Methods, systems, equipment, and media for processing dynamic monitoring data of construction waste

Abstract

This application provides a method, system, device, and medium for processing dynamic monitoring data of construction waste. The method acquires multi-source construction waste monitoring data, then constructs a training dataset based on this data and according to preset adjustment task types. The training dataset is cleaned and augmented to obtain an augmented training dataset. Then, a large language model is pre-trained and supervised fine-tuned using the augmented training dataset, combined with a reward learning strategy to optimize the large language model. A temporal IoT prediction model is trained simultaneously. Next, a reinforcement learning algorithm is used to update the strategy of the large language model. The output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data are integrated to generate the processing result of the construction waste dynamic adjustment task, thereby achieving dynamic control over the generation and flow of urban construction waste.

Description

Technical Field

[0001] This application relates to data processing technology, and in particular to a method, system, equipment and medium for processing dynamic monitoring data of construction waste. Background Technology

[0002] With the continuous growth of urban construction and residents' renovation needs, the amount of construction waste generated has shown a significant upward trend. The sources of waste are complex and the flow is scattered, with frequent occurrences of littering and illegal dumping.

[0003] Existing construction waste monitoring systems generally rely on manual inspections, fixed-point monitoring, and rule-driven judgment methods. They not only have limited processing capabilities for multi-source heterogeneous data such as unstructured text information, time-series IoT data, and map POIs, but also lack effective data fusion mechanisms and intelligent decision-making capabilities in tasks such as tracing waste sources, predicting littering hotspots, scheduling recycling vehicles, and planning inspection routes. As a result, it is difficult to achieve dynamic and refined management of the generation and flow of urban construction waste. Summary of the Invention

[0004] This application provides a method, system, equipment, and medium for processing dynamic monitoring data of construction waste, so as to realize dynamic control over the generation and flow of urban construction waste.

[0005] Firstly, this application provides a method for processing dynamic monitoring data of construction waste, including: Acquire multi-source construction waste monitoring data, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data; Based on the multi-source construction waste monitoring data and a training dataset constructed according to the preset adjustment task type, the training dataset is cleaned and augmented to obtain the augmented training dataset. The large language model is pre-trained and supervised fine-tuned using the augmented training dataset, and a reward learning strategy is combined to optimize the large language model, while a temporal IoT prediction model is trained simultaneously. The large language model is updated based on a reinforcement learning algorithm. The output of the large language model, the output of the time-series IoT prediction model, and the spatial features of the map POI data are integrated to generate the dynamic adjustment task processing result of construction waste.

[0006] Secondly, this application provides a dynamic monitoring data processing system for construction waste, comprising: The acquisition module is used to acquire multi-source construction waste monitoring data, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data. The augmentation module is used to construct a training dataset based on the multi-source construction waste monitoring data and according to a preset adjustment task type, and to clean and augment the training dataset to obtain an augmented training dataset. The training module is used to pre-train and supervised fine-tuning the large language model using the augmented training dataset, combining a reward learning strategy to optimize the large language model, and simultaneously training the temporal IoT prediction model. The generation module is used to perform policy updates on the large language model based on reinforcement learning algorithms, and to integrate the output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data to generate the processing results of the construction waste dynamic adjustment task.

[0007] Thirdly, this application provides an electronic device, comprising: Processor; and, Memory for storing the executable instructions of the processor; The processor is configured to perform any of the possible methods described in the first aspect by executing the executable instructions.

[0008] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement any of the possible methods described in the first aspect.

[0009] The construction waste dynamic monitoring data processing method, system, equipment, and medium provided in this application acquire multi-source construction waste monitoring data, then construct a training dataset based on the multi-source construction waste monitoring data and according to a preset adjustment task type, clean and augment the training dataset to obtain an augmented training dataset, then use the augmented training dataset to pre-train and supervised fine-tuning a large language model, combining a reward learning strategy to optimize the large language model, and simultaneously training a temporal IoT prediction model, then performing a strategy update on the large language model based on a reinforcement learning algorithm, and integrating the output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data to generate the construction waste dynamic adjustment task processing result, thereby realizing dynamic control of the generation and flow of urban construction waste. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] Figure 1 This is a schematic flowchart illustrating a method for processing dynamic monitoring data of construction waste according to an example embodiment of this application; Figure 2 This is a schematic flowchart of S100 according to an example embodiment of this application; Figure 3 This is a schematic flowchart of S300 according to an example embodiment of this application; Figure 4 This is a schematic diagram of the structure of a construction waste dynamic monitoring data processing system according to an example embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an example embodiment of this application.

[0012] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0014] Figure 1 This is a schematic flowchart illustrating a method for processing dynamic monitoring data of construction waste according to an example embodiment of this application. Figure 1 As shown, the method provided in this embodiment includes: S100, Obtain monitoring data on multi-source construction waste.

[0015] In this step, multi-source construction waste monitoring data may be acquired, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data.

[0016] Figure 2 This is a schematic flowchart of S100 according to an example embodiment of this application. Figure 2 As shown, the above-mentioned S100 includes: S110. For the source tracing task, construct training samples that include decoration registration text, decoration site POI features and video surveillance garbage feature descriptions. The labels of the training samples are the generating entity information.

[0017] Specifically, this could involve obtaining a renovation registration text containing information such as the construction address, construction type, construction period, and responsible party. Field extraction is then performed on the renovation registration text to obtain a construction entity description field and a construction location description field. Based on the construction location description field, the corresponding renovation site POI features are retrieved from a pre-defined POI database, and these features are associated with corresponding garbage feature descriptions identified by video surveillance. Finally, the construction entity description field, the renovation site POI features, and the video surveillance garbage feature descriptions are combined according to a pre-defined format to form training samples for the source tracing task.

[0018] S120. For the task of predicting littering hotspots, a training sample is constructed that includes historical littering time series data, POI distribution of construction sites, construction progress time series data, and meteorological data. The labels of the training samples are the probability of littering in the future target area and the time period.

[0019] Specifically, this involves acquiring historical littering time-series data within a preset time window in the target area. This data includes the timestamps and spatial locations of littering incidents. Next, it involves acquiring POI distribution data for construction sites within the corresponding time period and performing spatial density analysis to determine the density characteristics of these sites. Then, it involves acquiring construction progress time-series data and performing time alignment processing to ensure consistency between the construction progress time-series data and the historical littering time-series data in the time dimension. Finally, it involves extracting meteorological data for the target time period from a meteorological database, including temperature, precipitation, wind speed, and weather type. The historical littering time-series data, construction site density characteristics, aligned construction progress time-series data, and meteorological data are combined to form the training samples for the littering hotspot time-series prediction task.

[0020] S130. For the task of dispatching recycling vehicles, construct training samples that include prediction results of littering hotspots, time-series data of recycling vehicle status, and time-series data of road POI traffic.

[0021] Specifically, this involves acquiring high-risk areas and high-risk time periods from littering hotspot prediction results. It also includes acquiring time-series data on the status of reclaimed vehicles, including vehicle location, speed, load, and availability. Furthermore, it involves acquiring time-series data on road points of interest (POIs), including road capacity, traffic speed, traffic events, and congestion levels. Then, the high-risk areas, high-risk time periods, reclaimed vehicle status time-series data, and road POI time-series data are aligned to their corresponding times to form training samples for the reclaimed vehicle scheduling task.

[0022] S140. For the patrol route optimization task, construct training samples that include time-series data of high-risk POI distribution, time-series data of patrol personnel location, and road POI accessibility.

[0023] Specifically, this can involve acquiring time-series data on the distribution of high-risk Points of Interest (POIs) within the target area, output by a littering hotspot prediction model, and performing spatial indexing based on a pre-defined spatial grid to obtain a time-series feature matrix of high-risk POIs. It also involves acquiring time-series data on the locations of patrol personnel, including their coordinates and movement speed. Furthermore, it involves acquiring accessibility data for road POIs, including road connectivity, road traffic levels, and travel time. Finally, it involves performing spatial and temporal correlation processing on the high-risk POI time-series feature matrix, the patrol personnel location time-series data, and the road POI accessibility data to form training samples corresponding to the patrol route optimization task.

[0024] S200. Based on multi-source construction waste monitoring data and according to preset adjustment task types, a training dataset is constructed. The training dataset is then cleaned and augmented to obtain the augmented training dataset.

[0025] In this step, synonym replacement or sentence rewriting can be performed on the decoration filing text, time shifting and rate adjustment can be performed on the time-series IoT data, and neighborhood replacement operation can be performed on the map POI data to form diverse spatial feature samples.

[0026] S300: The large language model is pre-trained and supervised fine-tuned using the augmented training dataset. A reward learning strategy is combined to optimize the large language model, and a time-series IoT prediction model is trained simultaneously.

[0027] Figure 3 This is a schematic flowchart of S300 according to an example embodiment of this application. Figure 3 As shown, the above-mentioned S300 includes: S310. Encode the text description of the decoration filing to obtain a text representation vector.

[0028] Specifically, the process could involve performing word segmentation on the renovation registration text to divide it into multiple word sequences, converting the word sequences into corresponding word vector sequences based on a pre-defined vocabulary, then inputting the word vector sequences into the text encoder of a large language model to extract contextual semantic features, and finally generating corresponding text representation vectors based on the contextual semantic features.

[0029] S320. Generate the subject probability distribution based on text representation vector.

[0030] Specifically, the text representation vector can be input into a preset fully connected network layer, and Softmax calculation can be performed on the output of the fully connected network layer to obtain the probability values ​​corresponding to multiple candidate subjects. The probability values ​​corresponding to multiple candidate subjects can be combined to form a subject probability distribution.

[0031] S330. The cross-entropy loss function and regularization term are used to supervise the training of the prediction results of the target subject in order to improve the accuracy of source tracing.

[0032] Specifically, a target label vector is constructed based on the information of the generating entities labeled in the training samples. The cross-entropy loss value is calculated based on the target label vector and the probability distribution of the generating entities. A regularization term based on the model parameters is added to the cross-entropy loss value to suppress model overfitting. The model parameters of the large language model are updated by performing gradient backpropagation based on the cross-entropy loss value with the regularization term.

[0033] In addition, the above-mentioned reward learning strategy can include: constructing a cumulative reward function that includes the accuracy of source tracing, the accuracy of litter hotspot prediction, and the efficiency of recycling vehicle dispatching; and performing gradient updates based on the cumulative reward function to improve the overall performance of the model in multi-task scenarios.

[0034] Specifically, this can involve obtaining the predicted source of littering from the model output in the source tracing task and the labels of the source of littering from the training samples, and calculating the source tracing accuracy based on the prediction results and the labels of the source of littering. It can also involve obtaining the future littering probability output by the model in the littering hotspot prediction task and the actual littering occurrences labeled in the training samples, and calculating the littering hotspot prediction accuracy based on the future littering probability and the actual littering occurrences. Furthermore, it can involve obtaining the scheduling scheme generated by the model in the recycling vehicle scheduling task, and calculating the time consumption and path length for recycling vehicles to complete the collection task within the target time period based on the scheduling scheme, to obtain the recycling vehicle scheduling efficiency. Finally, it can combine the source tracing accuracy, littering hotspot prediction accuracy, and recycling vehicle scheduling efficiency based on a preset weighting rule to obtain the reward value of the cumulative reward function.

[0035] Then, the reward value of the cumulative reward function is used as the target reinforcement signal and input into the policy network of the large language model. The corresponding policy gradient is calculated based on the action probability distribution output by the policy network. Gradient ascent is then performed to update the policy parameters of the large language model's policy network based on the policy gradient and the reward value. Stability constraints are then applied to the updated policy parameters to improve the training stability of the large language model in multi-task scenarios.

[0036] Furthermore, the aforementioned time-series IoT prediction model can employ a hybrid structure of LSTM and attention mechanisms. By inputting time-series data from video surveillance and construction progress, it can generate predictions of the probability of littering and the amount of waste generated in the target area in the future.

[0037] S400: Based on reinforcement learning algorithm, the large language model is updated with a strategy. The output of the large language model, the output of the time series IoT prediction model, and the spatial features of map POI data are integrated to generate the dynamic adjustment task processing results of construction waste.

[0038] Specifically, a reinforcement learning algorithm based on proximal policy optimization is used to update the policy network of the large language model. More specifically, this can involve obtaining the action probability distribution of the large language model's output in the current state and sampling policy actions based on this distribution. The policy actions are then compared with the old policy distribution corresponding to the large language model to calculate the probability ratio between them. A pruning objective function is constructed based on this probability ratio to limit its deviation from a preset threshold range. The policy gradient is calculated based on the pruning objective function, and the parameters of the policy network are updated according to the policy gradient.

[0039] For the spatial features of the integrated large language model output, the temporal IoT prediction model output, and map POI data, the process involves obtaining the task text inference result from the large language model output and performing feature vectorization on this result. The probability of littering and the amount of waste generated in the future target area are obtained from the temporal IoT prediction model output, and these are combined into a temporal prediction feature. Spatial features from the map POI data are obtained, including road accessibility, area functional category, and spatial density features. Then, the feature-vectorized task text inference result, temporal prediction features, and spatial features from the map POI data are concatenated according to preset dimensions to form the comprehensive input features corresponding to the dynamic adjustment task of construction waste.

[0040] Then, the comprehensive input features are input into a pre-defined regulation decision generation network, which includes fully connected layers and nonlinear activation layers. Based on the output of the regulation decision generation network, regulation response information for the corresponding task is generated. This information includes at least one of the following: source tracing suggestions, littering hotspot warnings, recycling vehicle dispatch instructions, and patrol route optimization results. The regulation response information is then formatted according to different regulation task types to obtain the dynamic regulation task processing results for construction waste.

[0041] In this embodiment, multi-source construction waste monitoring data is acquired, and a training dataset is constructed based on the multi-source construction waste monitoring data and according to a preset adjustment task type. The training dataset is cleaned and augmented to obtain an augmented training dataset. Then, the augmented training dataset is used to pre-train and supervised fine-tuning a large language model. A reward learning strategy is combined to optimize the large language model, and a temporal IoT prediction model is trained simultaneously. Next, a reinforcement learning algorithm is used to update the strategy of the large language model. The output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data are integrated to generate the dynamic adjustment task processing results of construction waste, thereby realizing dynamic control of the generation and flow of urban construction waste.

[0042] For ease of understanding, in a preferred embodiment, the construction waste dynamic monitoring data processing method provided in this application is applied to a city's intelligent construction waste supervision platform to dynamically monitor and regulate the generation, dumping, transportation, and inspection of decoration construction waste throughout the city.

[0043] In this embodiment, the intelligent construction waste monitoring platform first acquires multi-source construction waste monitoring data. Specifically, the platform obtains target unstructured text data such as decoration registration documents and construction unit registration information from business systems such as urban management, housing and construction, and sanitation; target time-series IoT data from video surveillance equipment, vehicle GPS terminals, and personnel positioning terminals deployed in various communities, construction sites, and along roads; target map POI data, including spatial points of interest information such as communities, construction sites, disposal sites, and road nodes, from the urban basic geographic information system; and auxiliary decision-making data, including past penalty records, complaint records, historical scheduling plans, and their implementation effects, from the historical management decision-making system. The platform uniformly aggregates and preprocesses the above-mentioned multi-source construction waste monitoring data to form the basic data source for subsequent model training and inference.

[0044] In this embodiment, the platform constructs a training dataset based on multi-source construction waste monitoring data and according to preset adjustment task types. Specifically, for the source tracing task, the platform combines the renovation registration texts of the confirmed responsible parties in historical cases, the corresponding POI features of the renovation sites, and the garbage feature descriptions extracted from video surveillance into training samples, and uses the responsible parties who actually generate the garbage as labels to form a sample set for training the source tracing model. For the littering hotspot time-series prediction task, the platform constructs historical littering time-series data based on past littering reports, and combines it with the distribution of renovation site POIs, construction progress time-series data, and meteorological data within the same time period. It uses the probability of littering in the target area and the time period within a future time window as labels to form a time-series training sample for predicting littering hotspots. For the recycling vehicle dispatching task, the platform jointly constructs training samples with the littering hotspot prediction results, the status time-series data of each recycling vehicle, and the road POI traffic time-series data to learn dispatching strategies under different traffic and garbage generation situations. For the patrol route optimization task, the platform combines the high-risk POI distribution time-series data, the patrol personnel location time-series data, and road POI accessibility information to form a training sample for patrol route optimization to guide urban management personnel in planning patrol routes.

[0045] In this embodiment, to enhance the model's adaptability to complex real-world scenarios, the platform augments the constructed training dataset. Specifically, the platform performs synonym replacement or sentence rewriting on renovation registration text to generate semantically equivalent but differently expressed texts such as renovation entrustment instructions and project scope descriptions, thereby enriching the text samples for the source tracing task. Time-series IoT data undergoes time shifting and rate adjustment, such as simulating earlier or later construction progress and fluctuating vehicle arrival times, to create different time-mode scenarios for waste generation and collection. Neighborhood replacement is performed on map POI data, such as replacing real construction site locations with POIs of adjacent roads or neighborhoods, constructing samples with slightly offset spatial distribution but similar risk characteristics to form diverse spatial feature samples. Through these augmentation operations, the coverage of the training data is significantly expanded while ensuring data rationality, resulting in an augmented training dataset.

[0046] In this embodiment, the platform utilizes the augmented training dataset to pre-train and supervised fine-tune the large language model. Specifically, the platform inputs the decoration registration text description into the text encoding module of the large language model, encoding the description to obtain a text representation vector reflecting semantic information such as decoration type, construction scale, time schedule, and characteristics of the construction unit. Based on the text representation vector, the platform generates a corresponding probability distribution of the generating entity at the output layer of the large language model, indicating that the project is more likely to be carried out by the homeowner themselves, a small team, a professional decoration company, or a specific blacklisted entity. The platform uses a cross-entropy loss function and a regularization term to supervise the training of the target entity prediction results. That is, it uses verified generating entity information from historical cases as supervision labels, calculates the cross-entropy loss between the predicted probability distribution and the true label, adds a regularization term based on model parameters, performs gradient backpropagation, and updates the parameters of the large language model to improve the accuracy of source tracing. At the same time, the platform synchronously trains a time-series IoT prediction model based on the amplified training dataset. The time-series IoT prediction model adopts a hybrid structure of LSTM and attention mechanism. By inputting time-series data of video surveillance recognition and construction progress, it learns the patterns of garbage generation and littering behavior at different times and construction stages, and then generates prediction results of the probability of littering and the amount of garbage generated in the future target area.

[0047] In this embodiment, to further improve the overall performance of the large language model in multi-task scenarios, the platform introduces a reward learning strategy during training to update the parameters of the large language model. Specifically, the platform constructs a cumulative reward function that includes source tracing accuracy, littering hotspot prediction accuracy, and recycling vehicle dispatch efficiency. This quantifies the performance of the above tasks into a unified reward signal. Source tracing accuracy reflects the accuracy of identifying the responsible party; littering hotspot prediction accuracy reflects the time and location accuracy of the platform's warnings; and recycling vehicle dispatch efficiency reflects comprehensive indicators such as the time consumption and path length for vehicles to complete the cleanup task within a given time. The platform performs gradient updates based on the cumulative reward function, using the cumulative reward as a reinforcement signal to optimize the strategy part of the large language model. This results in a better overall effect when considering multiple business objectives.

[0048] In this embodiment, the platform further performs policy updates on the large language model based on reinforcement learning algorithms to adapt to the dynamically changing urban management environment. The platform uses a reinforcement learning algorithm based on proximal policy optimization to perform update operations on the policy network of the large language model. During the actual deployment phase, it continuously collects the decisions output by the model in the real environment and their feedback effects. By comparing the probability distribution of actions before and after the policy update, the platform limits the magnitude of the policy update to ensure the stability of the model during iterative training and avoid policy oscillations or performance degradation.

[0049] In this embodiment, when the platform receives new construction waste monitoring data during operation, it first generates task text reasoning results for the current scenario using a trained large language model. This includes analysis of potential entities and textual descriptions of abnormal behaviors. Simultaneously, the temporal IoT prediction model outputs the probability of littering and the amount of waste generated in each area over several future time periods based on the latest video surveillance identification time-series data and construction progress time-series data. Furthermore, the platform extracts spatial feature information of relevant areas from map POI data, including spatial attributes such as road accessibility, density distribution of surrounding construction sites and residential areas, and distance to disposal sites. The platform performs fusion processing on the output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data. The text reasoning features, temporal prediction features, and spatial features are concatenated into a comprehensive input feature, which is then input into a pre-defined adjustment decision generation network to generate the dynamic adjustment task processing results for construction waste. The results of dynamic management of construction waste can include: source tracing suggestions for suspected littering originating communities or construction sites, early warnings of potential littering hotspots in high-risk areas, specific dispatch instructions for recycling vehicles (including departure time, service area, and route), and optimized patrol routes for urban management personnel. Supervisory personnel execute on-site inspections, vehicle dispatch, and route planning based on the management results output by the platform, thereby achieving dynamic and refined control over the generation and flow of construction waste throughout the city.

[0050] Figure 4 This is a schematic diagram of the structure of a construction waste dynamic monitoring data processing system according to an example embodiment of this application. Figure 4 As shown, the construction waste dynamic monitoring data processing system 500 provided in this embodiment includes: The acquisition module 510 is used to acquire multi-source construction waste monitoring data, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data. The augmentation module 520 is used to construct a training dataset based on the multi-source construction waste monitoring data and according to a preset adjustment task type, and to clean and augment the training dataset to obtain an augmented training dataset. The training module 530 is used to pre-train and supervised fine-tuning the large language model using the amplified training dataset, combine a reward learning strategy to optimize the large language model, and simultaneously train the temporal IoT prediction model. The generation module 540 is used to perform policy updates on the large language model based on the reinforcement learning algorithm, and to integrate the output of the large language model, the output of the time-series IoT prediction model, and the spatial features of the map POI data to generate the processing result of the construction waste dynamic adjustment task.

[0051] Figure 5 This is a schematic diagram of the structure of an electronic device according to an example embodiment of this application. For example... Figure 5 As shown, the electronic device 600 provided in this embodiment includes: a processor 601 and a memory 602; wherein: Memory 602 is used to store computer programs, and the memory may also be flash memory.

[0052] Processor 601 is used to execute the execution instructions stored in the memory to implement the various steps in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0053] Alternatively, the memory 602 can be either standalone or integrated with the processor 601.

[0054] When the memory 602 is a device independent of the processor 601, the electronic device 600 may further include: Bus 603 is used to connect the memory 602 and the processor 601.

[0055] This embodiment also provides a readable storage medium storing a computer program, which, when executed by at least one processor of an electronic device, enables the electronic device to perform the methods provided in the various embodiments described above.

[0056] This embodiment also provides a program product including a computer program stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to cause the electronic device to perform the methods provided in the various embodiments described above.

[0057] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0058] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for processing dynamic monitoring data of construction waste, characterized in that, include: Acquire multi-source construction waste monitoring data, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data; Based on the multi-source construction waste monitoring data and a training dataset constructed according to the preset adjustment task type, the training dataset is cleaned and augmented to obtain the augmented training dataset. The large language model is pre-trained and supervised fine-tuned using the augmented training dataset, and a reward learning strategy is combined to optimize the large language model, while a temporal IoT prediction model is trained simultaneously. The large language model is updated based on a reinforcement learning algorithm. The output of the large language model, the output of the time-series IoT prediction model, and the spatial features of the map POI data are integrated to generate the dynamic adjustment task processing result of construction waste.

2. The method according to claim 1, characterized in that, The step of constructing a training dataset based on the multi-source construction waste monitoring data and according to preset adjustment task types includes: For the source tracing task, a training sample is constructed that includes decoration registration text, decoration site POI features and video surveillance garbage feature descriptions. The label of the training sample is the generating entity information. For the task of predicting littering hotspots, a training sample is constructed that includes historical littering time series data, POI distribution of construction sites, construction progress time series data, and meteorological data. The labels of the training sample are the probability of littering in the future target area and the time period. For the task of dispatching recycling vehicles, a training sample is constructed that includes prediction results of littering hotspots, time-series data of recycling vehicle status, and time-series data of road POI traffic. To optimize patrol routes, a training sample was constructed that includes time-series data on the distribution of high-risk points of interest (POIs), time-series data on the locations of patrol personnel, and accessibility data for road POIs.

3. The method according to claim 1, characterized in that, The augmentation of the training dataset includes: Perform synonym replacement or sentence rewriting on the decoration filing text; Perform time shifting and rate adjustment on time-series IoT data; Neighborhood replacement is performed on map POI data to generate diverse spatial feature samples.

4. The method according to claim 1, characterized in that, The pre-training and supervised fine-tuning of the large language model using the augmented training dataset includes: The text description of the decoration filing is encoded to obtain a text representation vector; The subject probability distribution is generated based on the text representation vector; The prediction results of the target entity are trained under supervision using the cross-entropy loss function and regularization term to improve the accuracy of source tracing.

5. The method according to claim 1, characterized in that, The reward learning strategy is used to perform parameter updates on the large language model, and the reward learning strategy includes: Construct a cumulative reward function that includes the accuracy of source tracing, the accuracy of litter hotspot prediction, and the efficiency of recycling vehicle dispatching; Gradient updates are performed based on the cumulative reward function to improve the overall performance of the model in multi-task scenarios.

6. The method according to claim 1, characterized in that, The time-series IoT prediction model adopts a hybrid structure of LSTM and attention mechanism. By inputting time-series data of video surveillance recognition and construction progress, it generates prediction results of the probability of littering and the amount of garbage generated in the future target area.

7. The method according to claim 1, characterized in that, The policy update for the large language model based on the reinforcement learning algorithm includes: A reinforcement learning algorithm based on proximal policy optimization is used to perform update operations on the policy network of a large language model.

8. A dynamic monitoring data processing system for construction waste, characterized in that, include: The acquisition module is used to acquire multi-source construction waste monitoring data, which includes target unstructured text data, target time-series IoT data, target map POI data, and auxiliary decision-making data. The augmentation module is used to construct a training dataset based on the multi-source construction waste monitoring data and according to a preset adjustment task type, and to clean and augment the training dataset to obtain an augmented training dataset. The training module is used to pre-train and supervised fine-tuning the large language model using the augmented training dataset, combining a reward learning strategy to optimize the large language model, and simultaneously training the temporal IoT prediction model. The generation module is used to perform policy updates on the large language model based on reinforcement learning algorithms, and to integrate the output of the large language model, the output of the temporal IoT prediction model, and the spatial features of the map POI data to generate the processing results of the construction waste dynamic adjustment task.

9. An electronic device, characterized in that, include: processor; as well as, Memory for storing the executable instructions of the processor; The processor is configured to execute the method of any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images