Air monitoring node intelligent deployment method based on crowd sensing

Abstract

The invention belongs to the cross technical field of crowd sensing, air monitoring and artificial intelligence, discloses an intelligent air monitoring node deployment method based on crowd sensing, and aims to solve the problems of insufficient node deployment density, response lag, monitoring blind areas and the like in an existing air quality monitoring system. Firstly, crowd sensing information related to air pollution is extracted from multi-source data such as social media and a public reporting platform, pollution events and spatial positions of the pollution events are recognized through natural language processing and image recognition technologies, and a dynamic pollution sensing thermodynamic diagram is constructed. On the basis, a deep reinforcement learning algorithm framework is designed, factors such as the perception coverage rate, the deployment cost and the communication connectivity are comprehensively considered, and an optimal deployment scheme of the air monitoring nodes is learned and output. The intelligent deployment method designed by the invention has self-adaptive capability, and dynamically adjusts the deployment strategy of the spatial quality monitoring nodes according to the crowd sensing data, thereby improving the response capability of the system to sudden pollution events.

Description

Technical Field

[0001] The present invention belongs to the cross-technical field of crowd intelligence perception, air monitoring and artificial intelligence, and specifically relates to an intelligent deployment method of air monitoring nodes based on crowd intelligence perception. Background Art

[0002] With the acceleration of urbanization and the increase of industrial emissions, air pollution is becoming increasingly serious and has become a key factor affecting public health and urban sustainable development. In order to achieve real-time perception and refined management of air quality, building a high-coverage, high-precision air monitoring network has become an important research direction in the field of environmental monitoring.

[0003] Traditional air monitoring methods mainly rely on fixed monitoring stations built by the government. Although they have high-precision sensing capabilities, they are difficult to meet the needs of air quality perception in large-scale, dynamically changing environments due to high construction costs and limited coverage. To this end, in recent years, researchers have proposed the use of low-cost sensor nodes (such as portable air sensors, mobile monitoring equipment, etc.) for auxiliary monitoring to improve the spatial resolution of monitoring.

[0004] At the same time, the development of the Internet of Things, artificial intelligence and data mining technology in recent years has provided new opportunities for environmental monitoring. As an emerging perception paradigm, crowd sensing can reflect the occurrence and spatial distribution of pollution incidents from the side by collecting text, pictures, location information and reporting records from the public on social media. It has the advantages of wide coverage, fast response speed and low deployment cost, providing rich temporal and spatial information supplements for environmental monitoring, and helping to reveal traditional monitoring blind spots and sudden pollution hotspots.

[0005] However, existing research still faces the following challenges in combining crowd-sensing information with monitoring node deployment: 1) Insufficient data acquisition and understanding: Most current systems only use structured location data or user feedback numbers, and do not fully explore the semantic information of pollution events contained in unstructured text and images in social media, resulting in inaccurate identification of pollution hotspots; 2) Deployment decision methods are static and inefficient: Existing deployment algorithms are mostly based on heuristic search or offline optimization, which makes it difficult to adjust deployment strategies in real time in dynamic pollution evolution scenarios; 3) It is difficult to balance coverage, cost and communication constraints: How to reduce the number of node deployments and communication costs while ensuring perception effects and maintaining stable system operation is a multi-objective optimization problem that needs to be solved urgently.

[0006] In response to the above technical challenges, the present invention aims to design an intelligent deployment method for air monitoring nodes based on crowd intelligence perception. By integrating multi-source crowd intelligence perception data and combining natural language processing, computer vision and deep reinforcement learning algorithms, an efficient deployment strategy and dynamic scheduling mechanism driven by perception data can be implemented, thereby achieving efficient utilization of monitoring resources and timely response to pollution hotspots. Summary of the Invention

[0007] The present invention provides an intelligent deployment method for air monitoring nodes based on crowd-sensing, which integrates natural language processing, computer vision, and deep reinforcement learning technologies. By extracting pollution-related information from social media and public reported data, it realizes the accurate identification of pollution hotspots, and designs a dynamically optimized air monitoring node deployment strategy based on this. This method not only improves the response ability and coverage effect of the monitoring system to pollution events, but also significantly reduces the deployment cost and communication energy consumption, and is applicable to the construction of intelligent sensing systems in large-scale complex urban environments.

[0008] Specifically, the present invention is implemented through the following technical solutions:

[0009] An intelligent deployment method for air monitoring nodes based on crowd-sensing includes the following steps:

[0010] Step S1: Based on social media texts, images, geolocation information, and public reported records, extract pollution keywords and geographical tags in the text through natural language processing technology, and use an image recognition model to identify pollution-related visual features, such as smoke, emissions, haze, etc., to achieve semantic-level recognition and spatial positioning of pollution events. Step S2: Construct a spatio-temporal mapping mechanism for pollution perception, perform geographical gridification and time aggregation processing on the extracted pollution events to form a dynamically updated pollution perception heat map; the pollution perception heat map reflects the high-incidence areas and their evolution trends of pollution events in the city, providing a perception basis for the node deployment strategy. Step S3: Based on the pollution heat map and environmental constraint information, by defining a multi-objective joint reward function such as maximizing coverage, minimizing deployment cost, and communication connectivity constraints, use a reinforcement learning agent to perform policy iteration optimization in a complex space state to generate a deployment plan with global optimality and output the node deployment points. Step S4: According to the real-time air data transmitted back by the sensor nodes and the newly added crowd-sensing information, dynamically update the pollution heat map and the reinforcement learning model to realize the online evolution of the deployment strategy and improve the adaptive ability and efficiency of the system in long-term operation.

[0011] In one embodiment, the image recognition model in step S1 adopts a pollution feature detection model based on a deep convolutional neural network to classify and label pollution phenomena in urban images, and further combines the position information in the images to accurately locate pollution events.

[0012] In one embodiment, a sliding time window mechanism is adopted in step S2 for time aggregation, and a spatial mapping algorithm based on a tree-like data structure combined with geographical grid division is used to generate a pollution heat map, and the heat value is jointly calculated through the pollution event frequency, credibility, and public report confidence weight.

[0013] In one embodiment, in step S3, a deep reinforcement learning model based on the combined policy gradient method and value function estimation is adopted. Using environmental state information such as pollution intensity distribution, node deployment density, and communication radius as inputs, through the collaborative training of the policy network and the value function network, the optimal deployment decision is output; the connectivity between nodes is used as a limiting term in policy optimization by constructing a communication graph and setting connection constraints.

[0014] In one embodiment, in step S4, an online reinforcement learning mechanism is introduced. By continuously interacting with the environment to correct the policy network parameters, and combining the actual observation data of the deployed nodes, the deployment structure is dynamically adjusted to achieve phased node migration, merging, and deployment density adjustment operations.

[0015] The present invention has the following advantages and beneficial effects:

[0016] 1. The present invention proposes a pollution hot spot recognition mechanism that integrates crowd-sourced sensing data. By using the joint analysis means of text, images, and geographical information, it realizes high-precision and multi-dimensional perception of pollution events, providing a more real and dynamic pollution distribution information basis for the deployment strategy. The present invention is applicable to large-scale urban monitoring scenarios and can be widely deployed in multiple fields such as smart cities, environmental governance, and public health protection.

[0017] 2. The present invention designs an intelligent deployment model based on a pollution perception heat map. By adopting deep reinforcement learning technology, it realizes the automatic generation of optimal deployment points in a multi-constraint environment, significantly improving the coverage rate and resource utilization efficiency of the monitoring system, and reducing the deployment and communication costs.

[0018] 3. The online update mechanism of the node deployment strategy designed by the present invention supports the dynamic adjustment ability of the monitoring network during operation. According to the actual environmental changes and perception feedback, it flexibly migrates or adjusts the node deployment, with strong long-term stability and adaptability.

[0019] The technical solutions / features disclosed above are intended to summarize the technical solutions and technical features described in the specific implementation part. Therefore, the scope recorded may not be exactly the same. However, these new technical solutions disclosed in this part also belong to a part of the numerous technical solutions disclosed in the present invention document. The technical features disclosed in this part, the technical features disclosed in the subsequent specific implementation part, and the parts of the drawings not clearly described in the specification are disclosed in a more reasonable combination to disclose more technical solutions.

[0020] The technical solutions formed by combining all the technical features disclosed at any position of the present invention are used to support the generalization of the technical solutions, the modification of the patent document, and the disclosure of the technical solutions. Brief Description of the Drawings

[0021] Figure 1 It is an example flowchart for explaining the intelligent deployment method of air monitoring nodes based on crowd sensing proposed by the present invention;

[0022] Figure 2 It is an example for explaining the deep reinforcement learning network based on proximal policy optimization designed by the present invention and its key elements. Detailed implementation manners

[0023] Since it is impossible to exhaustively describe all alternative solutions, the key points of the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. For other technical solutions and details not disclosed in detail below, generally they are all technical objectives or technical features that can be achieved by conventional means in the art. Due to space limitations, the present invention will not introduce them in detail.

[0024] Unless it means division, " / " at any position in the present invention represents logical "or". The serial numbers such as "first", "second", etc. at any position in the present invention are only used as distinguishing marks in description, and do not imply an absolute order in time or space, nor do they imply that the terms with such serial numbers must be different references from the same terms with other attributives.

[0025] The present invention will describe the key points for combining various different specific embodiments, and these key points will be combined into various methods and products. In the present invention, even if only the key points described when introducing the method / product solution are meant, it means that the corresponding product / method solution also clearly includes this technical feature.

[0026] When it is described in the present invention that there exists or includes a certain step, module, or feature at any position, it does not imply that this existence is an exclusive and unique existence. Those skilled in the art can completely obtain other embodiments by supplementing other technical means according to the technical solutions disclosed in the present invention. Based on the key points described in the specific embodiments of the present invention, those skilled in the art can completely apply means such as replacement, deletion, addition, combination, and order swapping to certain technical features to obtain a technical solution that still follows the concept of the present invention. These solutions that do not deviate from the technical concept of the present invention are also within the protection scope of the present invention.

[0027] Such as Figure 1, this application designs an intelligent deployment method for air monitoring nodes based on crowd-sensing. This deployment method includes: 1) a multi-source pollution event extraction mechanism based on crowd-sensing, which fuses unstructured sensing data from social media, user reporting platforms, etc., and extracts relevant information on air pollution events through natural language processing and image recognition technologies; 2) the construction of an air pollution perception heat map, which performs spatial mapping and temporal aggregation on the extracted pollution event information, and combines historical pollution distribution and perception density indicators to generate a dynamic air pollution perception heat map to depict potential pollution hotspots and their evolution trends; 3) a monitoring node deployment strategy based on deep reinforcement learning, which uses the air pollution heat map as the environmental state input, and designs a reward function in combination with factors such as deployment cost, node coverage, and connectivity to guide the deployment strategy to achieve dynamic optimal node layout in a complex urban space; 4) an online update mechanism for the node deployment strategy, which dynamically adjusts the deployment strategy to adapt to the spatio-temporal evolution characteristics of pollution distribution by real-time updating crowd perception data and node monitoring results, and realizes the intelligent, autonomous, and efficient operation of the air quality monitoring network. The above technical methods will be introduced in detail below.

[0028] The first stage: Design a multi-source pollution event extraction mechanism based on crowd-sensing. In this stage, by fusing unstructured sensing data from social media, user reporting platforms, etc., natural language processing and image recognition technologies are used to extract event information related to air pollution. The characteristics of this process are as follows:

[0029] 1. First, collect multi-source sensing data from social media platforms, mobile applications, and user reporting platforms. These data include information such as text, images, and locations. Assume the original sensing data set is , which includes social media text data , image data uploaded by users , and relevant geographical location information . For the image data uploaded by users , analyze it through computer vision models such as YOLOv11 and ResNet to identify the pollution type and degree therein. For the text data , analyze it through natural language processing technology, and based on the weighted value of term frequency and inverse document frequency index, extract important keywords of air pollution events. Given any word in the text of the spatial pollution event, the weighted value TF-IDF(j) is calculated as follows:

[0030] Among them, is the term frequency of the word in the document, and is the inverse document frequency of the word .

[0031] 2. Extract information related to pollution events from multi-source perception data through text analysis and image recognition technologies, and associate the event information with corresponding timestamps and location data to form a pollution event dataset. Any air pollution event , where represents the capture time of the air pollution event, represents the location where the air pollution event occurred, and represents the air pollution intensity. Each event in the dataset is labeled by spatial coordinates and pollution intensity, providing data support for the subsequent construction of the heat map.

[0032] 3. Since the information obtained from multiple data sources may contain noise or irrelevant data, data cleaning is required to filter out high-quality pollution event data. At this stage, a clustering-based filtering method is used to identify and remove abnormal events. Given a pollution event dataset , identify the clusters related to pollution events; assume is the th cluster, is the size of the cluster, then the cluster center is calculated as follows:

[0033] ;

[0034] Second stage: In this stage, spatial mapping and time aggregation will be performed on the structured pollution event data extracted in the first stage, and combined with historical pollution distribution characteristics and perception density to generate a dynamic air pollution perception heat map to reflect pollution hotspots and their evolution trends. The characteristics of this stage are as follows:

[0035] 1. First, perform spatial grid processing on the target monitoring area, divide the urban area into grid cells of equal size for spatial classification of pollution events. Given the two-dimensional space of the target area, this space can be divided into grid cells, and each grid is . According to the air pollution event set obtained in the first stage of the present invention, where any air pollution event , represents the capture time of the i-th air pollution event, represents the location where the i-th air pollution event occurred, and represents the i-th air pollution intensity. A pollution event mapping function can be defined to map air events to the geographical location grid cells of the target space.

[0036] 2. On each grid cell, the events are grouped according to the time window Perform aggregation processing to count the number and intensity of pollution events in each grid. Let the k-th time window be , denote the capture time of the k-th air pollution event. Then, the pollution intensity in the grid and the time window can be defined as:

[0037] where is the set of pollution events that fall into the grid within the time window , and is the pollution intensity of the pollution event .

[0038] 3. Based on the calculated spatio-temporal aggregated pollution intensity, to further improve the stability and prediction ability of the heat map, historical pollution data and the current perceived density are introduced and fused to generate an enhanced pollution heat value . Among them, , , are weight parameters with a weighted sum of 1. is the long-term pollution average calculated from historical monitoring data, and the current perceived density is the result of perceived density normalization, and its calculation method is .

[0039] The third stage: Design an air quality monitoring node deployment strategy based on deep reinforcement learning. Combining the spatial distribution characteristics of the pollution heat map, achieve the adaptive and efficient layout of air quality monitoring nodes. This method models the node deployment problem as a Markov decision process, uses the Proximal Policy Optimization (PPO) algorithm as the core learning framework, and perceives the urban environment through the policy network and outputs deployment actions, thereby gradually optimizing the overall performance of the monitoring network. The characteristics of this process are as follows:

[0040] 1. Construct the state space which is jointly composed of the air pollution heat map , the deployed monitoring nodes and the geographical information of the target area. The state is defined as ; Each action in the action space represents deploying a spatial quality monitoring node at the candidate position in the target space; The state transition probability It is obtained by integrating the pollution diffusion model based on the Gaussian regression process and the human flow dynamic distribution model; the reward function is used to measure the change in system performance after executing the deployment action, and its calculation method is , represents the contribution of the current deployment action in terms of pollution coverage rate, represents the deployment cost of the quality monitoring node, represents the connectivity between each air quality node, represents the newly deployed node the number of pollution hotspots covered, represents the total number of air pollution hotspots in the target area obtained in the first stage.

[0041] 2. Design of a deep reinforcement learning network based on proximal policy optimization. This deep reinforcement learning network uses proximal policy optimization as the training backbone algorithm and uses a policy gradient - value estimation structure to construct the behavior policy of the deep reinforcement learning network and value estimation . Specifically, the behavior policy indicates the probability distribution of the learning network taking action in state s, which is implemented by a multi - layer neural network including convolutional layers, flattening layers, fully connected layers, etc. The corresponding value estimation estimates the long - term expected return of the state through a regression model, and the state space set is , where are the neural network weights and bias parameters involved in the policy function of this deep reinforcement learning network. The loss function in the policy training process of this algorithm is the clipped objective function based on proximal policy optimization , the value network optimization objective and the policy entropy sum, is defined as follows: ; ;

[0042] Among them, at the th step of the interaction between the agent and the environment, the state of the deep reinforcement learning agent is , and the action decision taken by the agent in this state is . Let be the policy of the agent in the previous step, then is the ratio of the new and old policies, is the deployment advantage estimation of the policy at the th step obtained based on generalized advantage estimation, For the clipping coefficient of the proximal policy optimization, the optimization objective of the value network is the mean squared error of the minimum expected return of the deployment policy , is the actual return obtained by the deep reinforcement learning network in the actual environment, Figure 2 shows the deep reinforcement learning network based on proximal policy optimization and its key elements.

[0043] Phase 4: Design an online adaptive update mechanism to dynamically adjust the node deployment policy by monitoring and perceiving data updates in real time. Whenever new contaminated data arrives, the deep reinforcement learning algorithm recalculates and optimizes the deployment location and scheduling policy of the nodes according to the current environmental state and node deployment situation, so as to cope with the spatio-temporal evolution of pollution events.

Claims

1. An intelligent deployment method for air monitoring nodes based on crowd-sourced sensing, characterized in that, It includes the following steps: Step S1: Collection of spatial pollution events based on crowdsensing. Through social media platforms and public reporting systems, multi-source sensing data containing text, images, and geolocation information is obtained. Natural language processing and computer vision algorithms are used to extract semantic features and location information related to air pollution events from it, forming structured pollution perception samples; Step S2: Construction of an air pollution perception heat map. The extracted pollution event information is subjected to spatial mapping and temporal aggregation. Combining historical pollution distribution and perception density indicators, a dynamic air pollution perception heat map is generated to depict potential pollution hotspots and their evolution trends; Step S3: Construction of a deep reinforcement learning model based on the heat map and generation of a deployment strategy. A state-action-reward mechanism is designed, where the state includes the current heat map information and the node deployment status, the actions include node deployment, migration, and recovery, and the reward function comprehensively considers the perception coverage rate, deployment cost, and communication connectivity. The optimal deployment plan of air monitoring nodes is output through the deep reinforcement learning algorithm; Step S4: Implementation of the deployment strategy and online update. According to the deployment plan generated in Step S3, combined with the subsequent received real-time crowdsensing data, the deep reinforcement learning model is fine-tuned online and the parameters are updated to dynamically adjust the deployment strategy of spatial quality monitoring nodes in different time periods and different scenarios.

2. The method according to claim 1, wherein: In the said Step S1, the natural language processing includes pollution keyword recognition, entity naming recognition, and geographical location parsing, and the computer vision processing uses a convolutional neural network recognition model containing pollution image features to extract semantic features related to pollution events from images.

3. The method according to claim 1, characterized in that: In the said Step S2, the construction of the air pollution heat map is carried out based on a grid-based space. Specifically, the monitoring area is divided into spatial cells of a fixed size, and a pollution density indicator is generated based on the number of pollution events, perception intensity, and public reporting confidence level in each cell per unit time. A dynamically evolving pollution heat map is formed through a time sliding window mechanism.

4. The method according to claim 1, wherein: In the said Step S3, the deep reinforcement learning algorithm adopts the proximal policy optimization method. The state space of the deep reinforcement learning model includes the current heat map, the positions and perception ability distributions of the deployed nodes, the action space includes node deployment, migration, and recovery operations, and the reward function is a weighted combination function, including a coverage rate indicator, a deployment cost function, and a network connectivity constraint penalty term.

5. The method according to claim 1, characterized in that: In the said Step S4, the online update mechanism includes a real-time feedback acquisition module and a policy fine-tuning module. The former is used to update the heat map from newly added crowdsensing data, and the latter uses the online training mechanism in reinforcement learning to update the parameters of the deep reinforcement learning policy, so as to dynamically adapt to changes in pollution distribution.

Images