to a waste sorting equipment remote control method based on wireless communication. BACKGROUND

[0002] In recent years, industrial internet of things and intelligent control technology have made significant progress in the field of waste sorting, promoting the transformation from traditional mechanical processing to intelligent remote management. In terms of visual analysis, object detection algorithms based on deep learning (such as YOLOv5, Faster R-CNN) have been able to realize real-time classification of waste materials, with an accuracy of over 90%, but most solutions still rely on local computing resources, making it difficult to process high-concurrency video stream data. The evolution of wireless communication technology provides a new path for remote control, and the commercialization of 5G network slicing technology makes differentiated quality of service possible. Existing research transmits control instructions through uRLLC (ultra-reliable low-latency communication) slices and processes image data through eMBB (enhanced mobile broadband) slices. In terms of decision optimization, reinforcement learning algorithms are introduced into the sorting process, but single-agent models have the problems of slow policy convergence and poor adaptability in complex dynamic environments.

[0003] Despite the continuous progress of related technologies, there are still two key deficiencies: first, in terms of resource allocation, current methods lack global optimization capabilities across network, computing, and control domains. Traditional solutions usually independently design communication slice strategies and visual analysis processes, resulting in low resource utilization - for example, when sorting equipment processes multiple types of waste simultaneously, fixed slice configurations cannot adapt to sudden traffic changes, causing transmission delays of high-priority data (such as control instructions) to fluctuate by more than 30 milliseconds, while low-priority data (such as historical logs) occupies redundant bandwidth. Second, in terms of decision adaptability, existing remote control methods have limited response capabilities to dynamic environments. Most solutions use preset thresholds to trigger transmission modes, without considering the real-time coupling relationship between network state and data characteristics. When wireless channel quality changes suddenly, a single threshold mechanism can easily lead to misjudgment, such as high-confidence data being lost due to network congestion, or low-confidence images being transmitted with high slice load. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a waste sorting equipment remote control method based on wireless communication to solve the problems of rigid resource allocation and insufficient decision adaptability in the prior art.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The application provides a waste sorting equipment remote control method based on wireless communication, which comprises the following steps of: a remote control center receiving a sorting task work order, acquiring an optimal resource allocation strategy through a multi-agent reinforcement learning framework, and generating a task resource demand list by using a random optimization algorithm; based on the task resource demand list, the remote control center generates a network slice instantiation request and sends it to a mobile network operator to obtain an access identifier and configuration parameters of a special network slice matched with the sorting task; the remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment to perform the sorting task, and obtains a network environment access state customized for the sorting task; based on the network environment access state, the sorting equipment starts a visual recognition function, and an integrated edge artificial intelligence model is used to perform real-time visual analysis on the waste to generate structured data; the sorting equipment performs real-time analysis on the structured data through a built-in communication decision maker, and when a high confidence condition is met, the structured data is transmitted through a low-latency slice, otherwise, alarm signaling and key image data are synchronously sent through a composite transmission mode, and are sent to the remote control center through a corresponding slice channel; the remote control center fuses the structured data, the alarm signaling and the key image data to generate a control instruction, which is issued to the sorting equipment to perform a sorting action, and the execution state of the sorting equipment is fed back to the remote control center.

[0008] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, the remote control center receives the sorting task work order, and acquires the optimal resource allocation strategy through the multi-agent reinforcement learning framework, and the specific steps are as follows,

[0009] The remote control center receives the sorting task work order and uniformly converts it into a high-dimensional feature vector to establish a work order feature library.

[0010] Based on the work order feature library, a preliminary control strategy is generated through a strategy network of the multi-agent reinforcement learning framework, and a state evaluation quantitative index set is generated through a value network for state evaluation.

[0011] The preliminary control strategy and the state evaluation quantitative index set are coordinated for resource allocation to obtain the optimal resource allocation strategy.

[0012] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, the remote control center receives the sorting task work order, and acquires the optimal resource allocation strategy through the multi-agent reinforcement learning framework, and the specific steps are as follows,

[0013] The network service quality parameters are extracted from the optimal resource allocation strategy, and a dynamic evolution framework is constructed by combining historical operation data to predict the dynamic change trend of the service quality in the future time period.

[0014] Based on the dynamic change trend of service quality in the future time period, the minimum expected total cost problem of resource constraint condition is obtained by using a random optimization algorithm, and a task resource demand list is generated.

[0015] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, wherein: based on the task resource demand list, the remote control center generates a network slice instantiation request and sends it to the mobile network operator, and the specific steps are as follows,

[0016] The remote control center generates a network slice instantiation request based on the network service quality parameters and resource constraints in the task resource demand list.

[0017] The network slice instantiation request is sent to the core network control unit of the mobile network operator through a secure communication protocol.

[0018] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, wherein: the access identifier and configuration parameters of the special network slice matched with the sorting task are obtained, and the specific steps are as follows,

[0019] The core network control unit of the mobile network operator performs slice resource allocation on the network slice instantiation request, and dynamically creates a special network slice matched with the sorting task demand;

[0020] The remote control center obtains the access identifier and configuration parameters from the special network slice through an authorized interface.

[0021] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, wherein: the remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment to perform the sorting task, and obtains the network environment access state customized for the sorting task, and the specific steps are as follows,

[0022] The remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment through an encrypted communication channel;

[0023] The waste sorting equipment initializes the network connection and accesses the special network slice according to the access identifier and configuration parameters;

[0024] After accessing the special network slice, the waste sorting equipment performs the sorting task and transmits the sorting operation data stream in real time through the special network slice;

[0025] The sorting operation data stream is captured in real time through an integrated monitoring interface, and performance analysis and state evaluation are performed to generate network environment access state information customized for the sorting task.

[0026] As a preferred scheme of the waste sorting device remote control method based on wireless communication according to the application, wherein: based on the network environment access state, the sorting device starts the visual recognition function, and the integrated edge artificial intelligence model is used for real-time visual analysis of the waste to generate structured data, and the specific steps are as follows,

[0027] The sorting device generates a visual recognition start signal based on the network environment access state;

[0028] Based on the visual recognition start signal, the visual recognition function is started and the integrated edge artificial intelligence model is initialized;

[0029] According to the edge artificial intelligence model, the sorting device performs real-time visual analysis on the waste to generate a category probability distribution;

[0030] The confidence score is obtained by calculating the maximum value in the category probability distribution, and the structured data is generated.

[0031] As a preferred scheme of the waste sorting device remote control method based on wireless communication according to the application, wherein: the sorting device analyzes the structured data in real time through the built-in communication decision maker, and when the high confidence condition is met, the structured data is transmitted through the low latency slice transmission structure, otherwise the composite transmission mode is triggered to send alarm signaling and key image data synchronously, and sent to the remote control center through the corresponding slice channel, and the specific steps are as follows,

[0032] The sorting device analyzes the structured data through the built-in communication decision maker to obtain the confidence score and the timestamp;

[0033] According to the timestamp, the sorting device obtains the historical confidence sequence, the network state index and the data importance sequence through the communication decision maker;

[0034] Based on the historical confidence sequence, the network state index and the data importance sequence, the integral-period formula is used to calculate the dynamic threshold value;

[0035] Compare the confidence score with the dynamic threshold value to determine the transmission mode;

[0036] When the transmission mode decision is high confidence transmission, the communication decision maker triggers the high confidence transmission mode, and uploads the structured data to the remote control center through the ultra-high reliability low latency communication slice channel;

[0037] When the transmission mode decision is low confidence transmission, the communication decision maker triggers the composite transmission mode, sends the alarm signaling through the low latency slice, and uploads the associated key image data to the remote control center through the large bandwidth slice.

[0038] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, wherein: the remote control center fuses structured data, alarm signaling and key image data to generate a control instruction set, and the specific steps are as follows,

[0039] Based on structured data, alarm signaling and key image data, the remote control center performs data alignment and noise filtering, and generates a unified environment state estimation vector by using a weighted fusion algorithm.

[0040] Based on the environment state estimation vector, the optimal control action vector is calculated through a reinforcement learning model to generate a control instruction set.

[0041] As a preferred scheme of the waste sorting equipment remote control method based on wireless communication, wherein: the sorting action is executed by the sorting equipment through a dedicated network slice, and the execution state of the sorting equipment is fed back to the remote control center,

[0042] The remote control center encapsulates the control instruction set into a protocol data unit and transmits it to the sorting equipment through a secure downlink channel of the dedicated network slice.

[0043] After receiving and analyzing the control instruction set, the sorting equipment executes the sorting action and feeds back the sorting result and device state parameters to the remote control center.

[0044] The present application has the advantages that: through the dynamic policy optimization mechanism of the multi-agent reinforcement learning framework, the precise matching of sorting task requirements and network computing resources is realized, the cross-domain resource collaborative optimization problem existing in the traditional static resource configuration is solved, and the resource allocation error rate is reduced; the dynamic decision threshold is calculated through the integral-period formula, the intelligent switching of the data transmission mode is realized, and the defect of slow response of the fixed threshold mechanism in the complex network environment is solved. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0046] Fig. 1 The flowchart of the waste sorting equipment remote control method based on wireless communication.

[0047] Fig. 2 The flowchart for obtaining the optimal resource allocation strategy.

[0048] Fig. 3 The flowchart of network slice instantiation request and processing.

[0049] Fig. 4 Flow chart for communicating decisions to a decision maker. DETAILED DESCRIPTION

[0050] In order to make the above objectives, features and advantages of the present application more clear and comprehensible, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0051] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details given herein, that the present application can be practiced with other than the described embodiments and that variations from the particular embodiments described herein can be made and still be within the scope of the present application.

[0052] Secondly, the "one embodiment" or "an embodiment" referred to herein means that a specific feature, structure or characteristic described can be included in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not all refer to the same embodiment, nor is it an alternative or exclusive embodiment to other embodiments.

[0053] Reference Signs List Figs. 1-4 For one embodiment of the present application, the embodiment provides a wireless communication-based waste sorting equipment remote control method, comprising the following steps:

[0054] S1, the remote control center receives the sorting task work order, obtains the optimal resource allocation strategy through the multi-agent reinforcement learning framework, and generates the task resource demand list by using the random optimization algorithm.

[0055] The remote control center receives the sorting task work order and uniformly converts it into a high-dimensional feature vector to establish a work order feature library.

[0056] The specific process includes that after the remote control center receives the sorting task work order, the multi-dimensional information contained in each sorting task work order, such as task type, waste type, processing priority, geographical location, time constraint and equipment demand, is structurally parsed, and the multi-dimensional information is converted into a high-dimensional feature vector according to the feature coding rule. Each dimension of the high-dimensional feature vector corresponds to a specific attribute or its combination in the work order, so as to completely retain the semantics and operation meaning of the sorting task work order in the vector space; all the high-dimensional feature vectors converted are sequentially stored according to time sequence or task category to form a work order feature library, providing input basis for the multi-agent reinforcement learning framework to support the generation of the optimal resource allocation strategy.

[0057] It should be noted that the feature coding rule refers to converting various attributes in the sorting task order into a numerical vector dimension according to a unified mapping method. The feature coding rule is defined and fixed before the method is implemented, which is used to ensure that different orders can be consistently converted into high-dimensional feature vectors.

[0058] Based on the order feature library, a preliminary control strategy is generated by the policy network of the multi-agent reinforcement learning framework, and a state evaluation quantitative index set is generated by using the value network.

[0059] The specific process includes: based on the order feature library, the policy network of the multi-agent reinforcement learning framework inputs the high-dimensional feature vector in the order feature library as the current environment state, the policy network performs layer-by-layer nonlinear transformation on the input high-dimensional feature vector according to the internal neural network weight, and finally generates an action probability distribution or a deterministic action vector in the output layer. This output is the preliminary control strategy, and the preliminary control strategy includes initial decisions for current sorting tasks in terms of computing resource scheduling, communication resource configuration, and sorting device execution actions; the value network of the multi-agent reinforcement learning framework receives the same high-dimensional feature vector, obtains the expected cumulative return estimate value corresponding to the state represented by the current input high-dimensional feature vector through its own neural network structure, forms a plurality of dimensional numerical results, and constitutes a state evaluation quantitative index set.

[0060] The training process of the multi-agent reinforcement learning framework: taking the high-dimensional feature vector in the order feature library as the environment state input, in each training time step, the policy network outputs an action according to the current high-dimensional feature vector, which corresponds to a resource allocation or device control decision. The environment feeds back a new high-dimensional feature vector and the corresponding immediate reward after executing the action. The value network estimates the expected cumulative return based on the new high-dimensional feature vector, and compares it with the actual cumulative reward to obtain the time difference error. The time difference error is used to update the neural network weights of the policy network and the value network synchronously; multiple agents interact in the shared or independent experience replay buffer to sample historical trajectories, and the policy network parameters are optimized by the policy gradient method while minimizing the mean square Bellman error of the value network to improve the state value estimation accuracy. After multiple rounds of iterative training, the policy network gradually converges to a stable and effective preliminary control strategy, and the value network simultaneously forms an accurate state evaluation quantitative index set.

[0061] It should be noted that the immediate reward refers to a numerical feedback obtained by comprehensively considering the resource utilization rate, task completion timeliness, and sorting accuracy during the execution of the sorting task.

[0062] The preliminary control strategy and the state evaluation quantitative index set are coordinated for resource allocation to obtain an optimal resource allocation strategy.

[0063] The specific process includes, when coordinating resource allocation for the preliminary control strategy and the state evaluation quantitative index set, jointly comparing and comprehensively evaluating the initial decision in the preliminary control strategy about the scheduling of computing resources, the configuration of communication resources and the action of sorting equipment with the multiple dimension values in the state evaluation quantitative index set representing the degree of advantage and disadvantage of the current task state, dynamically adjusting the allocation proportion of different sorting tasks on computing resources, communication resources and equipment actions according to the optimization mechanism inside the multi-agent reinforcement learning framework, so that the adjusted resource allocation scheme not only meets the actual needs of the sorting task order in terms of delay, bandwidth, reliability and equipment capacity, but also obtains higher cumulative return in long-term operation, and finally outputs the optimal resource allocation strategy that meets the re-target.

[0064] It should be noted that the resource scheduling suggestion refers to the initial allocation scheme of computing resources, communication resources and sorting equipment actions in the current sorting task in the preliminary control strategy.

[0065] The network service quality parameters are extracted from the optimal resource allocation strategy, and a dynamic evolution framework is constructed combined with historical operation data to predict the dynamic change trend of service quality in the future time period.

[0066] The specific process includes extracting network service quality parameters from the optimal resource allocation strategy, which includes end-to-end delay requirement, required bandwidth capacity, connection reliability level and data packet loss rate tolerance; aligning the network service quality parameters with the historical operation data in the time dimension, which includes the network load, channel state, slice usage and task completion quality observed in the past sorting task execution process; on this basis, a dynamic evolution framework is constructed, and the specific construction method is: taking the network service quality parameters at each time step as the state variable, using the observed values of consecutive time steps in the historical operation data to establish the state transition equation, which describes how the network service quality parameters at the current time are determined by the parameter values at the previous time and external disturbance factors; fitting the coefficients in the state transition equation by using existing parameter identification methods such as least squares method or maximum likelihood estimation; in the process of continuously arriving new observation data, recursively update these coefficients to make the dynamic evolution framework adaptively reflect the influence of task load change and external environmental disturbance on network service quality parameters, and finally realize the prediction of the dynamic change trend of service quality in the future time period.

[0067] It should be noted that the historical operation data is the data obtained by monitoring and recording in the past sorting task execution process, including network load, channel state, slice usage and task completion quality.

[0068] The recursive updating mechanism refers to using the current new observation value and the existing estimation result to correct the parameters of the state transition equation through Kalman filtering or exponential weighting method in a dynamic evolution framework, which can continuously track the changes of network service quality parameters and adapt to environmental disturbances.

[0069] Based on the dynamic change trend of service quality in the future time period, a random optimization algorithm is used to obtain the minimum expected total cost problem of resource constraint conditions, and a task resource demand list is generated.

[0070] The specific process includes: based on the dynamic change trend of service quality in the future time period, the time-varying demand of network resources and uncertainty are modeled as random variables, and combining with multi-dimensional resource constraints such as computing resources, communication resources and device availability, an optimization problem is constructed with the expected total cost as the objective function, which comprehensively considers the resource rental cost, task delay penalty and service default risk; on this basis, a random optimization algorithm is used to solve the optimization problem, and through iterative approximation, a resource allocation scheme that meets all resource constraints and minimizes the expected total cost is generated, and a task resource demand list is generated.

[0071] It should be noted that the random optimization algorithm is a mathematical method for solving optimization problems with random variables or uncertain factors in the objective function, the core idea is that when there is probability distribution or noise interference in the objective function or constraint condition, through sampling, expectation approximation or scenario analysis, the decision variable that optimizes the expected performance index is found; the random optimization algorithm is used to solve the problem of minimizing the expected total cost under the resource constraint condition, and a task resource demand list is generated.

[0072] S2, based on the task resource demand list, the remote control center generates a network slice instantiation request and sends it to the mobile network operator, and obtains the access identifier and configuration parameters of the special network slice matched with the sorting task.

[0073] The remote control center generates a network slice instantiation request of slice configuration demand based on the network service quality parameters and resource constraints in the task resource demand list.

[0074] The specific process includes that the remote control center structures and arranges each parameter according to the network service quality parameters and resource constraint conditions contained in the task resource demand list, wherein the network service quality parameters include end-to-end delay requirement, required bandwidth capacity, connection reliability level and data packet loss rate tolerance, the resource constraint conditions cover the restrictive indexes such as available computing unit quantity, maximum concurrent connection number and upper limit of energy consumption, and are encoded according to the interface specification defined by the mobile network operator to form the slice configuration requirements of the network slice in aspects of isolation level, service level agreement, resource reservation granularity and life cycle; the remote control center encapsulates these slice configuration requirements according to the standard communication protocol format to generate a complete network slice instantiation request for applying to the mobile network operator to create a dedicated network slice matched with the sorting task.

[0075] The network slice instantiation request is sent to the core network control unit of the mobile network operator through a secure communication protocol.

[0076] The specific process includes that the remote control center encapsulates the network slice instantiation request according to the encryption and authentication mechanism specified by the secure communication protocol to ensure that the content of the network slice instantiation request has confidentiality, integrity and identity verifiability in the transmission process, and sends the encapsulated network slice instantiation request to the core network control unit of the mobile network operator through the wireless communication network, which receives and parses the network slice instantiation request to perform subsequent slice resource allocation operations.

[0077] The core network control unit of the mobile network operator performs slice resource allocation on the network slice instantiation request to dynamically create a dedicated network slice matched with the sorting task demand.

[0078] The specific process includes that after receiving the network slice instantiation request, the core network control unit of the mobile network operator parses the slice configuration requirements contained therein, and performs slice resource allocation on the physical network infrastructure according to the network service quality parameters and resource constraint conditions specified in the network slice instantiation request to dynamically divide and configure independent computing, storage and communication resources (for example, reserving a dedicated virtual machine or container in an edge data center for processing visual data uploaded by sorting equipment, allocating exclusive radio resource blocks on the 5G base station side to guarantee uplink bandwidth, establishing a low-latency forwarding path in the transport network and isolating other business traffic) for the sorting task, and simultaneously invokes the network function virtualization management platform according to the network slice orchestrator to deploy user plane function, control plane function and policy control unit on demand, and configures end-to-end virtual link and forwarding rule through the software-defined network controller to set quality of service policy, security policy and life cycle management parameters, thereby establishing a virtual network topology and policy control rule matched with the sorting task demand to create a dedicated network slice in real time.

[0079] The remote control center obtains the access identifier and configuration parameters from the special network slice through the authorization interface.

[0080] The specific process includes that the remote control center initiates an access request to the special network slice through the authorization interface, the authorization interface ensures the legality of access based on the identity authentication and permission verification mechanism, and after the authentication is passed, the remote control center obtains the access identifier and configuration parameters for device access from the special network slice, wherein the access identifier is used to uniquely identify the special network slice, and the configuration parameters include network address, security key, service quality policy and connection endpoint information.

[0081] S3, the remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment to perform the sorting task, and obtains the network environment access state customized for the sorting task.

[0082] The remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment through an encrypted communication channel.

[0083] The specific process includes that the remote control center encrypts the access identifier and configuration parameters of the special network slice according to the encryption algorithm (such as AES-256 or SM4) and key (such as session key dynamically distributed by the key management server or shared symmetric key based on device identity) adopted by the encrypted communication channel, forms a ciphertext data packet, and transmits the ciphertext data packet to the waste sorting equipment through the encrypted communication channel, so as to ensure that the access identifier and configuration parameters are not stolen, tampered or forged during transmission, and enable the waste sorting equipment to safely receive the necessary information for subsequent network connection initialization.

[0084] The waste sorting equipment initializes the network connection and accesses the special network slice according to the access identifier and configuration parameters.

[0085] The specific process includes that after the waste sorting equipment receives the access identifier and configuration parameters of the special network slice, it identifies the target special network slice according to the access identifier, and performs parameter setting and authentication process of the network protocol stack according to the network address, security key, service quality policy and connection endpoint information contained in the configuration parameters, completes the activation of the communication interface and the establishment of the link, so as to initialize the network connection and successfully access the special network slice.

[0086] After the waste sorting equipment accesses the special network slice, it performs the sorting task and transmits the sorting operation data stream in real time through the special network slice.

[0087] The specific process includes that after the waste sorting equipment accesses the special network slice, the execution process of the sorting task is started, the features of the real-time collected waste images are extracted and semantically analyzed, the prediction probabilities of various categories are generated and the belonging categories are determined, and the recognition and classification are completed; according to the classification result, the mechanical arm or the sorting execution mechanism is driven to perform the corresponding physical action, such as grabbing, pushing or putting into the designated recycling channel, and the sorting operation data stream related to the sorting action, the equipment state and the environmental perception is continuously generated during the execution process, and the sorting operation data stream is uploaded to the remote control center in real time through the established special network slice.

[0088] The sorting task customized network environment access state information is generated by integrating the monitoring interface to capture the sorting operation data stream in real time and performing performance analysis and state evaluation.

[0089] The specific process includes that the sorting operation data stream uploaded by the waste sorting equipment in the special network slice is continuously monitored and captured in real time through the integrated monitoring interface, the sorting operation data stream includes device action instruction execution timestamp, visual recognition result transmission delay, control feedback interval, data packet serial number and link layer state information, and the performance analysis of the sorting operation data stream is performed to obtain network performance indicators such as end-to-end transmission delay, effective throughput, data packet loss rate and connection jitter, and the state evaluation is performed in combination with the requirements of the sorting task on the quality of communication service to judge whether the current network connection meets the reliability and timeliness requirements of task execution (for example: if it is a high-precision metal sorting task, the end-to-end delay is required to be less than 10 milliseconds and the packet loss rate is less than 0.1%; if it is a general plastic classification task, 20 milliseconds of delay and 1% of packet loss rate can be tolerated), and the performance analysis result and the state evaluation conclusion are fused to generate the sorting task customized network environment access state information which can accurately describe the adaptation degree of the current communication environment.

[0090] S4, based on the network environment access state, the sorting equipment starts the visual recognition function, and the integrated edge artificial intelligence model performs real-time visual analysis on the waste to generate structured data.

[0091] The sorting equipment generates a visual recognition start signal based on the network environment access state.

[0092] The specific process includes that the sorting equipment determines whether the communication quality of the current special network slice meets the requirements of bandwidth, delay and reliability of the visual recognition function according to the network environment access state information customized for the sorting task, when it is confirmed that the network environment has the ability to stably transmit image data and model inference results, the sorting equipment triggers the internal logic to generate a visual recognition start signal.

[0093] Based on the visual recognition start signal, the visual recognition function is started and the integrated edge artificial intelligence model is initialized.

[0094] The specific process includes: based on the visual recognition start signal, the sorting device activates the visual recognition function carried, calls the built-in image acquisition unit to start acquiring waste image data, and synchronously loads the integrated edge artificial intelligence model to the local computing unit, completes the parameter loading, memory allocation and inference engine initialization of the integrated edge artificial intelligence model, makes the integrated edge artificial intelligence model enter the executable state, and prepares for subsequent real-time visual analysis of waste.

[0095] Further, the integrated edge artificial intelligence model adopts a lightweight convolutional neural network architecture, and its hierarchical structure includes an input layer, a plurality of convolution-batch normalization-activation function combination layers, a pooling layer, a global average pooling layer, and a fully connected output layer in sequence; the input layer receives the preprocessed waste image; the convolution layer extracts local spatial features through a learnable filter, each convolution layer is followed by a batch normalization layer to stabilize the training process and speed up convergence, and then a nonlinear activation function (such as ReLU) is used to enhance the model expression ability; the pooling layer performs down-sampling on the feature map in some stages to reduce the computational complexity and preserve the main semantic information; after multi-level convolution and pooling, the global average pooling layer compresses the spatial dimension to a channel statistic, eliminating the dependence on fixed input size; finally, the fully connected output layer maps high-dimensional features to a probability distribution of predefined waste categories; the layers are sequentially connected in a feedforward manner, with the output of the previous layer serving as the input of the next layer, and the overall structure is subjected to lightweight processing such as pruning, quantization and knowledge distillation to adapt to the edge computing resource constraints of the sorting device.

[0096] The training process of the edge artificial intelligence model: is completed in a remote control center or offline training environment, a large amount of labeled image data containing various types of waste is collected, an image data set for training is constructed, a deep learning method is used to train the convolutional neural network structure end-to-end, the prediction class probability distribution is obtained through forward propagation, and the difference between the prediction result and the true label is measured by using a loss function, and the network weights of the integrated edge artificial intelligence model are updated through a back propagation algorithm; during the training process, data augmentation, regularization and early stopping mechanism are combined to improve the generalization ability; after training, the integrated edge artificial intelligence model is subjected to lightweight processing, including pruning, quantization and knowledge distillation, and meets the local computing resource and real-time constraints of the sorting device; finally, the optimized integrated edge artificial intelligence model is deployed to the sorting device for executing real-time visual analysis tasks.

[0097] According to the edge artificial intelligence model, the sorting device performs real-time visual analysis on the waste to generate a class probability distribution.

[0098] The specific process includes that the sorting equipment calls the initialized integrated edge artificial intelligence model, inputs the real-time collected waste image into the integrated edge artificial intelligence model, the integrated edge artificial intelligence model performs layer-by-layer feature extraction and semantic reasoning on the image through internal convolution layers, pooling layers and fully connected layers, and finally obtains probability values of the waste belonging to each predefined category in the output layer. The probability values form a normalized vector, that is, a category probability distribution.

[0099] It should be noted that the categories corresponding to the probability values of the predefined categories are manually divided before model training according to the types of waste to be identified in the sorting task, such as metal, plastic, paper and glass, and are realized by labeling sample labels in the training data set.

[0100] The confidence score is obtained by calculating the maximum value in the category probability distribution, and the structured data is generated, and the expression is:

[0101] ;

[0102] Among them, represents the confidence score, represents the category probability distribution vector, represents the maximum value taken from the probability distribution vector , represents the entropy adjustment coefficient, represents the Shannon entropy function of the category probability distribution vector , represents the enhancement factor coefficient, represents the natural exponential function, represents the sharpness adjustment coefficient, represents the total number of categories in the classification task, represents the category index in the probability distribution vector , represents the predicted probability value of the th category.

[0103] The specific process includes that the sorting equipment selects the element with the maximum value from the category probability distribution as the basic confidence after obtaining the category probability distribution, and combines the Shannon entropy function of the category probability distribution to correct the uncertainty, and then applies nonlinear adjustment to the corrected result through the enhancement factor coefficient and the sharpness adjustment coefficient. Finally, the confidence score reflecting the recognition reliability is calculated, and the confidence score, together with the predicted label of the category and other related meta information, constitutes the structured data.

[0104] It should be noted that the entropy adjustment coefficient is an adjustable parameter for controlling the influence degree of the Shannon entropy on the confidence score, and the value is set and fixed according to offline verification before the method is implemented.

[0105] The enhancement factor coefficient is a parameter used to adjust the strength of the nonlinear correction of the confidence score. It is obtained through offline optimization on historical running data and remains fixed during execution.

[0106] S5. The sorting equipment performs real-time analysis of structured data through its built-in communication decision-maker. When the high confidence condition is met, the structured data is transmitted through low-latency slices. Otherwise, the composite transmission mode is triggered to synchronously send alarm signals and key image data, and then send them to the remote control center through the corresponding slice channel.

[0107] The sorting equipment parses structured data through a built-in communication decision-maker to obtain confidence scores and timestamps.

[0108] The specific process includes the sorting equipment parsing the structured data field by field through its built-in communication decision-maker, identifying and separating the confidence score and timestamp contained therein; the confidence score is a value derived from the category probability distribution, used to characterize the credibility of the current waste identification result, while the timestamp is recorded synchronously by the sorting equipment when generating the structured data, accurately identifying the time when the structured data was generated, and the two together serve as the input basis for subsequent transmission mode decisions.

[0109] It should be noted that the communication decision unit is an embedded logic unit deployed in the sorting equipment. It is composed of a lightweight neural network. The decision logic and parameters are obtained through offline training and rule induction on historical operating data and are solidified in the equipment firmware for real-time transmission mode judgment.

[0110] Based on the timestamp, the sorting equipment obtains historical confidence sequences, network status indicators, and data importance sequences through the communication decision unit.

[0111] The specific process includes: based on the timestamp, the communication decision-maker built into the sorting equipment retrieves the historical confidence sequence within the time window associated with the current moment from the locally stored historical records. The historical confidence sequence consists of confidence scores in structured data arranged in chronological order. At the same time, the communication decision-maker obtains network status indicators within the same time window from the network monitoring unit, including latency, bandwidth utilization, and packet loss rate. In addition, the communication decision-maker also determines the relative importance of each piece of data based on the priority of the sorting task and the type of waste, and forms a data importance sequence in chronological order.

[0112] Based on historical confidence sequences, network state indicators, and data importance sequences, a dynamic decision threshold is calculated using an integral-period formula, expressed as:

[0113] ;

[0114] in, Indicates time The dynamic determination threshold, Represents the normalization factor. This indicates the starting and ending point of the historical data integration process. Indicates a point in history The dynamic weight function at the location, Indicates a historical point in time. Indicates a point in history The historical confidence sequence at that location, Represents the exponential decay coefficient. Indicates the volatility weighting coefficient. This represents the standard deviation of the historical confidence series at time t. This represents the periodic adjustment coefficient. This represents the periodic parameter.

[0115] The specific process includes calculating a dynamic judgment threshold using an integral-period formula based on historical confidence sequences, network state indicators, and data importance sequences. The historical confidence sequences are weighted and accumulated through an integral term, where the weights are determined by a dynamic weight function and decay over time. Simultaneously, the fluctuation weight coefficient is multiplied by the standard deviation of the historical confidence sequences to reflect the uncertainty changes in the current confidence. A sine term consisting of a periodic adjustment coefficient and a periodic parameter is then superimposed to capture the periodic characteristics of confidence changes, generating a dynamic judgment threshold that changes over time.

[0116] It should be noted that the exponential decay coefficient is a parameter used to determine how quickly the confidence weight at each moment in the historical confidence sequence decays over time. It is obtained through offline optimization on historical running data and remains fixed during method execution.

[0117] The fluctuation weighting coefficient is a parameter used to adjust the influence of the standard deviation of the historical confidence sequence on the dynamic judgment threshold. It is obtained through offline optimization on historical running data and remains fixed during the method execution.

[0118] The periodic adjustment coefficient is a parameter used to control the periodic fluctuation amplitude in the dynamic judgment threshold. It is obtained through offline optimization on historical running data and remains fixed during method execution.

[0119] The confidence score is compared with a dynamic threshold, and the transmission mode is determined.

[0120] It should be noted that the built-in communication decision-maker compares the confidence score in the current structured data with the dynamic decision threshold calculated by the integral-period formula. If the confidence score is greater than or equal to the dynamic decision threshold, the decision is to use the high-confidence transmission mode; if the confidence score is less than the dynamic decision threshold, the decision is to use the low-confidence transmission mode.

[0121] When the transmission mode decision is high-confidence transmission, the communication decision trigger high-confidence transmission mode, and structured data is uploaded to the remote control center through the ultra-high reliable and low latency communication slice channel.

[0122] The specific process includes that when the transmission mode decision is high-confidence transmission, the built-in communication decision trigger high-confidence transmission mode, and the current generated structured data is encapsulated according to the protocol format specified by the remote control center, forming a standard data frame, and an established ultra-high reliable and low latency communication slice channel is selected as the transmission path, which provides end-to-end low latency guarantee and high reliability service level, and then the encapsulated structured data is uploaded to the remote control center in real time through the ultra-high reliable and low latency communication slice channel, ensuring that the recognition result is delivered under the premise of meeting the task timeliness and transmission integrity.

[0123] When the transmission mode decision is low-confidence transmission, the communication decision trigger composite transmission mode, and the alarm signaling is sent through the low latency slice, and the associated key image data is uploaded to the remote control center through the large bandwidth slice.

[0124] The specific process includes that when the transmission mode decision is low-confidence transmission, the built-in communication decision trigger composite transmission mode, on the one hand, the alarm signaling containing the recognition uncertainty information is sent to the remote control center through the low latency slice to ensure that the alarm information reaches in the shortest time, and on the other hand, the key image data associated with the current structured data is extracted from the local cache and uploaded to the remote control center through the large bandwidth slice, and the high throughput provided by the large bandwidth slice guarantees the integrity and clarity of the image data, thereby supporting the remote control center to perform manual review or advanced intelligent analysis.

[0125] S6, the remote control center fuses the structured data, the alarm signaling and the key image data, generates a control instruction, and issues the control instruction to the sorting equipment through a special network slice to execute a sorting action, and feeds back the execution state of the sorting equipment to the remote control center.

[0126] Based on the structured data, the alarm signaling and the key image data, the remote control center performs data alignment and noise filtering, and generates a unified environment state estimation vector using a weighted fusion algorithm.

[0127] The specific process includes that the remote control center maps information from different transmission channels to a unified time reference according to time alignment of the structured data, the alarm signaling and the key image data based on the received structured data, alarm signaling and key image data, and respectively filters noise in the confidence score in the structured data, the anomaly identifier in the alarm signaling and the visual feature in the key image data to eliminate abnormal values introduced by transmission errors or identification interference. Then, the cleaned multi-source information is weighted and fused according to a preset weight distribution strategy, wherein the structured data is given a basic decision weight, the alarm signaling is given an anomaly correction weight, and the key image data is given a visual verification weight, to generate a unified environment state estimation vector that can comprehensively reflect the real situation of the current sorting environment.

[0128] It should be noted that the weight distribution strategy is preset by offline optimization before the method is implemented according to the sorting task type, historical identification accuracy and network transmission reliability.

[0129] Based on the environment state estimation vector, an optimal control action vector is calculated through a reinforcement learning model to generate a control instruction set, and the expression is:

[0130] ;

[0131] Among them, represents the optimal control action vector, represents the action space, represents any candidate action in the action space, represents the temperature parameter, represents the state-action value function, represents the environment state.

[0132] The specific process includes that based on the environment state estimation vector, an optimal control action vector is calculated through a reinforcement learning model, the reinforcement learning model uses the state-action value function to evaluate the long-term return of the current environment state and possible actions, and combines the temperature parameter to adjust the exploration and utilization balance of action selection. The exponential weighted probability of all candidate actions in the action space is normalized through the integral form to generate the optimal control action vector with the highest expected return.

[0133] The training process of the reinforcement learning model: carried out in the remote control center, based on the structured data accumulated in the historical sorting tasks, alarm signaling, key image data and corresponding execution results to build an experience replay buffer; then sample state transition sequences from the experience replay buffer, where the state is a unified environment state estimation vector, the action is the actual issued control instruction, and the reward is a scalar value calculated by the task completion quality, energy consumption and delay; use these samples to update the state-action value function in the reinforcement learning model through the time difference method, so that it gradually approaches the true cumulative return; during the training process, the neural network parameters of the reinforcement learning model are optimized using the policy gradient or deep Q network algorithm, and the target network mechanism is used to stabilize the training process; after multiple iterations, the reinforcement learning model can output the optimal control action vector with high long-term return according to the current environment state estimation vector.

[0134] The remote control center encapsulates the control instruction set as a protocol data unit and transmits it to the sorting device through the secure downlink channel of the dedicated network slice.

[0135] The specific process includes that the remote control center encapsulates the control instruction set according to the pre-defined communication protocol format to form a structured protocol data unit, and sends the protocol data unit to the waste sorting device through the secure downlink channel configured in the dedicated network slice, so that the control instruction has identity authentication, data encryption and integrity verification capabilities during transmission, thereby reliably delivering to the waste sorting device for executing sorting actions.

[0136] After the sorting device receives and parses the control instruction set, it executes the sorting action and feeds back the sorting result and device state parameters to the remote control center.

[0137] The specific process includes that after the sorting device receives the control instruction set issued through the dedicated network slice, it parses the control instruction set according to the format of the protocol data unit, extracts the sorting action instructions and resource configuration parameters contained therein, drives the execution mechanism to complete physical operations such as grabbing, moving or dropping to execute the sorting action, and collects sorting result information and device state parameters after the action is completed, including waste actual dropping category, execution completion time, mechanical arm position state and energy consumption data, then encapsulates the above information as feedback data and uploads it to the remote control center through the dedicated network slice.

[0138] In summary, the present application realizes precise matching of sorting task demand and network computing resources through the dynamic strategy optimization mechanism of the multi-agent reinforcement learning framework, solves the cross-domain resource collaborative optimization problem existing in traditional static resource configuration, and reduces the resource allocation error rate; through integral-period formula calculation of dynamic determination threshold, intelligent switching of data transmission mode is realized, and the defect of slow response of fixed threshold mechanism in complex network environment is solved.

[0139] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and all these modifications and equivalents should be included in the scope of the claims of the present application.

Claims

1. A wireless communication-based remote control method for a waste sorting apparatus, characterized by: The remote control center receives a sorting task work order, obtains an optimal resource allocation strategy through a multi-agent reinforcement learning framework, and generates a task resource demand list using a stochastic optimization algorithm; Based on the task resource demand list, the remote control center generates a network slice instantiation request and sends it to a mobile network operator to obtain an access identifier and configuration parameters of a dedicated network slice matched with the sorting task, and the specific steps are as follows, The remote control center generates a network slice instantiation request of the slice configuration demand based on the network service quality parameters and resource constraint conditions in the task resource demand list; The network slice instantiation request is sent to the core network control unit of the mobile network operator through a secure communication protocol; The core network control unit of the mobile network operator performs slice resource allocation on the network slice instantiation request, and dynamically creates a dedicated network slice matched with the sorting task demand; The remote control center obtains the access identifier and configuration parameters from the dedicated network slice through an authorization interface; The remote control center sends the access identifier and configuration parameters of the dedicated network slice to the waste sorting equipment to perform the sorting task, and obtains the network environment access state customized for the sorting task; Based on the network environment access state, the sorting equipment starts the visual recognition function, and the integrated edge artificial intelligence model performs real-time visual analysis on the waste to generate structured data, and the specific steps are as follows, The sorting equipment generates a visual recognition start signal based on the network environment access state; Based on the visual recognition start signal, the visual recognition function is started and the integrated edge artificial intelligence model is initialized; According to the edge artificial intelligence model, the sorting equipment performs real-time visual analysis on the waste to generate a class probability distribution; The maximum value in the class probability distribution is calculated to obtain a confidence score, and structured data is generated; The sorting equipment analyzes the structured data in real time through the built-in communication decision maker, and when the high confidence condition is met, the structured data is transmitted through the low-latency slice, otherwise the composite transmission mode is triggered to send alarm signaling and key image data simultaneously, and sent to the remote control center through the corresponding slice channel; The remote control center fuses the structured data, alarm signaling and key image data to generate a control instruction, which is sent to the sorting equipment to perform the sorting action, and the execution state of the sorting equipment is fed back to the remote control center. The remote control center receives a sorting task work order, obtains an optimal resource allocation strategy through a multi-agent reinforcement learning framework, and the specific steps are as follows, 2. The wireless communication-based waste sorting apparatus remote control method according to claim 1, characterized by: The remote control center receives a sorting task work order and uniformly converts it into a high-dimensional feature vector to establish a work order feature library; Based on the work order feature library, the strategy network of the multi-agent reinforcement learning framework generates a preliminary control strategy, and the value network is used for state evaluation to generate a set of state evaluation quantitative indicators; The preliminary control strategy and the set of state evaluation quantitative indicators are coordinated for resource allocation to obtain an optimal resource allocation strategy. The task resource demand list is generated using a stochastic optimization algorithm, and the specific steps are as follows, 3. The wireless communication-based waste sorting apparatus remote control method of claim 2, wherein: ​ Extract network service quality parameters from the optimal resource allocation strategy, and combine historical operation data to build a dynamic evolution framework to predict the dynamic change trend of service quality in the future time period; Based on the dynamic change trend of service quality in the future time period, the minimum expected total cost problem of resource constraint condition is obtained by using a stochastic optimization algorithm, and a task resource demand list is generated.

4. The wireless communication-based waste sorting apparatus remote control method of claim 3, wherein: The remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment to perform the sorting task, and obtains the access state of the network environment customized for the sorting task, the specific steps are as follows, The remote control center issues the access identifier and configuration parameters of the special network slice to the waste sorting equipment to perform the sorting task, and obtains the access state of the network environment customized for the sorting task, the specific steps are as follows, The waste sorting equipment initializes the network connection and accesses the special network slice according to the access identifier and configuration parameters; After accessing the special network slice, the waste sorting equipment performs the sorting task and transmits the sorting operation data stream in real time through the special network slice; Through the integrated monitoring interface, the sorting operation data stream is captured in real time, and performance analysis and state evaluation are performed to generate network environment access state information customized for the sorting task.

5. The wireless communication-based waste sorting apparatus remote control method of claim 4, wherein: The sorting equipment analyzes the structured data in real time through the built-in communication decision maker, and when the high confidence condition is met, the structured data is transmitted through the low latency slice, otherwise the composite transmission mode is triggered to send alarm signaling and key image data synchronously, and the corresponding slice channel is sent to the remote control center, the specific steps are as follows, The sorting equipment analyzes the structured data through the built-in communication decision maker to obtain the confidence score and timestamp; According to the timestamp, the sorting equipment obtains the historical confidence sequence, network state index and data importance sequence through the communication decision maker; Based on the historical confidence sequence, network state index and data importance sequence, the dynamic threshold is calculated using the integral-period formula; Compare the confidence score with the dynamic threshold to determine the transmission mode; When the transmission mode decision is high confidence transmission, the communication decision maker triggers the high confidence transmission mode to upload the structured data to the remote control center through the ultra-high reliability and low latency communication slice channel; When the transmission mode decision is low confidence transmission, the communication decision maker triggers the composite transmission mode to send alarm signaling through the low latency slice, and uploads the associated key image data to the remote control center through the large bandwidth slice.

6. The wireless communication-based waste sorting apparatus remote control method of claim 5, wherein: The remote control center fuses the structured data, alarm signaling and key image data to generate a control instruction set, the specific steps are as follows, Based on the structured data, alarm signaling and key image data, the remote control center performs data alignment and noise filtering, and generates a unified environment state estimation vector using a weighted fusion algorithm; Based on the environment state estimation vector, the optimal control action vector is calculated through the reinforcement learning model to generate the control instruction set.

7. The wireless communication-based waste sorting apparatus remote control method of claim 6, wherein: The control instruction set is encapsulated as a protocol data unit by the remote control center, and is transmitted to the sorting equipment through the secure downlink channel of the special network slice, and the specific steps are as follows, ​ After receiving and analyzing the control instruction set, the sorting device performs the sorting action and feeds back the sorting result and device state parameters to the remote control center.

Images