An artificial intelligence-based solid waste treatment intelligent optimization system

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and solid waste management, specifically providing an AI-based intelligent optimization system for solid waste management. The system includes: a multi-source sensing module that collects and preprocesses raw waste images; an AI recognition and classification module that outputs waste category prediction results and confidence levels; a priority coding module that couples waste category value, value decay function, and classification confidence levels in a product form to generate dynamic collection priority codes; a route optimization module that integrates collection cost and priority satisfaction to solve for the optimal collection route; and a closed-loop feedback optimization module that filters samples with low confidence levels and inconsistent sorting verification, triggering incremental fine-tuning of the AI ​​recognition and classification module. This invention, through confidence-value decay coupling and a closed-loop incremental feedback mechanism, achieves coordinated optimization of the recognition and collection processes, significantly improving the accuracy of collection resource allocation and the system's continuous optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and solid waste management, specifically to an intelligent optimization system for solid waste management based on artificial intelligence. Background Technology

[0002] The generation of urban solid waste continues to rise, and solid waste management faces two major challenges: inaccurate front-end identification and classification, and inefficient back-end collection and scheduling. Regarding identification and classification: existing technologies generally employ deep learning models, such as ResNet, VGG, and YOLO architectures, to identify waste images. Attention mechanisms or multi-sensor fusion are being introduced to improve accuracy. Visual similarities exist between different categories of waste; for example, cardboard and paper have a high cross-classification misclassification rate, with cardboard having a recall rate of only 59%. Existing methods only output category labels and do not quantify or utilize the uncertainty (confidence) of the classification results. Regarding collection route planning: existing solutions are mostly based on the Vehicle Routing Problem (VRP) model, using metaheuristic methods such as genetic algorithms, particle swarm optimization, or simulated annealing to solve the problem. The optimization objectives are usually transportation costs, carbon emissions, or time window constraints. Fixed collection frequencies are set according to waste type, but priority calculation relies on static category weights or physical overflow.

[0003] The existing technology has several major flaws: First, there is an information gap between the AI ​​identification and collection scheduling stages. Confidence information generated during identification is discarded, and the collection system cannot perceive the reliability of AI judgments. This can lead to high-value recyclables missing the optimal collection window due to misjudgment or uncertainty. Second, the collection priority calculation does not mathematically couple AI identification confidence with the time-dependent decay of waste value, lacking a dynamic scheduling mechanism. Third, once the classification model is deployed, its parameters are fixed, and manual verification results during the collection and sorting stage cannot provide closed-loop feedback to optimize the model. Fourth, the path optimization objective function does not incorporate the time-dependent decay of waste value, resulting in a disconnect between the collection plan and the goal of maximizing the economic value of recycling. Summary of the Invention

[0004] This invention proposes an AI-based intelligent optimization system for solid waste management. Addressing issues in existing technologies such as information gaps between identification and collection stages, lack of confidence-driven dynamic scheduling mechanisms, missing closed-loop feedback, and failure to consider the time-degradation characteristics of waste value in path optimization, a closed-loop optimization system is constructed integrating multi-source visual perception, confidence assessment, priority dynamic coding, and hybrid intelligent path optimization. Visual perception units are deployed at multiple waste disposal points to collect waste images in real time, which are then preprocessed by embedded edge computing devices. A pre-trained deep residual network is used in the cloud to extract deep visual features and output category predictions and confidence distributions. The priority coding module couples the waste category baseline value, category-differentiated value decay function, and classification confidence in a product form to generate dynamic collection priority codes. The path optimization module uses a hybrid genetic-simulated annealing algorithm to solve for the optimal collection path, aiming to minimize overall cost. The verification results from the collection and sorting stages are filtered through low-confidence associations, triggering incremental fine-tuning of the AI ​​identification module. The system deeply integrates the uncertainty of AI identification and the timeliness of waste value into the collection and transportation decision-making through the "confidence-value decay coupling drive + closed-loop incremental feedback" mechanism. This forms an end-to-end collaborative closed loop from perception and identification to scheduling and execution and then to model optimization, which significantly improves the accuracy of collection and transportation resource allocation, the recycling efficiency of high-value recyclables, and the system's continuous optimization capability in complex environments, providing a new technical approach for the intelligent upgrading of solid waste management.

[0005] The technical solution adopted by this invention is as follows: This solution provides an intelligent optimization system for solid waste treatment based on artificial intelligence, including a multi-source sensing module, an AI recognition and classification module, a classification confidence assessment module, a priority coding module, a path optimization module, a collection and sorting verification module, and a closed-loop feedback optimization module.

[0006] Multi-source sensing modules are deployed at various waste disposal points to collect visual data of solid waste, perform preprocessing operations on the visual data, standardize image specifications, and generate images to be recognized.

[0007] The AI ​​recognition and classification module has a built-in pre-trained deep residual network model. It receives the image to be recognized, extracts deep visual features through multi-layer convolution operations, and outputs the waste category prediction result and the corresponding prediction probability distribution based on the deep visual features.

[0008] The classification confidence assessment module receives the predicted probability distribution, calculates the classification confidence based on the degree of difference between the probability values ​​in the predicted probability distribution, and generates low confidence sample labels according to the preset confidence judgment rules.

[0009] The priority coding module receives the waste category prediction results and classification confidence levels, and calls a preset value decay function. This function sets a category-specific decay coefficient based on the differences in the physicochemical properties of the waste categories indicated by the prediction results. The priority coding module determines a baseline value weight based on the waste category, calculates a time decay factor based on the value decay function and the pre-obtained residence time of the waste at the disposal point, calculates a confidence correction factor based on the classification confidence level, and couples the baseline value weight, time decay factor, and confidence correction factor in a product form to generate a dynamic collection priority code for each waste disposal point.

[0010] The route optimization module receives dynamic collection priority codes from each waste disposal point and retrieves pre-stored geographical location information of each waste disposal point, as well as pre-acquired information on the status of currently available collection vehicle resources. The module constructs a comprehensive objective function to minimize overall cost, including a collection cost term and a priority satisfaction term. The priority satisfaction term is constructed by weighting and accumulating the dynamic collection priority codes of each waste disposal point using a time-related satisfaction decay factor. The module then calls a pre-defined hybrid optimization solver to optimize the comprehensive objective function and outputs the optimal collection route.

[0011] The waste collection execution and sorting verification module receives the optimal collection route and dispatches collection vehicles to visit each waste disposal point sequentially along the optimal route to execute the waste collection task. After the waste is collected to the waste transfer station, the waste collection execution and sorting verification module performs secondary sorting on the waste to be sorted, obtains the true category information of each waste sample, and compares the true category information with the waste category prediction results to generate sorting verification results.

[0012] The closed-loop feedback optimization module receives the sorting verification results and obtains low-confidence sample identifiers. It then performs correlation analysis between the sorting verification results and the low-confidence sample identifiers. If a waste sample simultaneously meets the criteria of having an inconsistent true category information in the sorting verification results with the waste category prediction output by the AI ​​recognition and classification module, and is labeled as a low-confidence sample, it is included in the incremental training dataset. When the accumulated number of samples in the incremental training dataset reaches a preset trigger condition, the closed-loop feedback optimization module triggers the incremental fine-tuning operation of the AI ​​recognition and classification module, updating the network weight parameters of the AI ​​recognition and classification module.

[0013] Furthermore, when generating dynamic collection priority codes, the priority coding module couples the baseline value weight of the waste category, the time decay factor derived from the value decay function, and the confidence correction factor in a product form. This product coupling method creates a non-linear amplification effect among the three factors: when the waste category corresponding to a waste disposal point has a high economic value, a long retention time, and a low classification confidence, the product of the three factors causes the dynamic collection priority code to increase sharply. This forces the disposal point to be placed at the front of the collection sequence when the path optimization module solves for the optimal collection path. This mechanism introduces the identification uncertainty information, represented by the classification confidence, accompanying the waste category prediction results into the collection scheduling decision-making link for the first time, solving the problem of information gap between identification and collection. Simultaneously, the value decay function uses a differentiated decay coefficient strongly correlated with the waste category, enabling the system to automatically adjust the slope of priority growth over time based on the waste's own physicochemical characteristics (such as moisture sensitivity and oxidation rate).

[0014] Furthermore, the route optimization module constructs a comprehensive objective function that integrates the transportation cost and priority satisfaction terms when solving for the optimal transportation route. The priority satisfaction term is composed of the weighted sum of the dynamic transportation priority codes of each collection point and a satisfaction decay factor. The satisfaction decay factor is a negative exponential function of arrival time, meaning that even if a collection point has a high dynamic transportation priority code, if the transportation vehicle arrives late due to a later route arrangement, the contribution of that collection point to the satisfaction objective will be exponentially weakened. This design forces the optimization algorithm to not only place high-priority collection points at the beginning of the sequence when generating transportation routes, but also to minimize the travel distance from the originating station to that collection point, thereby achieving priority coverage of "high-value, high-timeliness, and high-uncertainty" waste collection points globally. The classification confidence score output by the classification confidence assessment module characterizes the identification uncertainty; the value decay function and its category-specific decay coefficient describe the characteristics of waste value decaying over time. This strategy integrates identification uncertainty, the characteristics of waste value decaying over time, and vehicle transportation economic costs into a unified mathematical programming framework, achieving the organic integration of multi-dimensional optimization objectives.

[0015] Furthermore, the closed-loop feedback optimization module employs a dual screening strategy when selecting incremental training samples: only waste samples that simultaneously meet the criteria of "inconsistency between sorting verification results and waste category prediction results" and "being marked as low-confidence samples" are included in the incremental training dataset. The synergistic effect of this screening strategy is reflected in two aspects: targeted screening of low-confidence samples with inconsistent verification, accurately locating difficult samples on the classification decision boundary, and avoiding the inclusion of high-confidence samples with occasional errors or low-confidence samples with correct verification in training, thus preventing model oscillation and overfitting; since the priority encoding module has assigned higher collection priority to low-confidence samples, these difficult samples are collected, sorted, and verified first in the collection process, thus entering the incremental training pool in a shorter cycle, forming a positive acceleration closed loop of "low confidence → high priority → priority collection → priority verification → priority feedback → priority optimization". This mechanism enables the AI ​​recognition and classification module to continuously optimize the classification decision boundary with minimal training cost, achieving self-evolution of the system in long-term operation.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] (1) To address the information gap between the identification and collection stages, this invention uses a priority coding module to couple the classification confidence level output by the AI ​​identification and classification module with the value decay function to generate a dynamic collection priority code. The dynamic collection priority code integrates three types of information: waste category value, retention time decay, and identification confidence level. This enables the collection system to automatically identify and prioritize waste disposal points that are "uncertain in AI but may be of high value," thus achieving information connectivity and collaborative linkage between the identification and collection stages.

[0018] (2) To address the lack of a dynamic coupling scheduling mechanism in the calculation of waste collection priority, this invention constructs a comprehensive objective function that integrates waste collection cost and priority satisfaction through a path optimization module. The priority satisfaction is composed of a weighted sum of the dynamic waste collection priority code and the negative exponential decay factor of the arrival time. This design forces high-priority disposal points to be placed at the front of the path, and effectively suppresses the satisfaction decay caused by the later path, thus achieving global priority coverage of high-value, high-timeliness, and high-uncertainty waste disposal points.

[0019] (3) To address the problem of the inability to perform closed-loop optimization after the classification model is deployed, this invention implements a dual screening strategy through a closed-loop feedback optimization module, which only includes samples that are "inconsistent between sorting verification and AI prediction" and "marked as low confidence" into the incremental training dataset. This strategy accurately locates difficult samples at the classification boundary, and at the same time uses the priority encoding module to assign high priority to low confidence samples, forming a positive acceleration closed loop of "low confidence → priority sorting → priority verification → priority feedback", so that the AI ​​recognition and classification module can continuously optimize the classification decision boundary with minimal training cost;

[0020] (4) To address the issue that the path optimization objective does not incorporate the time decay difference of waste value, this invention introduces a decay coefficient set according to the differences in the physicochemical properties of waste categories in the priority coding module, and incorporates this decay characteristic as a priority satisfaction term in the path optimization objective function. During the solution process, the optimization algorithm simultaneously weighs the economic cost of transportation with the priority satisfaction weighted by time decay, so that the collection plan maximizes the recycling benefits of high-value waste while controlling costs, and realizes refined and differentiated dynamic scheduling. Attached Figure Description

[0021] Figure 1 This is an overall architecture diagram of an artificial intelligence-based intelligent optimization system for solid waste treatment according to the present invention.

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Example:

[0025] Please see Figure 1 The artificial intelligence-based intelligent optimization system for solid waste management in this embodiment includes a multi-source sensing module, an AI recognition and classification module, a classification confidence assessment module, a priority coding module, a path optimization module, a collection and sorting verification module, and a closed-loop feedback optimization module.

[0026] The multi-source sensing module collects visual data of solid waste and generates images to be identified.

[0027] The AI ​​recognition and classification module processes the image to be recognized and outputs the waste category prediction results and the predicted probability distribution.

[0028] The classification confidence assessment module calculates the classification confidence based on the predicted probability distribution and generates low-confidence sample labels.

[0029] The priority coding module obtains the baseline value weight based on the waste category prediction results, calculates the confidence correction factor based on the classification confidence level, and calls the preset value decay function to calculate the time decay factor. The baseline value weight, time decay factor, and confidence correction factor are coupled to generate a dynamic waste collection priority code.

[0030] The route optimization module receives the dynamic collection priority code of each collection point, obtains the geographical location information of the collection point and the vehicle resource status information, and solves the optimal collection route through a hybrid genetic-simulated annealing algorithm.

[0031] The waste collection and sorting verification module executes the collection according to the optimal collection route, performs secondary sorting at the waste transfer station, and generates sorting verification results.

[0032] The closed-loop feedback optimization module receives the sorting verification results and associates them with the low-confidence sample identifier. When the conditions are met that "the sorting verification result is inconsistent with the waste category prediction result" and "the low-confidence sample identifier is 1", the sample is included in the incremental training dataset, triggering the incremental fine-tuning operation of the AI ​​recognition and classification module.

[0033] Preferably, the multi-source sensing module is deployed at each waste disposal point. Each disposal point is equipped with a high-definition camera to collect visual data of solid waste. The camera is connected to an edge computing unit. The edge computing unit performs preprocessing operations on the visual data. The visual data is the raw image. The raw image is uniformly scaled to a resolution of 1024×768 pixels to ensure that it matches the input layer size of the AI ​​recognition and classification module. Each pixel in the scaled image has a pixel value, and the pixel values ​​are normalized, linearly mapping the value range from [0, 255] to the interval [0, 1]. The normalized image is output as the image to be recognized.

[0034] Preferably, the AI ​​recognition and classification module receives the image to be recognized, extracts deep visual features, and outputs the waste category prediction result and the predicted probability distribution. The AI ​​recognition and classification module uses a deep residual network based on ResNet-50 as the backbone network. The residual structure of ResNet-50 alleviates the gradient vanishing problem in deep networks through skip connections, enabling the network to learn deeper feature representations. In waste classification tasks, ResNet-50 has better generalization performance compared to deeper architectures. Deeper architectures are prone to overfitting on finite-scale datasets due to their large number of parameters, while ResNet-50 performs better in terms of training stability and convergence speed.

[0035] The network structure is as follows:

[0036] The input layer receives the image to be recognized. The image to be recognized is represented in the form of a three-dimensional tensor, with the three dimensions corresponding to the image height in sequence. pixels, image width Pixels and number of color channels .

[0037] The backbone network is a ResNet-50 convolutional base (with the original top fully connected layer removed), containing Each residual block and Multiple convolutional layers are used to extract multi-level visual features.

[0038] Add a after the convolution base The hidden layer has 3D dimensions, and the activation function is ReLU. Let the input vector of the hidden layer be... The ReLU function is defined as follows: .

[0039] The output layer is The fully connected layer corresponds to four waste categories: cardboard, metal, paper, and plastic. The raw linear fraction of the output is called the logits value. Let the output layer be... The logits value of each neuron is Then the first Predicted probability of class Calculate using the following formula:

[0040] ;

[0041] The waste category prediction result is the category corresponding to the highest predicted probability:

[0042] ;

[0043] The predicted probability distribution is a probability vector of four categories. .

[0044] The training strategy for the AI ​​recognition and classification module is as follows:

[0045] Transfer learning was employed. The image sample set consisting of images to be identified was divided into an 80% training set and a 20% validation set. During training, the weights of the shallow convolutional layers of the ResNet-50 convolutional base were frozen, and only the weights of the deep convolutional layers, the newly added 512-dimensional fully connected hidden layer, and the 4-dimensional output layer were fine-tuned. The loss function used was sparse classification cross-entropy. Let the... The true class of each sample is (Integer encoding), predicted probability is Then the loss value for:

[0046] ;

[0047] in The batch size is set to [value]. The optimizer uses Adam, and the initial learning rate is set to [value]. Batch size set to The training rounds are set to .

[0048] Preferably, the classification confidence assessment module receives the predicted probability distribution, calculates the classification confidence, and generates low-confidence sample identifiers. This embodiment uses the probability difference method to calculate the classification confidence. Let the probability values ​​in the predicted probability distribution be arranged in descending order as follows: Then the classification confidence level Calculate using the following formula:

[0049] ;

[0050] This metric reflects the difference between the highest probability and the second highest probability. When the AI ​​recognition and classification module is highly certain about the category of the image to be recognized, the probability distribution exhibits a unimodal shape. Much larger The classification confidence level approaches 100%. When there is a high degree of uncertainty in the AI ​​recognition and classification module, the probability distribution exhibits a multi-peaked shape. and The classification confidence level is close to that of the classifier. Therefore, classification confidence can effectively quantify the degree of ambiguity of classification boundaries.

[0051] Low confidence sample identification The generation rule is as follows: set a confidence threshold. In this embodiment .when hour, (Indicates a low-confidence sample); otherwise .

[0052] The low-confidence sample identifiers serve a dual purpose: on the one hand, they are sent to the priority encoding module to adjust the collection priority; on the other hand, they are sent to the closed-loop feedback optimization module to filter incremental training samples.

[0053] Preferably, the priority coding module receives the waste category prediction results and classification confidence scores, calls the value decay function, and generates a dynamic collection priority code. The specific working steps of the priority coding module include:

[0054] Step S401: Receive the waste category prediction results and classification confidence scores.

[0055] Step S402: Obtain the benchmark value weight The values ​​are retrieved from a preset weight table based on the waste category prediction results. This preset weight table is stored in the local cache of the priority encoding module and includes: metals. ,plastic paper ,cardboard The benchmark value weight reflects the average economic value of recycling per unit weight of various types of waste.

[0056] Step S403: Calculate the time decay factor Waste retention time at disposal points (Unit: hours) The longer the time, the more severe the quality degradation due to factors such as moisture, oxidation, and contamination. This embodiment uses an exponential decay function for modeling:

[0057] ;

[0058] in This is the attenuation coefficient. Differentiated settings based on waste category. Paper and cardboard are susceptible to moisture. The value is relatively large; metals and plastics have good weather resistance. The value is relatively small. In this embodiment, the cardboard... paper ,plastic ,Metal .

[0059] Step S404: Calculate the confidence correction factor Classification confidence The lower the confidence level, the higher the identification uncertainty. For low-confidence samples, the system increases their clearance priority so that the true category can be confirmed through manual sorting. Calculate using the following formula:

[0060] ;

[0061] in In this embodiment, the confidence penalty coefficient is used. .when hour, ;when hour, .

[0062] Step S405: Couple and generate dynamic cleaning priority codes The benchmark value weight, time decay factor, and confidence correction factor are coupled in a product form:

[0063] ;

[0064] The synergistic effect mechanism of this coupling method lies in:

[0065] (a) Benchmark Value Weights Provide benchmark differences in the economic value of different categories of waste;

[0066] (b) The reciprocal of the time decay factor Depending on the length of stay Exponential growth, and the rate of growth is class-dependent. This control allows high-decay categories to receive a higher priority boost than low-decay categories for the same residence time, thus automatically securing a higher position in path optimization;

[0067] (c) Confidence correction factor When identification is highly uncertain, priority is increased, enabling the system to respond quickly to suspected high-value but uncertain waste disposal sites;

[0068] (d) The three factors are coupled by product rather than nonlinear weighting, resulting in a nonlinear amplification effect among them. When a sample simultaneously meets the three conditions of high economic value of waste category, long retention time, and low classification confidence, The value increases sharply, and the forced path optimization module places the delivery point at the beginning of the path. This nonlinear amplification effect cannot be achieved by simple linear weighting.

[0069] Step S406: Output dynamic cleaning priority code To the path optimization module.

[0070] Preferably, the route optimization module receives the dynamic collection priority code of each collection point, obtains geographical location information and vehicle resource status information, and solves the optimal collection route with the goal of minimizing the overall cost.

[0071] The path optimization module uses a hybrid genetic-simulated annealing algorithm to solve the problem. Combining the global search capability of the genetic algorithm with the local escape capability of the simulated annealing algorithm, it can obtain a balanced solution that takes into account both the transportation cost and the priority satisfaction under complex constraints.

[0072] The overall objective function is:

[0073] ;

[0074] in To synthesize the objective function value, For the total removal cost, Prioritize satisfaction level. and As the weighting coefficient, in this embodiment .

[0075] Total waste removal cost The calculation formula is:

[0076] ;

[0077] in: In this embodiment, the fixed operating cost of a single waste collection vehicle is... (Unit: Yuan / vehicle) For the first Vehicle activation sign ( To enable, (Not enabled) In this embodiment, the unit distance transportation cost is... (Unit: Yuan / km); For delivery points arrive The actual road distance; For decision variables (representing the first...) The car from Drive to ); In this embodiment, the total number of available vehicles is [number]. ; This represents the total number of collection points. Indicates the originating station of the vehicle.

[0078] Priority Satisfaction The calculation formula is:

[0079] ;

[0080] in: For the first Dynamic collection priority coding for each collection point; For the vehicle to arrive at the The timing of each delivery point; In this embodiment, the satisfaction attenuation coefficient is used. .

[0081] Satisfaction decay factor Its function is: even if a certain delivery point Very high, if the vehicle arrives at the time If it's too late, its contribution to the satisfaction goal will also be exponentially decayed. This prompts optimization algorithms to consider not only high-risk paths but also other factors when planning routes. The system prioritizes delivery points and minimizes the total travel time from the originating station to these delivery points. This mechanism maps the three-dimensional information of "value-time-confidence" into the spatial order of the waste collection route, achieving an end-to-end closed-loop drive from AI-identified uncertainty to waste collection execution.

[0082] The constraints include:

[0083] Each delivery point is visited only once by a single vehicle.

[0084] ;

[0085] Each vehicle's driving distance shall not exceed its maximum range. :

[0086] ;

[0087] In this embodiment (Unit: kilometers)

[0088] The vehicle departs from and returns to the originating station:

[0089] ;

[0090] The specific steps for finding the optimal waste disposal route include:

[0091] Step S501: Population initialization. Randomly generated. Individual, in this embodiment Each individual represents a vehicle routing scheme. The encoding uses a double-stranded chromosome: the access sequence chain is a random permutation of the drop-off point numbers; the segment point chain is... An incrementing integer is used to divide the access sequence into... Sub-path. The originating station number will be used during decoding. Add it to the beginning and end of each sub-path.

[0092] Step S502: Fitness Calculation. Calculate the total waste disposal cost for each individual. and priority satisfaction Then calculate the comprehensive objective function value. fitness Values .

[0093] Step S503: Genetic operation. The crossover operation randomly selects a region on the visited sequence chain. The gene sequence of parent generation 1 within the interval is retained, and the rest are filled in the order of parent generation 2. The mutation operation adopts a dynamic mutation rate (the mutation rate is higher in the high temperature stage to enhance the global search, and lower in the low temperature stage to promote local convergence). The mutation methods include exchange mutation and segment point mutation.

[0094] Step S504: Simulated annealing selection. Calculate the fitness difference between offspring and parents. .like Accept offspring; if With probability Accepting offspring, This is the current temperature.

[0095] Step S505: Temperature update. The initial temperature Temperature decay coefficient When the temperature Or reach the maximum number of iterations The event will end at that time.

[0096] The optimal path optimization module outputs the optimal individual, which is then decoded to obtain the optimal cleaning path and transmitted to the cleaning execution and sorting verification module.

[0097] Preferably, the waste collection execution and sorting verification module executes the waste collection task according to the optimal collection route, performs secondary sorting at the waste transfer station, and records the sorting verification results.

[0098] The collection vehicles visit each collection point sequentially along the optimal collection route, collect waste, and load it onto the collection point. After completing all collection tasks, they return to the waste transfer station.

[0099] Set up sorting and verification stations at waste transfer stations. For each batch of waste collected at each disposal point, samples are drawn for secondary sorting according to the following rules:

[0100] The total weight of the batch is less than When kg, extract One sample;

[0101] Total batch weight kg to When the weight is between kg, extract One sample;

[0102] Batch total weight greater than When kg, extract One sample.

[0103] Secondary sorting employs a spectral-based material identification device. The sorting verification results for each sample are recorded, including the original sample image identifier, predicted waste category, actual category information, consistency indicator ("T" for consistency, "F" for inconsistency), disposal point number, and collection timestamp. The sorting verification results are then transmitted to the closed-loop feedback optimization module.

[0104] Preferably, the closed-loop feedback optimization module receives the sorting verification results and associates them with the low-confidence sample identifiers, triggering incremental fine-tuning of the AI ​​recognition and classification module.

[0105] The specific steps of the closed-loop feedback optimization module include:

[0106] Step S701: Receive the sorting verification results and low-confidence sample identifiers.

[0107] Step S702: Filter incremental training samples. Filtering strategy: A sample is included in the incremental training dataset only if both of the following conditions are met: the consistency indicator in the sorting verification results is "F"; the sample is identified as a low-confidence sample. for" ".

[0108] The synergistic mechanism of this dual-screening strategy lies in: only screening samples with low confidence and inconsistent validation, accurately locating difficult samples with ambiguous classification boundaries, and avoiding including samples with high confidence but occasional errors or low confidence but consistent validation in the training, thereby preventing model oscillation and overfitting. Since the priority encoding module assigns higher priority to low-confidence samples, these difficult samples are prioritized for removal, sorting, and entry into the incremental training pool. This forms a positive acceleration loop of "low confidence → high priority → priority removal → priority validation → priority feedback → priority optimization".

[0109] Step S703: Accumulation and Triggering. Accumulate the selected samples into the incremental training data pool. Set the trigger threshold. In this embodiment Incremental fine-tuning is triggered when the cumulative number of samples reaches the trigger threshold.

[0110] Step S704: Incremental fine-tuning. When the cumulative number of samples reaches the trigger threshold... At that time, the closed-loop feedback optimization module performs the following incremental fine-tuning operations:

[0111] (a) Prepare the incremental training dataset. Package the image data of all samples in the incremental training data pool and their corresponding validation categories into a TensorFlow TFRecord format file. The image of each sample has been preprocessed by the multi-source perception module. The tensor is labeled with an integer encoding corresponding to the verification category.

[0112] (b) Fine-tune the training parameters. Keep all convolutional layer weights of the ResNet-50 convolutional base in the AI ​​recognition and classification module frozen, and only adjust the weights of newly added layers. Fully connected hidden layers and The weights of the output layer are set to trainable states. The Adam optimizer is used, with an initial learning rate of [value missing]. Batch size set to The training rounds are set to The loss function continues to use sparse classification cross-entropy.

[0113] (c) Perform fine-tuning training. In each training epoch, randomly sample from the incremental training dataset according to the batch size for forward and backward propagation, updating the trainable weights. Monitor the change in the loss value during training; if it is continuous... If the loss value does not decrease after a certain number of rounds, training will be terminated early.

[0114] (d) Model Update. After fine-tuning the training, the updated hidden layer weights and output layer weights replace the corresponding weight parameters in the AI ​​recognition and classification module. The original convolutional basis weights remain unchanged.

[0115] Step S705: After updating the weights, clear the incremental training data pool and trigger the counter to zero, then enter the next round of sample accumulation cycle.

[0116] This closed-loop feedback mechanism enables the AI ​​recognition and classification module to continuously improve in actual operation, and the recognition accuracy gradually increases with the increase of system running time, solving the technical problem that the model cannot be continuously optimized after deployment in the existing technology.

[0117] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent optimization system for solid waste treatment based on artificial intelligence, characterized in that, The system includes a multi-source sensing module, an AI recognition and classification module, a classification confidence assessment module, a priority coding module, a route optimization module, a collection execution and sorting verification module, and a closed-loop feedback optimization module. The multi-source sensing module collects raw images of solid waste and performs preprocessing to generate images to be identified. The AI ​​recognition and classification module receives the images to be identified and outputs the waste category prediction results and the corresponding prediction probability distribution. The classification confidence assessment module calculates the classification confidence based on the prediction probability distribution and generates low-confidence sample identifiers. The priority coding module receives the waste category prediction results and classification confidence, calls a preset value decay function, and generates dynamic collection priority codes for each waste disposal point. The route optimization module receives dynamic collection priority codes, obtains geographical location information and vehicle resource status information, and solves the optimal collection route with the goal of minimizing the overall cost. The collection execution and sorting verification module executes the collection task according to the optimal collection route and performs secondary sorting to generate sorting verification results. The closed-loop feedback optimization module receives the sorting verification results and low-confidence sample labels. When a waste sample is determined to be inconsistent with the sorting verification results and the waste category prediction results and is marked as a low-confidence sample, it is included in the incremental training dataset and triggers the incremental fine-tuning operation of the AI ​​recognition and classification module.

2. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 1, characterized in that: The value decay function sets a category-specific decay coefficient based on the waste category; the priority coding module determines the benchmark value weight based on the waste category, calculates the time decay factor based on the value decay function and the pre-acquired waste retention time at the disposal point, calculates the confidence correction factor based on the classification confidence, and couples the benchmark value weight, time decay factor and confidence correction factor in a product form to generate a dynamic collection priority code.

3. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 2, characterized in that: The value decay function adopts an exponential decay form and includes a decay coefficient; The attenuation coefficient is set differently based on the waste category.

4. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 2, characterized in that: The confidence correction factor is calculated based on the classification confidence level and the preset confidence penalty coefficient. The lower the classification confidence level, the larger the confidence level correction factor.

5. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 2, characterized in that: The route optimization module constructs a comprehensive objective function, which includes a waste collection cost item and a priority satisfaction item. The priority satisfaction item is composed of dynamic waste collection priority codes weighted by a time-related satisfaction decay factor.

6. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 5, characterized in that: The overall objective function aims to minimize the weighted difference between collection cost and maximize priority satisfaction. In the priority satisfaction term, the dynamic collection priority code is accumulated after being weighted by a satisfaction decay factor in the form of a negative exponential function of arrival time.

7. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 5, characterized in that: The path optimization module uses a hybrid genetic-simulated annealing algorithm to solve for the optimal waste disposal path.

8. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 4, characterized in that: The classification confidence assessment module uses the probability difference method to calculate the classification confidence, which is equal to the difference between the highest probability value and the second highest probability value in the predicted probability distribution.

9. The intelligent optimization system for solid waste treatment based on artificial intelligence according to claim 1, characterized in that: The AI ​​recognition and classification module has a built-in pre-trained deep residual network model, which includes convolutional bases and fully connected layers; the incremental fine-tuning operation keeps the weights of the convolutional bases frozen and only updates the weights of the fully connected layers.

Images