A Deep Learning-Based Method and System for Environmental Gas Monitoring

Abstract

This invention discloses a deep learning-based method and system for environmental gas monitoring. The method includes: constructing an environmental monitoring mobile robot to acquire obstacle image data of the target environment and performing data augmentation processing to obtain enhanced wooden obstacle image data; introducing a TAA strategy to train a YOLOv8 target detection model to construct an improved YOLOv8 target detection model; performing recognition processing on the enhanced wooden obstacle image data to obtain wooden furniture information data of the target environment; controlling the environmental monitoring mobile robot to perform global movement to monitor gases in the target environment and obtain the monitoring results of harmful gases in the target environment. This invention can accurately identify the location of wooden furniture and effectively monitor the concentration of harmful gases in real time, ensuring the safety and health of the home environment. As a deep learning-based environmental gas monitoring method and system, this invention can be widely applied in the field of environmental gas monitoring technology.

Description

Technical Field

[0001] This invention relates to the field of environmental gas monitoring technology, and in particular to an environmental gas monitoring method and system based on deep learning. Background Technology

[0002] Wooden furniture is popular in modern homes and offices for its aesthetics and durability. However, this furniture may release harmful gases such as formaldehyde and TVOCs (total volatile organic compounds), posing a potential threat to the health of residents. Although traditional air quality monitoring devices, such as handheld detectors or fixed monitoring stations, can measure indoor air quality, they lack automation and positioning capabilities. Therefore, they cannot accurately identify pollution sources or continuously monitor changes in air quality across the entire environment. Most related technologies lack automatic navigation and positioning functions, resulting in incomplete monitoring results and an inability to track the source and concentration changes of harmful gases in real time. Furthermore, these devices often rely on manual operation, leading to low efficiency and an inability to adapt to complex indoor environments, especially in scenarios with dense furniture, where they struggle to accurately detect harmful gases released from furniture surfaces. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide an environmental gas monitoring method and system based on deep learning, which can accurately identify the location of wooden furniture and monitor the concentration of harmful gases in real time, thereby ensuring the safety and health of the home environment.

[0004] The first technical solution adopted in this invention is: an environmental gas monitoring method based on deep learning, comprising the following steps:

[0005] An environmental monitoring mobile robot was constructed to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data;

[0006] The TAA strategy is introduced to train the YOLOv8 object detection model, thus constructing an improved YOLOv8 object detection model;

[0007] Based on the improved YOLOv8 target detection model, the enhanced wooden material obstacle image data is identified and processed to obtain wooden furniture information data of the target environment;

[0008] Based on the information data of wooden furniture in the target environment, the environmental monitoring mobile robot is controlled to move globally to monitor the gas in the target environment and obtain the monitoring results of harmful gases in the target environment.

[0009] Furthermore, the environmental monitoring mobile robot specifically includes a fisheye camera, an ultrasonic distance sensor, and an ambient air quality detector, wherein:

[0010] The fisheye camera is used to acquire obstacle information data in the target environment;

[0011] The ultrasonic distance sensor is used to obtain the distance value between the environmental monitoring mobile robot and obstacles;

[0012] The ambient air quality detector is used to monitor the gas quality in the target environment.

[0013] Furthermore, the step of constructing an environmental monitoring mobile robot to acquire obstacle image data of the target environment and performing data augmentation processing to obtain enhanced wooden material obstacle image data specifically includes:

[0014] Construct an environmental monitoring mobile robot and a target scenario, wherein the target scenario includes several sampling points;

[0015] The fisheye camera of the environmental monitoring mobile robot collects sampling points in the target scene from multiple angles to obtain obstacle image data of the target environment.

[0016] The obstacle image data of the target environment is classified and processed to obtain obstacle image data of wooden materials and obstacle image data of non-wood materials;

[0017] Digital augmentation processing is performed on the image data of wooden obstacles to obtain augmented image data of wooden obstacles.

[0018] Furthermore, the step of performing digital augmentation on the wooden obstacle image data to obtain enhanced wooden obstacle image data specifically includes:

[0019] The image data of wooden obstacles is processed by digital cropping to obtain cropped image data of wooden obstacles;

[0020] The cropped wooden obstacle image data is scaled to obtain scaled wooden obstacle image data;

[0021] Exposure enhancement processing is performed on the scaled-down image data of wooden obstacles to obtain exposed-enhanced image data of wooden obstacles;

[0022] The enhanced wooden obstacle image data is obtained by randomly rotating and flipping the image data.

[0023] Furthermore, the step of introducing the TAA strategy to train the YOLOv8 object detection model and constructing the improved YOLOv8 object detection model specifically includes:

[0024] The YOLOv8 object detection model is pre-trained using a pre-training dataset to obtain the pre-trained YOLOv8 object detection model.

[0025] Based on the pre-trained YOLOv8 object detection model, a position loss function and a classification loss function are introduced to construct an optimized YOLOv8 object detection model.

[0026] The TAA strategy is introduced to dynamically adjust the ratio of positive to negative samples in the optimized YOLOv8 object detection model, resulting in an improved YOLOv8 object detection model.

[0027] Furthermore, the specific expression of the TAA strategy is as follows:

[0028] t = s α *u β

[0029] In the above formula, t represents the weighted score, s represents the prediction score of the category corresponding to the predicted box, u represents the intersection-union ratio between the predicted box and the ground truth box, and α and β represent hyperparameters.

[0030] Furthermore, the step of identifying and processing the enhanced wooden obstacle image data based on the improved YOLOv8 target detection model to obtain wooden furniture information data of the target environment specifically includes:

[0031] The enhanced image data of wooden obstacles is input into the improved YOLOv8 target detection model;

[0032] Based on the improved YOLOv8 target detection model backbone network module, feature extraction processing is performed on the enhanced wooden obstacle image data to obtain wooden obstacle image feature data;

[0033] Based on the improved YOLOv8 target detection model, the neck network module performs feature fusion processing on the feature data of wooden obstacle images to obtain fused wooden obstacle image feature data;

[0034] The head network module based on the improved YOLOv8 target detection model is used to classify and detect the feature data of the fused wooden material obstacle image to obtain the wooden furniture information data of the target environment.

[0035] Furthermore, the step of controlling the environmental monitoring mobile robot to move globally and monitor the gas in the target environment based on the information data of wooden furniture in the target environment, and obtaining the monitoring results of harmful gases in the target environment, specifically includes:

[0036] Based on the information data of wooden furniture in the target environment, the environmental monitoring mobile robot's ambient air quality detector collects gases in the target environment to obtain the monitoring results of harmful gases at the current sampling point;

[0037] The distance between the environmental monitoring mobile robot and obstacles is obtained by the ultrasonic distance sensor of the environmental monitoring mobile robot. Global movement planning is carried out to control the environmental monitoring mobile robot to move to the next sampling point to collect target environmental gases.

[0038] The process continues until all sampling points are visited to obtain the monitoring results of harmful gases in the target environment.

[0039] The second technical solution adopted in this invention is: an environmental gas monitoring system based on deep learning, comprising:

[0040] The first module is used to build an environmental monitoring mobile robot to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data;

[0041] The second module is used to introduce the TAA strategy to train the YOLOv8 object detection model and build an improved YOLOv8 object detection model.

[0042] The third module is used to identify and process the enhanced wooden material obstacle image data based on the improved YOLOv8 target detection model to obtain wooden furniture information data of the target environment.

[0043] The fourth module is used to control the environmental monitoring mobile robot to move globally and monitor the gas in the target environment based on the information data of wooden furniture in the target environment, and obtain the monitoring results of harmful gases in the target environment.

[0044] The beneficial effects of the method and system of this invention are as follows: This invention acquires obstacle image data of the target environment by constructing an environmental monitoring mobile robot and performs data augmentation processing, thereby increasing the diversity of the target environment data. Furthermore, it introduces the TAA strategy to train the YOLOv8 target detection model, and adopts the Task-AlignedAssigner (TAA) strategy to dynamically adjust the ratio of positive and negative samples, making the model pay more attention to high-quality prediction boxes. Based on the improved YOLOv8 target detection model, it performs recognition processing on the enhanced wooden material obstacle image data, which can achieve accurate detection and classification of targets in the image. Finally, based on the wooden furniture information data of the target environment, it controls the environmental monitoring mobile robot to move globally to monitor the gas in the target environment, effectively monitor the concentration of harmful gases in real time, and ensure the safety and health of the home environment. Attached Figure Description

[0045] Figure 1This is a flowchart of the steps of an environmental gas monitoring method based on deep learning according to the present invention;

[0046] Figure 2 This is a structural block diagram of an environmental gas monitoring system based on deep learning according to the present invention;

[0047] Figure 3 This is a schematic diagram of the wood identification results provided in a specific embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of formaldehyde detection data for wooden furniture provided in a specific embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram illustrating formaldehyde data from wood products in an office environment, provided by a specific embodiment of the present invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0051] Reference Figure 1 This invention provides a deep learning-based method for monitoring environmental gases, which includes the following steps:

[0052] S100. Construct an environmental monitoring mobile robot to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data;

[0053] First, it should be noted that the environmental monitoring mobile robot specifically includes a fisheye camera, an ultrasonic distance sensor, and an ambient air quality detector. The fisheye camera is used to acquire obstacle information data of the target environment; the ultrasonic distance sensor is used to acquire the distance value between the environmental monitoring mobile robot and the obstacle; and the ambient air quality detector is used to monitor the gas quality in the target environment.

[0054] In this embodiment, the robot is equipped with a fisheye camera and an ultrasonic distance sensor. The former captures images with a wide field of view, while the latter is used for environmental perception and obstacle detection. MasterPi features a 120-degree wide-angle telephoto camera with an automatic 180° surround view and an automatic tilt exceeding 180°. Combined with a flexible vehicle, the visual range reaches 360° omnidirectional coverage, ensuring the robot can acquire panoramic images of the indoor environment and provide high-quality image data for subsequent wooden furniture identification. Equipped with an IIC interface-based luminescent ultrasonic sensor, it not only achieves detection and obstacle avoidance without consuming excessive CPU resources but also allows for real-time adjustment of the automatic obstacle avoidance distance and ultrasonic emission color via a mobile app. This enables the robot to effectively perceive obstacles in its surroundings, achieve autonomous obstacle avoidance, and ensure safe movement in complex environments. A highly sensitive ambient air quality detector is integrated, capable of real-time monitoring of harmful gas concentrations such as formaldehyde, TVOC, and C6H6, while also displaying temperature and humidity data. This ensures the robot can monitor and report the concentration of harmful gases in the home environment in real time, promptly identifying potential health risks and providing users with a safer and healthier living environment.

[0055] Furthermore, it should be noted that an environmental monitoring mobile robot and a target scene are constructed, the target scene including several sampling points; the fisheye camera of the environmental monitoring mobile robot collects multi-angle images of the sampling points in the target scene to obtain obstacle image data of the target environment; the obstacle image data of the target environment is classified and processed to obtain wooden material obstacle image data and non-wood material obstacle image data; the wooden material obstacle image data is augmented to obtain augmented wooden material obstacle image data.

[0056] In this embodiment, a random sampling method is used to ensure data diversity and representativeness during the construction of the furniture dataset. First, random sampling is performed, with furniture randomly selected for data collection at each sampling point. This method helps avoid subjective bias and ensures the dataset accurately reflects the diversity of furniture on the market. Furthermore, to ensure data comprehensiveness, sampling is conducted in different environments (such as living rooms, bedrooms, and offices), covering furniture of different styles and materials. Multi-angle data collection is then performed on each piece of furniture, including front, side, back, and top views. This ensures the model can accurately identify furniture from different angles in practical applications. High-resolution cameras are used to capture images, ensuring image quality so the model can extract more detailed features. Finally, the data is classified into wood and non-wood categories. The wood dataset contains images of various wooden furniture, while the non-wood dataset contains other types of furniture (such as metal and plastic). This method allows the model to be trained to more accurately distinguish between wooden furniture and other types of furniture.

[0057] Furthermore, the YOLOv8 object detection algorithm was used in the model. To further improve the model's performance, this embodiment of the invention also employed data augmentation methods. First, data cropping was performed, randomly cropping a portion of the image to simulate furniture in different positions and sizes. Next, data scaling was performed, enlarging or shrinking the image to simulate the visual effect of the furniture at different distances. Then, data exposure enhancement was performed, adjusting the image's brightness and contrast to simulate scenes under different lighting conditions. Next, data rotation was performed, randomly rotating the image to simulate the visual effect of the furniture at different angles. Finally, data flipping was performed, horizontally or vertically flipping the image to increase data diversity.

[0058] S200. Introduce the TAA strategy to train the YOLOv8 object detection model and construct an improved YOLOv8 object detection model.

[0059] Specifically, the YOLOv8 object detection model is pre-trained using a pre-training dataset to obtain a pre-trained YOLOv8 object detection model. Based on the pre-trained YOLOv8 object detection model, a position loss function and a classification loss function are introduced to construct an optimized YOLOv8 object detection model. The TAA strategy is introduced to dynamically adjust the ratio of positive and negative samples in the optimized YOLOv8 object detection model to obtain an improved YOLOv8 object detection model.

[0060] In this embodiment, the YOLOv8 model is first pre-trained on a large dataset (such as COCO) to learn the features of general objects. The model is then fine-tuned on a collected furniture dataset to improve its accuracy in recognizing wooden furniture. Next, the loss function is optimized using a comprehensive loss function, including positional loss (box_loss) and classification loss (cls_loss), to ensure accuracy in both position and category. Finally, task alignment is performed using a Task-AlignedAssigner (TAA) strategy to dynamically adjust the ratio of positive to negative samples, making the model focus more on high-quality predicted boxes.

[0061] Further model evaluation and optimization were conducted. First, performance testing of the model was carried out for several days in different indoor environments to evaluate the robot's performance and stability in complex environments. Then, data analysis was performed to record the detection results under different spaces, heights, and concentrations of harmful gases to analyze the robot's accuracy and stability. Finally, the model was optimized and improved based on experimental results and user feedback to improve the accuracy and speed of recognition.

[0062] S300: Based on the improved YOLOv8 target detection model, the enhanced wooden material obstacle image data is identified and processed to obtain wooden furniture information data of the target environment;

[0063] Specifically, the enhanced wooden obstacle image data is input into the improved YOLOv8 target detection model; the backbone network module of the improved YOLOv8 target detection model performs feature extraction processing on the enhanced wooden obstacle image data to obtain wooden obstacle image feature data; the neck network module of the improved YOLOv8 target detection model performs feature fusion processing on the wooden obstacle image feature data to obtain fused wooden obstacle image feature data; the head network module of the improved YOLOv8 target detection model performs classification and detection on the fused wooden obstacle image feature data to obtain wooden furniture information data of the target environment, such as... Figure 3 As shown.

[0064] First, the structure of the improved YOLOv8 target detection model in this embodiment of the invention is described. The YOLOv8 network structure mainly consists of three parts: Backbone, Neck, and Head. These three parts work together to achieve the target detection function.

[0065] The backbone is responsible for extracting features from the input image. YOLOv8's backbone borrows the idea of ​​the CSP (Compact and Separate) module, but replaces the C3 module from YOLOv5 with the C2f module to achieve further weight reduction while maintaining the CSP concept. In addition, YOLOv8 also employs the SPPF (SPPFast) module to enhance feature extraction capabilities. In the backbone, convolutional modules (Conv) and C2f modules are used alternately. Features are extracted through convolutional operations, and feature fusion and dimensionality reduction are performed through the Bottleneck structure in the C2f module. Compared to the C3 module, the C2f module has fewer parameters and superior feature extraction capabilities, contributing to improved model performance and efficiency.

[0066] The Neck component is responsible for fusing feature maps of different scales extracted by the Backbone to enhance feature representation capabilities. YOLOv8's Neck employs a PANet (Path Aggregation Network) structure, a bidirectional pathway network that introduces a bottom-up path, making it easier for lower-level information to reach higher levels. In PANet, the feature maps output by the Backbone undergo upsampling and channel fusion to form feature maps of multiple scales. These feature maps are then fed into the DetectHead for object detection and classification.

[0067] The Head section is responsible for the final object detection and classification tasks. YOLOv8 adopts a decoupled Head structure similar to YOLOX, separating the regression and prediction branches, making network training and inference more efficient. In the decoupled Head, the regression branch uses the DFL (Distribution Focal Loss) strategy for coordinate prediction, transforming the coordinates into a distribution to address the problem of blurred boundaries. Simultaneously, the classification branch uses global average pooling to classify each feature map.

[0068] The YOLOv8 network architecture diagram illustrates its complete process from input image to output detection result. Through the collaborative work of the Backbone, Neck, and Head, YOLOv8 achieves accurate detection and classification of objects in images. Its innovations include replacing the C3 module with the C2f module, employing the SPPF module to enhance feature extraction capabilities, and introducing a DecoupledHead structure to improve training and inference efficiency. These improvements give YOLOv8 higher performance and flexibility in the field of object detection.

[0069] Furthermore, the improved YOLOv8 object detection model algorithm of this invention needs to be described. First, it should be noted that YOLOv8 automatically adjusts the shape it uses to predict the object's position, thus more accurately matching the shape of the actual object. Further, to reduce false detections, by optimizing these shapes, YOLOv8 can reduce false alarms, i.e., mistaking the background for an object.

[0070] Further, including the comprehensive loss function, YOLOv8 uses a special formula to measure prediction accuracy. This formula considers the correctness of object location and category, as well as regularization strategies. YOLOv8 also employs techniques to avoid overfitting, preventing the model from becoming too reliant on training data and unable to adapt well to new data. Among these, Task-Aligned Assigner (TAA) is a strategy that dynamically adjusts the ratio of positive to negative samples. Its core idea is to dynamically select positive samples based on classification and regression performance during training. This approach is called "task alignment" because it attempts to make the selection of positive samples more aligned with the needs of the final task.

[0071] To further explain the principles of task alignment:

[0072] TAA uses a weighted score to measure the alignment of each predicted bounding box with the ground truth (GT). The weighted score is determined by the classification score and the intersection-over-union (IoU) ratio between the predicted and GT boxes. The formula for calculating the weighted score is:

[0073] t = s α *u β

[0074] In the above formula, t represents the weighted score, s represents the prediction score of the category corresponding to the predicted box, u represents the intersection-union ratio between the predicted box and the ground truth box, and α and β represent hyperparameters.

[0075] The weighted score combines the classification score and the IoU value, reflecting the degree of task alignment. Hyperparameters α and β are set to adjust the relative importance of the classification score and IoU. In this description, the embodiments of the present invention do not directly mention these two hyperparameters, but instead calculate the weighted score directly using a formula.

[0076] Specific execution steps:

[0077] 1) Calculate the alignment score:

[0078] Based on the classification score of the predicted bounding box and the IoU between the predicted bounding box and the ground truth bounding box, the alignment score t of each predicted bounding box is calculated.

[0079] 2) Screening candidate positive samples:

[0080] Determine which predicted boxes have their center points inside the ground truth (GT) boxes; only these predicted boxes are eligible to be positive samples.

[0081] 3) Select positive samples:

[0082] From the candidate positive samples, select the top K predicted boxes as positive samples based on the alignment score t.

[0083] The remaining predicted boxes are considered negative samples.

[0084] By employing this strategy, TAA can help the model focus more on high-quality bounding boxes, namely those with high classification scores and good localization, thereby improving the overall performance of the model.

[0085] Finally, transfer learning is performed. The pre-trained model, YOLOv8, is first trained on a large dataset to learn how to recognize various types of objects, and then fine-tuned. Then, for specific tasks (such as furniture recognition), YOLOv8 is additionally trained on a small amount of relevant data to improve its performance on that task.

[0086] S400: Based on the information data of wooden furniture in the target environment, control the environmental monitoring mobile robot to move globally and monitor the gas in the target environment to obtain the monitoring results of harmful gases in the target environment.

[0087] Specifically, based on the information data of wooden furniture in the target environment, the environmental monitoring mobile robot is controlled to collect ambient air quality detectors to collect gases in the target environment and obtain the monitoring results of harmful gases at the current sampling point; the distance between the environmental monitoring mobile robot and obstacles is obtained through the ultrasonic distance sensor of the environmental monitoring mobile robot, global movement planning is performed, and the environmental monitoring mobile robot is controlled to move to the next sampling point to collect gases in the target environment; until all sampling points are traversed, the monitoring results of harmful gases in the target environment are obtained.

[0088] In this embodiment, multi-sensor fusion is first performed, combining a fisheye camera and an ultrasonic distance sensor to obtain comprehensive environmental information and improve the robot's ability to perceive the environment. Then, motion control optimization is performed to optimize the motion control algorithm and achieve smooth turning, acceleration and deceleration, thereby improving the robot's stability and safety in complex environments.

[0089] More specifically:

[0090] 1) Route planning: Calculate the best path from the current location to the target location based on the environmental information provided by the YOLO model and ultrasonic sensors, avoiding obstacles such as tables and chairs.

[0091] 2) Motion Decision: Based on the path calculation results, the robot's next action is determined, such as moving straight, turning, accelerating, or decelerating. Simultaneously, drives and shock absorbers are installed in the power system to further reduce shock and vibration. Drives provide more precise motor control and responsiveness, while shock absorbers reduce the transmission of vibrations during robot movement to the surrounding environment, optimizing the smoothness of motion execution.

[0092] 3) Obstacle avoidance strategy: During movement, if the YOLO model or ultrasonic sensor detects a new obstacle, the path planning is updated in real time and an obstacle avoidance strategy is adopted to ensure that the robot passes safely.

[0093] 4) Feedback control: Through the feedback control system and PID controller, the speed and direction of the robot are adjusted to ensure that it travels stably along the predetermined path and responds to changes in the external environment in a timely manner.

[0094] Finally, the method of the present invention is summarized and described in detail as follows:

[0095] 1) Motion and Positioning: The robot uses a fisheye camera to acquire images of its surroundings and an ultrasonic distance sensor to detect obstacles, thereby planning its movement path and achieving autonomous navigation and obstacle avoidance. The robot will move autonomously in an indoor environment, searching for and approaching wooden furniture.

[0096] 2) Wooden furniture recognition: When the robot approaches wooden furniture, it uses a fisheye camera to take pictures of the furniture. These pictures are then processed by the YOLOv8 algorithm to identify the wooden furniture in the images.

[0097] 3) Gas concentration detection: Once wooden furniture is detected, the robot will stop and use an integrated high-sensitivity air quality detector to detect the concentration of harmful gases near the furniture, such as formaldehyde and TVOC.

[0098] 4) Data processing and output: The detected data will be transmitted to the data processing center in real time or sent directly to the user's mobile device via wireless network. Users can view the identification results of wooden furniture and the concentration data of harmful gases through the application.

[0099] 5) Repeat the steps. After completing the inspection of the current furniture, the robot will continue to move to the next target location and repeat the above process until the entire room or set area has been inspected.

[0100] 6) Remote control: Users can also remotely control the robot via mobile devices, such as adjusting the monitoring area or replanning the path.

[0101] Furthermore, experimental simulations were conducted to verify the embodiments of the present invention:

[0102] First, home environment monitoring and furniture identification are performed. In this embodiment, the intelligent robot is deployed in a typical home environment, including the living room, bedroom, and kitchen. The robot is equipped with a fisheye camera and ultrasonic distance sensors to ensure it can obtain comprehensive visual information and perceive surrounding obstacles. Simultaneously, the robot integrates a high-precision air quality detector to monitor the concentration of harmful gases such as formaldehyde, TVOC, and C6H6 in real time, as well as environmental parameters such as temperature and humidity. The robot begins to move autonomously along a predetermined path, using the fisheye camera to capture images while the ultrasonic sensor continuously scans ahead to detect any obstacles. The captured images are processed using the YOLO deep learning model to identify wooden furniture in the images. Meanwhile, the harmful gas concentration data collected by the air quality detector is also analyzed in real time. The processed data is transmitted to a data center via a wireless network in real time. Users can view the identification results of wooden furniture and the concentration data of harmful gases through a mobile device application. Figure 4 As shown, in addition, when an anomaly is detected, the user can manually control the robot through the application, such as adjusting the monitoring area or replanning the path.

[0103] This is followed by office environment monitoring. Intelligent robots are programmed to patrol regularly to monitor air quality in employee work areas. During patrols, the robots use fisheye cameras and ultrasonic distance sensors to avoid collisions with desks, chairs, filing cabinets, and other objects. Simultaneously, a high-precision air quality detector continuously monitors the concentration of harmful gases such as formaldehyde and TVOCs. The robots autonomously move along a predetermined route within the office, collecting environmental data. Deep learning algorithms are used to identify wooden furniture in the office, and the air quality detector monitors the concentration of harmful gases. The monitoring data is uploaded to a server in real time, and office managers can view monitoring reports through computer software to understand the health status of the office environment. Figure 5 As shown, when necessary, managers can adjust the robot's movement path through software or immediately dispatch the robot to a specific area for detailed inspection.

[0104] In summary, the accuracy of wooden furniture identification in this embodiment of the invention is as follows: In a series of tests, the robot's accuracy in identifying wooden furniture reached over 95%. Regarding the monitoring accuracy of harmful gas concentrations, the robot's monitoring error under different environmental conditions was less than ±5%, indicating excellent performance in monitoring accuracy.

[0105] Reference Figure 2 An environmental gas monitoring system based on deep learning, comprising:

[0106] The first module 201 is used to construct an environmental monitoring mobile robot to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data;

[0107] The second module 202 is used to introduce the TAA strategy to train the YOLOv8 object detection model and build an improved YOLOv8 object detection model.

[0108] The third module 203 is used to identify and process the enhanced wooden material obstacle image data based on the improved YOLOv8 target detection model to obtain wooden furniture information data of the target environment.

[0109] The fourth module 204 is used to control the environmental monitoring mobile robot to move globally and monitor the gas in the target environment based on the information data of wooden furniture in the target environment, so as to obtain the monitoring results of harmful gases in the target environment.

[0110] In summary, the robot constructed in this embodiment of the invention can not only move freely in indoor environments and accurately identify the type and location of wooden furniture, but also features advanced environmental monitoring sensors capable of detecting the concentration of harmful gases such as formaldehyde and benzene in indoor air in real time. When abnormally high levels of harmful gases are detected, the robot automatically issues an early warning to the user, reminding them to take appropriate measures. Through continuous monitoring and data analysis, the device can generate detailed air quality reports, helping users understand trends in the indoor environment. This integrated solution not only improves the efficiency and accuracy of detection but also reduces reliance on manual operation, offering significant advantages, especially in situations requiring coverage of large areas and complex environments. Users can not only obtain real-time air quality data but also improve indoor air quality through suggestions provided by the robot. This innovative design provides enhanced safety for families and offices seeking a healthy living environment, creating a safer and healthier living and working space for users.

[0111] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0112] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A deep learning-based method for monitoring environmental gases, characterized in that, Includes the following steps: An environmental monitoring mobile robot was constructed to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data; The YOLOv8 object detection model is pre-trained using a pre-training dataset to obtain the pre-trained YOLOv8 object detection model. Based on the pre-trained YOLOv8 object detection model, a position loss function and a classification loss function are introduced to construct an optimized YOLOv8 object detection model. The TAA strategy is introduced to dynamically adjust the ratio of positive to negative samples in the optimized YOLOv8 object detection model, resulting in an improved YOLOv8 object detection model. The specific expression for the TAA strategy is as follows: In the above formula, Indicates the weighted score. This represents the predicted score for the category corresponding to the predicted bounding box. This represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. , Indicates hyperparameters; Based on the improved YOLOv8 target detection model, the enhanced wooden material obstacle image data is identified and processed to obtain wooden furniture information data of the target environment; Based on the information data of wooden furniture in the target environment, the environmental monitoring mobile robot's ambient air quality detector collects gases in the target environment to obtain the monitoring results of harmful gases at the current sampling point; The distance between the environmental monitoring mobile robot and obstacles is obtained by the ultrasonic distance sensor of the environmental monitoring mobile robot. Global movement planning is carried out to control the environmental monitoring mobile robot to move to the next sampling point to collect target environmental gases. The process continues until all sampling points are visited to obtain the monitoring results of harmful gases in the target environment.

2. The environmental gas monitoring method based on deep learning according to claim 1, characterized in that, The environmental monitoring mobile robot specifically includes a fisheye camera, an ultrasonic distance sensor, and an ambient air quality detector, wherein: The fisheye camera is used to acquire obstacle information data in the target environment; The ultrasonic distance sensor is used to obtain the distance value between the environmental monitoring mobile robot and obstacles; The ambient air quality detector is used to monitor the gas quality in the target environment.

3. The environmental gas monitoring method based on deep learning according to claim 2, characterized in that, The step of constructing an environmental monitoring mobile robot to acquire obstacle image data of the target environment and performing data augmentation processing to obtain enhanced wooden material obstacle image data specifically includes: Construct an environmental monitoring mobile robot and a target scenario, wherein the target scenario includes several sampling points; The fisheye camera of the environmental monitoring mobile robot collects sampling points in the target scene from multiple angles to obtain obstacle image data of the target environment. The obstacle image data of the target environment is classified and processed to obtain obstacle image data of wooden materials and obstacle image data of non-wood materials; Digital augmentation processing is performed on the image data of wooden obstacles to obtain augmented image data of wooden obstacles.

4. The environmental gas monitoring method based on deep learning according to claim 3, characterized in that, The step of performing digital augmentation processing on the wooden obstacle image data to obtain augmented wooden obstacle image data specifically includes: The image data of wooden obstacles is processed by digital cropping to obtain cropped image data of wooden obstacles; The cropped wooden obstacle image data is scaled to obtain scaled wooden obstacle image data; Exposure enhancement processing is performed on the scaled-down image data of wooden obstacles to obtain exposed-enhanced image data of wooden obstacles; The enhanced wooden obstacle image data is obtained by randomly rotating and flipping the image data.

5. The environmental gas monitoring method based on deep learning according to claim 4, characterized in that, The step of identifying and processing the enhanced wooden obstacle image data based on the improved YOLOv8 target detection model to obtain the wooden furniture information data of the target environment specifically includes: The enhanced image data of wooden obstacles is input into the improved YOLOv8 target detection model; Based on the improved YOLOv8 target detection model backbone network module, feature extraction processing is performed on the enhanced wooden obstacle image data to obtain wooden obstacle image feature data; Based on the improved YOLOv8 target detection model, the neck network module performs feature fusion processing on the feature data of wooden obstacle images to obtain fused wooden obstacle image feature data; The head network module based on the improved YOLOv8 target detection model is used to classify and detect the feature data of the fused wooden material obstacle image to obtain the wooden furniture information data of the target environment.

6. An environmental gas monitoring system based on deep learning, characterized in that, Includes the following modules: The first module is used to build an environmental monitoring mobile robot to acquire obstacle image data of the target environment and perform data augmentation processing to obtain enhanced wooden material obstacle image data; The second module is used to pre-train the YOLOv8 object detection model using a pre-training dataset to obtain the pre-trained YOLOv8 object detection model. Based on the pre-trained YOLOv8 object detection model, a position loss function and a classification loss function are introduced to construct an optimized YOLOv8 object detection model. The TAA strategy is introduced to dynamically adjust the ratio of positive to negative samples in the optimized YOLOv8 object detection model, resulting in an improved YOLOv8 object detection model. The specific expression for the TAA strategy is as follows: In the above formula, Indicates the weighted score. This represents the predicted score for the category corresponding to the predicted bounding box. This represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. , Indicates hyperparameters; The third module is used to identify and process the enhanced wooden material obstacle image data based on the improved YOLOv8 target detection model to obtain wooden furniture information data of the target environment. The fourth module is used to control the environmental monitoring mobile robot's ambient air quality detector to collect gases in the target environment based on the information data of wooden furniture in the target environment, and to obtain the monitoring results of harmful gases at the current sampling point. The distance between the environmental monitoring mobile robot and obstacles is obtained by the ultrasonic distance sensor of the environmental monitoring mobile robot. Global movement planning is carried out to control the environmental monitoring mobile robot to move to the next sampling point to collect target environmental gases. The process continues until all sampling points are visited to obtain the monitoring results of harmful gases in the target environment.

Images