Soil respiration dynamic monitoring method based on autonomous mobile robot

Abstract

The invention provides a soil respiration dynamic monitoring method based on an autonomous mobile robot, and relates to the technical field of soil monitoring, and the method comprises the steps: collecting related information of a to-be-monitored area, carrying out grid method environment modeling based on the related information, obtaining a grid model, and carrying out path planning through an improved A * algorithm based on the grid model, the autonomous mobile robot moves according to the planned path, soil respiration data are collected in the moving process, the collected soil respiration data are preprocessed, and soil respiration dynamic monitoring is conducted on the basis of the preprocessed soil respiration data. Through a sensing-planning-analysis-feedback closed-loop system, the core defect of the traditional technology is overcome, an efficient, accurate and extensible solution is provided for soil respiration monitoring, and the method has important practical significance on ecological system management and global carbon cycle research.

Description

Technical Field

[0001] The present invention relates to the technical field of soil monitoring, and in particular to a soil respiration dynamics monitoring method based on an autonomous mobile robot. Background Art

[0002] Soil respiration is a core process in the terrestrial ecosystem carbon cycle. Its dynamic monitoring is crucial for understanding global climate change and assessing ecosystem carbon budgets. Traditional soil respiration monitoring techniques rely primarily on fixed measurement devices (such as static respiration chambers) and manual sampling. However, these methods have significant drawbacks, including:

[0003] 1. Insufficient spatial coverage: Fixed devices can only obtain single-point data and cannot reflect the spatial heterogeneity of soil respiration.

[0004] 2. Low temporal resolution: Manual sampling requires frequent on-site operations, which is time-consuming and inefficient.

[0005] 3. Limitations of data processing models: Traditional methods rely on empirical formulas or simple statistical models and are difficult to process multi-source heterogeneous data.

[0006] Therefore, it is necessary to design a soil respiration dynamic monitoring method based on an autonomous mobile robot. Summary of the Invention

[0007] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a soil respiration dynamics monitoring method based on an autonomous mobile robot.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] The present invention provides a soil respiration dynamic monitoring method based on an autonomous mobile robot, comprising:

[0010] Step 1: Collect relevant information of the area to be monitored, and perform grid-based environmental modeling based on it to obtain a grid model;

[0011] Step 2: Based on the grid model, path planning is performed using the improved A* algorithm;

[0012] Step 3: The autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement;

[0013] Step 4: Preprocess the collected soil respiration data;

[0014] Step 5: Dynamic monitoring of soil respiration based on the preprocessed soil respiration data.

[0015] Preferably, in step 1, relevant information of the area to be monitored is collected, and grid method environmental modeling is performed based on the information to obtain a grid model, specifically:

[0016] Based on lidar, multispectral camera and ground penetrating radar, information of the monitored area is collected to obtain 1cm precision DEM, NDVI and underground obstacles. Based on the 1cm precision DEM, NDVI and underground obstacles, a grid model is constructed using the grid method.

[0017] Preferably, the specific improvement method of the improved A* algorithm is:

[0018] Optimize the search neighborhood selection strategy and child node selection strategy of the traditional A* algorithm;

[0019] Quantify the environmental information of the traditional A* algorithm;

[0020] Improve the heuristic function in the evaluation function of the traditional A* algorithm;

[0021] The path is optimized using a bidirectional smoothing strategy.

[0022] Preferably, the search neighborhood selection strategy and child node selection strategy of the traditional A* algorithm are optimized, specifically:

[0023] The search neighborhood selection strategy of the traditional A* algorithm is optimized as follows:

[0024] Discard 3 search directions based on the eight-neighborhood search;

[0025] The child node selection strategy of the traditional A* algorithm is optimized as follows:

[0026] According to the positional relationship between child nodes and parent nodes, they are divided into different categories, including up-down direction child nodes, left-right direction child nodes and other direction child nodes, and corresponding optimization selection processing is performed based on the child nodes of different categories.

[0027] Preferably, the environmental information of the traditional A* algorithm is quantified, specifically as follows:

[0028] The obstacle ratio is introduced to quantify the environmental information, where the obstacle ratio K is the ratio of the number of grid cells occupied by the current point and the target point to the number of cells in the rectangular environment.

[0029] Preferably, the heuristic function in the evaluation function of the traditional A* algorithm is improved, specifically:

[0030] The obstacle ratio K is introduced into the evaluation function to dynamically adjust the weight of the heuristic function, thereby improving the evaluation function of the traditional A* algorithm.

[0031] Preferably, in step 3, the autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement, specifically:

[0032] Key grids were set at the grid model. When the autonomous mobile robot moved to the key grids, it continuously sampled soil CO2 concentration distribution using an NDIR-CO2 sensor at a frequency of 10 Hz for 3 minutes. The soil state characteristic sensor also collected soil state characteristics, including soil moisture, soil temperature, and soil conductivity.

[0033] When the autonomous mobile robot is in a non-key grid, it obtains the continuous spatial CO2 concentration distribution based on an open laser CO2 analyzer;

[0034] During the movement of the autonomous mobile robot, environmental disturbance characteristics, including wind speed, air pressure and light, are collected through a micro-meteorological station, and thermal infrared temperature fields are collected through a thermal imager.

[0035] Preferably, in step 4, the collected soil respiration data is preprocessed, specifically by:

[0036] Obtain soil respiration data and construct a multidimensional time series data matrix;

[0037] Calculate correlations on multidimensional time series matrices;

[0038] Determine the abnormal dimension based on the calculated correlation coefficient;

[0039] Locate abnormal points on the abnormal dimension;

[0040] Repair abnormal data at abnormal points;

[0041] The soil respiration data after abnormal data repair is obtained and combined with the coordinates of the data collection to generate the data to be monitored.

[0042] Preferably, in step 5, soil respiration dynamic monitoring is performed based on the pre-processed soil respiration data, specifically:

[0043] Obtain the data to be monitored;

[0044] A soil respiration dynamic monitoring model was constructed based on the improved Transformer network structure;

[0045] The soil respiration dynamic monitoring model is trained based on a preset data set to obtain a trained soil respiration dynamic monitoring model;

[0046] The data to be monitored is input into the trained soil respiration dynamic monitoring model to obtain the monitoring results of each coordinate, that is, to determine which coordinates have abnormalities.

[0047] Preferably, the specific improvement method of the improved Transformer network structure is:

[0048] A sparse attention mechanism is added to the traditional Transformer network structure, and a complete encoder-decoder structure is used.

[0049] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] The present invention provides a soil respiration dynamic monitoring method based on an autonomous mobile robot. The method includes collecting relevant information of a monitored area, performing grid-based environmental modeling based on the information to obtain a grid model, performing path planning based on the grid model using an improved A* algorithm, and the autonomous mobile robot moving along the planned path. During the movement, soil respiration data is collected, the collected soil respiration data is preprocessed, and soil respiration dynamic monitoring is performed based on the preprocessed soil respiration data. The present invention has the following beneficial effects:

[0051] 1. This invention uses high-precision dynamic environment modeling and path planning. The grid modeling method reduces storage requirements by 67% compared to traditional 3D modeling. The improved A* algorithm improves path planning efficiency through search neighborhood optimization, subnode classification selection, and dynamic weight adjustment of the heuristic function. The bidirectional smoothing strategy further reduces the number of path turns and reduces the robot's energy consumption.

[0052] 2. The present invention adopts multimodal data collaborative collection and preprocessing, which can collect multimodal data and preprocess it, thereby improving data accuracy.

[0053] 3. The present invention adopts an improved Transformer model, which can realize the monitoring of multimodal data, and finally calculate the monitoring results to determine whether an abnormality occurs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 This is a flow chart of a soil respiration dynamic monitoring method based on an autonomous mobile robot according to an embodiment of the present invention;

[0056] Figure 2 This is a five-neighborhood search diagram;

[0057] Figure 3 Schematic diagram of the selection method of child nodes;

[0058] Figure 4 Schematic diagram of the improved Transformer network structure. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] The purpose of this invention is to provide a soil respiration dynamic monitoring method based on an autonomous mobile robot. Through the "perception-planning-analysis-feedback" closed-loop system, it solves the core defects of traditional technologies and provides an efficient, accurate and scalable solution for soil respiration monitoring, which has important practical significance for ecosystem management and global carbon cycle research.

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] like Figure 1 As shown, the present invention provides a soil respiration dynamic monitoring method based on an autonomous mobile robot, which is applied to an autonomous robot. This application does not limit the autonomous robot and only provides one embodiment, wherein the autonomous robot includes a robot body and a monitoring unit, and the robot body includes:

[0063] Low-disturbance mobile chassis: adopts a six-wheel independent suspension structure, equipped with wide low-pressure tires (diameter ≥ 40cm), ground contact pressure < 5kPa, and achieves a soil surface disturbance rate of < 3%;

[0064] Liftable sensor pod: electric lifting mechanism (travel 30-150cm), integrated sensor stabilization platform, equipped with anti-vibration bracket and electromagnetic shielding layer;

[0065] Multimodal positioning system: multispectral camera, ground-penetrating radar, RTK-GNSS module (positioning accuracy ±1cm), lidar SLAM system (mapping resolution 2cm), UWB seamless indoor and outdoor positioning module;

[0066] The monitoring unit includes NDIR-CO2 sensor, open laser CO2 analyzer, soil three-parameter probe, micro-weather station, and thermal imager;

[0067] Additionally, it includes:

[0068] Energy management unit: lithium battery module (48V / 20Ah) + solar power supplement system (200W);

[0069] Communication module: 5G / Beidou dual-mode transmission, support MQTT protocol;

[0070] Edge computing unit: Embedded industrial computer (Intel NUC, 16GB RAM), running real-time data preprocessing algorithms.

[0071] The method comprises the following steps:

[0072] Step 1: Collect relevant information of the area to be monitored, and perform grid-based environmental modeling based on it to obtain a grid model;

[0073] Step 2: Based on the grid model, path planning is performed using the improved A* algorithm;

[0074] Step 3: The autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement;

[0075] Step 4: Preprocess the collected soil respiration data;

[0076] Step 5: Dynamic monitoring of soil respiration based on the preprocessed soil respiration data.

[0077] In step 1, relevant information of the area to be monitored is collected, and grid-based environmental modeling is performed based on the information to obtain a grid model, specifically:

[0078] Based on the information collection of the monitoring area using lidar, multispectral camera and ground penetrating radar, 1cm precision DEM, NDVI and underground obstacles are obtained. Based on the 1cm precision DEM, NDVI and underground obstacles, a grid model is constructed using the grid method.

[0079] This application adopts the rectangular coordinate method. The lower left corner of the grid is the coordinate origin (0, 0), the horizontal direction is the x-axis, the vertical direction is the y-axis, and the right and upward directions are positive directions. Each unit grid is uniquely marked with coordinates (x, y) to facilitate direct understanding and analysis of the grid area. The specific setting method is a conventional technical means and will not be described in detail here.

[0080] The specific improvement method of the improved A* algorithm is:

[0081] The A* algorithm is an efficient search algorithm for finding the shortest path in static road networks. It combines the concepts of the Dijkstra algorithm and the greedy algorithm and employs a heuristic search strategy. The advantage of the A* algorithm lies in its integration of actual costs and heuristic cost estimates, which not only improves search efficiency but also ensures the optimal path. Guided by a heuristic function, the A* algorithm can selectively expand nodes, accelerating the path to the target. Since the A* algorithm is a conventional technique, it will not be discussed in detail here.

[0082] The A* algorithm is popular for its simple calculation method, shorter planning path, and accelerated search performance after the introduction of heuristic functions. However, the A* algorithm has low search efficiency, uneven paths, and potential safety hazards. Therefore, in response to the problems of the A* algorithm, the traditional A* algorithm is improved. First, the search neighborhood selection strategy is optimized to improve search efficiency; second, the environmental information is quantified to better represent the complexity of the environment; third, the heuristic function in the evaluation function is improved, the obstacle ratio K is introduced, and the weight of the heuristic function is dynamically adjusted; then, the child node selection strategy is optimized to avoid collisions; finally, the path is optimized through a two-way smoothing strategy to enhance the smoothness of the path. They are introduced separately:

[0083] The search neighborhood selection strategy of the traditional A* algorithm is optimized as follows:

[0084] In a grid map, the computational complexity of the four-neighborhood search is small, so the search time is shorter. The eight-neighborhood search allows movement along the grid diagonal, so the path length is shorter. Based on the eight-neighborhood search, three search directions are discarded and a five-neighborhood search is adopted. The schematic diagram is as follows: Figure 2 As shown, the angle between the line connecting the target point and the current point and the N2 direction is set to According to the corresponding relationship in Table 1, the search direction can be flexibly selected to improve the search efficiency and path planning quality;

[0085] Table 1 Angle The correspondence between the three discarded directions

[0086]

[0087] The environmental information of the traditional A* algorithm is quantified as follows:

[0088] In order to estimate the length of the minimum path more accurately and take into full account the distribution of obstacles, the obstacle ratio K is introduced. Assuming that the obstacle ratio K is in a rectangular scene composed of the current point and the target point, the value is the ratio of the number of grid cells occupied to the number of cells in the rectangular environment. Assuming that the number of grid cells occupied in the rectangular environment composed of the current position and the target point position is N, the starting point is (x s ,y s ), the target point is (x t ,y t ), then the expression of obstacle ratio K is as follows:

[0089]

[0090] The heuristic function in the evaluation function of the traditional A* algorithm is improved as follows:

[0091] The obstacle ratio K is affected by the number and distance of obstacle grids. By introducing the obstacle ratio K into the evaluation function f(n), the weight of h(n) can be dynamically adjusted to achieve adaptive adjustment of the evaluation function f(n). h(n) and the improved f(n) are shown in the following formula:

[0092]

[0093] f(n)=g(n)+(1-lnK)h(n);

[0094] The child node selection strategy of the traditional A* algorithm is optimized as follows:

[0095] The present invention optimizes the selection method of child nodes by considering the relative position of child nodes and obstacles to prevent the path from obliquely crossing the obstacle vertex. According to the positional relationship between child nodes and parent nodes, they are divided into different categories, including Class I child nodes (up and down direction), Class II child nodes (left and right direction) and Class III child nodes (other directions). Figure 3 As shown;

[0096] According to the three types of sub-nodes, Ⅰ, Ⅱ, and Ⅲ, the corresponding optimization selection strategies are formulated, as shown in Table 2;

[0097] Table 2 Child node selection rules

[0098]

[0099] The path is optimized through a two-way smoothing strategy, specifically:

[0100] The present invention provides a bidirectional smoothing strategy that gradually optimizes the path from the starting point and the end point at the same time. By judging whether the straight-line distance between non-adjacent nodes is less than the specified distance and ensuring that this straight line does not intersect with obstacles, the algorithm can intelligently delete redundant paths between these nodes.

[0101] In step 3, the autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement, specifically:

[0102] Key grids are set at the grid model. When the autonomous mobile robot moves to the key grid, it deploys a liftable sensor pod and presses the breathing chamber into the soil surface to form a confined space. Continuous sampling is performed using an NDIR-CO2 sensor at a frequency of 10 Hz for 3 minutes to obtain the soil CO2 concentration distribution. Soil state characteristics, including soil moisture, soil temperature, and soil conductivity, are collected using a soil state characteristic sensor.

[0103] When the autonomous mobile robot is in a non-key grid, it retracts the breathing chamber and activates the open laser CO2 analyzer to obtain the continuous spatial CO2 concentration distribution;

[0104] During the movement of the autonomous mobile robot, environmental disturbance characteristics, including wind speed, air pressure and light, are collected through a micro-meteorological station, and thermal infrared temperature fields are collected through a thermal imager.

[0105] In step 4, the collected soil respiration data is preprocessed, specifically:

[0106] Obtain soil respiration data and construct a multidimensional time series data matrix;

[0107] The correlation calculation is performed on the multidimensional time series matrix, specifically:

[0108] Using sliding windows to segment the stored complete multidimensional time series data, we attempt to transform the global anomaly detection problem into a local anomaly detection problem. Each segmented local multidimensional time series data is called a double sliding window, and correlation calculations are performed on the data represented as arrays in each double sliding window in chronological order. During the correlation calculation process, since the correlation coefficient ranges from [-1 to 1], where a negative value indicates a negative correlation between the two dimensions of data and a positive value indicates a positive correlation between the two dimensions of data, whether the obtained correlation coefficient is positive or negative will not affect the anomaly detection effect.

[0109] The abnormal dimension is determined based on the calculated correlation coefficient, specifically:

[0110] In order to identify suspicious dimensions with anomalies, the complete sequence is divided into n / w multidimensional time series data segments through a double sliding window, where n represents the number of data points in the entire sequence and w represents the size of the double window. First, the correlation coefficient of the multidimensional time series data segment in the first double window is calculated to obtain the correlation coefficient adjacency matrix Corr_Matrix_Window_1. In order to find the correlation coefficient that marks anomalies in the double window, assuming that the historical data in the same window size is correct, the historical data can be obtained through correlation calculation to obtain the correlation coefficient adjacency matrix Corr_Matrix_Origin.

[0111] Secondly, the correlation coefficient adjacency matrix Corr_Matrix_Window_1 calculated by the first double sliding window is compared with the correlation coefficient adjacency matrix Corr_Matrix_Origin calculated by the data in the historical double sliding window. Since an undirected graph is used to express the logical structure of multidimensional time series data, and for the edge E of the undirected graph, its two end points have the same effect when performing forward or reverse correlation calculations, when storing the correlation coefficient adjacency matrix, only the upper triangle or lower triangle matrix needs to be saved. At the same time, in the process of determining the abnormal dimension, only the upper triangle or lower triangle of the adjacency matrix is ​​considered. For the upper triangular matrix, the relationship between the element subscript and the corresponding number is Where i and j are the row and column coordinates of the elements respectively, and k is the number. Once the difference between the k correlation coefficients in the i-th row of the upper triangular adjacency matrix in the first double sliding window and the corresponding values ​​in the correlation coefficient adjacency matrix in the historical double sliding window is greater than the set threshold, the i-th dimension data is considered to be an abnormal dimension. It should be noted that this threshold is an empirical value obtained through experiments in different scenarios. After determining the abnormal dimension of the first double sliding window, the first double sliding window is cleaned by accurately locating and repairing the abnormal points. The anomaly detection of the second double sliding window depends on the correlation between the multidimensional time series data in the cleaned first double sliding window. And so on. After completing the anomaly detection of each double sliding window, it means that the cleaning of the entire sequence is completed.

[0112] Locate abnormal points on the abnormal dimension, specifically:

[0113] After determining the abnormal dimension, we need to focus on how to locate the abnormal point in the abnormal dimension. Since the concept of dynamic speed change rate constraint is introduced when performing anomaly detection on single-dimensional time series data, that is, historical data is learned and simulated through ELM, and finally a reasonable threshold is given. Once the speed change rate calculated for two data points in the previous and next time in the single-dimensional time series data exceeds or falls below this threshold, it is considered an abnormal point. Based on this idea, the complex relationship between multidimensional time series data is considered. It is believed that there is a certain correlation between dimensions. Once this correlation shows a strong correlation, the change trends of the two dimensions with strong correlation should be relatively consistent. Therefore, at this time, the similarity of the speed change rate ratio between the data in the abnormal dimension and the dimension most correlated with it can be used to describe this consistency. When the speed change rate ratio between the data in the abnormal dimension is significantly lower or higher than the speed change rate ratio of the data in the other dimension (exceeding or falling below the threshold), the abnormal point can be located. Among them, the threshold for judging whether the speed change rate fluctuates significantly needs to be set by yourself. According to experience, its value is set to 0.5.

[0114] Perform abnormal data repair on abnormal points, specifically including generating a candidate set of repair values ​​and determining the final repair value;

[0115] Generating a candidate set of repair values ​​is the first step in anomaly repair. It first receives anomalies discovered in the anomaly detection phase in the order of the dual sliding windows, and stores and represents the anomalies and their related dimensions in each dual sliding window. Secondly, it defines the correlation between the multidimensional time series data within the current dual sliding window, and combines the total impact of other dimensions on the dimension to be repaired to calculate the impact value of each dimension. Finally, it randomly extracts several dimensions to assist in generating a candidate set of repair values ​​for the dimension to be repaired.

[0116] Determining the final repair value is the second and final step in anomaly repair. First, a clustering-based method is used to optimally classify the candidate set of repair values ​​generated during anomaly repair, thereby obtaining several clusters. Second, the cluster with the highest data density is selected according to the "minority obeys the majority" rule (i.e., points with low data density are outliers or anomalies, while points with high density form a class or cluster). For such a cluster, its cluster center is considered to be the most representative of all data, so the cluster center is often used as the final repair value. However, it may happen that the cluster center does not belong to the candidate set of repair values. Therefore, the data that exists in the candidate set of repair values ​​and is closest to the cluster center is generally selected as the final repair value.

[0117] The soil respiration data after abnormal data repair is obtained and combined with the coordinates of the data collection to generate the data to be monitored.

[0118] In step 5, soil respiration dynamic monitoring is performed based on the preprocessed soil respiration data, specifically:

[0119] Obtain the data to be monitored;

[0120] A soil respiration dynamic monitoring model was constructed based on the improved Transformer network structure;

[0121] The soil respiration dynamic monitoring model is trained based on a preset data set to obtain a trained soil respiration dynamic monitoring model;

[0122] The data to be monitored is input into the trained soil respiration dynamic monitoring model to obtain the monitoring results of each coordinate, that is, to determine which coordinates have abnormalities.

[0123] The specific improvement method of the improved Transformer network structure is:

[0124] Since the Transformer network structure is a conventional technical means, it will not be described in detail here. Instead, the improvement methods are mainly described. The improvement ideas are mainly based on the following aspects:

[0125] Add a sparse attention mechanism to the traditional Transformer network structure and use a complete encoder-decoder structure;

[0126] Among them, the improved Transformer network structure diagram is as follows Figure 4 As shown;

[0127] Describe the sparse attention mechanism:

[0128] Sparse Attention is an optimization of the traditional attention mechanism, aiming to reduce computational complexity and memory consumption;

[0129] The sparse attention mechanism reduces computational complexity by limiting the calculation of attention scores to only a small number of elements in the sequence. A common sparse attention formula can be expressed as:

[0130]

[0131] Among them, Q, K, V and are query, key and value matrices respectively, d k is the dimension of the key vector, Used to scale the dot product to prevent excessive values. M is a mask matrix used to achieve sparsity. Its elements are either 0 (indicating that attention between positions is allowed) or negative infinity (indicating that attention between positions is blocked);

[0132] The decoder and encoder of the present invention are introduced in detail:

[0133] The encoder part is the key to the model and is responsible for converting the input sequence for subsequent tasks. It consists of a series of encoder layers, each of which implements a multi-head self-attention mechanism, followed by a feedforward network consisting of two linear layers. This design makes the output of each position depend not only on its own characteristics, but also on the information of other positions in the sequence. The sparse attention technology used in the encoder selectively calculates the attention weights instead of calculating them for all possible input pairs, thereby reducing the amount of computation while retaining the ability to model long-distance dependencies and reducing the number of attention weights that must be calculated, thereby reducing computational complexity, especially for long sequence data.

[0134] Similar to the encoder, the decoder consists of a stack of decoder layers. Starting from some initial value, the output is built layer by layer through the decoder. Each decoder layer applies self-attention to the output and then applies attention between the output and the sequence output after encoder processing. Like the encoder layer, it is followed by two layers of feed-forward network.

[0135] The input to the decoder is generated through a feed-forward network consisting of four linear layers. These four linear layers are designed to transform the encoder output sequence into a form suitable for the decoder to process. Specifically, the functions of these four linear layers are as follows:

[0136] The first linear layer reduces the encoder output from the embedding dimension to one dimension. This step is to reduce the dimensionality of the data to make it more suitable for subsequent processing. The second linear layer stretches the sequence to a larger intermediate size (in the time dimension). This step is to expand the sequence in preparation for the subsequent nonlinear transformation and compression;

[0137] Introducing nonlinear transformations between the second and third layers, usually using the GELU activation function, helps increase the expressive power of the model;

[0138] The third linear layer compresses the sequence to the length of the prediction horizon. This step compresses the expanded sequence back to a more compact form to match the length of the prediction horizon.

[0139] The fourth linear layer expands the dimension of the sequence back to the embedding dimension, corresponding to the input linear layer. This step is to restore the dimension of the sequence so that it can be further processed by the decoder;

[0140] The introduction of these four linear layers is to perform necessary transformations and adjustments on the encoder output before the decoder processes it.

[0141] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0142] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A soil respiration dynamic monitoring method based on an autonomous mobile robot, characterized in that: include: Step 1: Collect relevant information of the area to be monitored, and perform grid-based environmental modeling based on it to obtain a grid model; Step 2: Based on the grid model, path planning is performed using the improved A* algorithm; Step 3: The autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement; Step 4: Preprocess the collected soil respiration data; Step 5: Dynamic monitoring of soil respiration based on the preprocessed soil respiration data.

2. The method according to claim 1, characterized in that In step 1, relevant information of the area to be monitored is collected, and grid-based environmental modeling is performed based on the information to obtain a grid model, specifically: Based on lidar, multispectral camera and ground penetrating radar, information of the monitored area is collected to obtain 1cm precision DEM, NDVI and underground obstacles. Based on the 1cm precision DEM, NDVI and underground obstacles, a grid model is constructed using the grid method.

3. The method according to claim 2, characterized in that The specific improvement method of the improved A* algorithm is: Optimize the search neighborhood selection strategy and child node selection strategy of the traditional A* algorithm; Quantify the environmental information of the traditional A* algorithm; Improve the heuristic function in the evaluation function of the traditional A* algorithm; The path is optimized using a bidirectional smoothing strategy.

4. The method according to claim 3, characterized in that The search neighborhood selection strategy and child node selection strategy of the traditional A* algorithm are optimized as follows: The search neighborhood selection strategy of the traditional A* algorithm is optimized as follows: Discard 3 search directions based on the eight-neighborhood search; The child node selection strategy of the traditional A* algorithm is optimized as follows: According to the positional relationship between child nodes and parent nodes, they are divided into different categories, including up-down direction child nodes, left-right direction child nodes and other direction child nodes, and corresponding optimization selection processing is performed based on the child nodes of different categories.

5. The method according to claim 3, characterized in that The environmental information of the traditional A* algorithm is quantified as follows: The obstacle ratio is introduced to quantify the environmental information, where the obstacle ratio K is the ratio of the number of grid cells occupied by the current point and the target point to the number of cells in the rectangular environment.

6. The method according to claim 3, characterized in that The heuristic function in the evaluation function of the traditional A* algorithm is improved as follows: The obstacle ratio K is introduced into the evaluation function to dynamically adjust the weight of the heuristic function, thereby improving the evaluation function of the traditional A* algorithm.

7. The method according to claim 3, characterized in that In step 3, the autonomous mobile robot moves according to the planned path and collects soil respiration data during the movement, specifically: Key grids were set at the grid model. When the autonomous mobile robot moved to the key grids, it continuously sampled soil CO2 concentration distribution using an NDIR-CO2 sensor at a frequency of 10 Hz for 3 minutes. The soil state characteristic sensor also collected soil state characteristics, including soil moisture, soil temperature, and soil conductivity. When the autonomous mobile robot is in a non-key grid, it obtains the continuous spatial CO2 concentration distribution based on an open laser CO2 analyzer; During the movement of the autonomous mobile robot, environmental disturbance characteristics, including wind speed, air pressure and light, are collected through a micro-meteorological station, and thermal infrared temperature fields are collected through a thermal imager.

8. The method according to claim 7, characterized in that In step 4, the collected soil respiration data is preprocessed, specifically: Obtain soil respiration data and construct a multidimensional time series data matrix; Calculate correlations on multidimensional time series matrices; Determine the abnormal dimension based on the calculated correlation coefficient; Locate abnormal points on the abnormal dimension; Repair abnormal data at abnormal points; The soil respiration data after abnormal data repair is obtained and combined with the coordinates of the data collection to generate the data to be monitored.

9. The method according to claim 8, characterized in that In step 5, soil respiration dynamics monitoring is performed based on the preprocessed soil respiration data, specifically: Obtain the data to be monitored; A soil respiration dynamic monitoring model was constructed based on the improved Transformer network structure; The soil respiration dynamic monitoring model is trained based on a preset data set to obtain a trained soil respiration dynamic monitoring model; The data to be monitored is input into the trained soil respiration dynamic monitoring model to obtain the monitoring results of each coordinate, that is, to determine which coordinates have abnormalities.

10. The method according to claim 9, characterized in that The specific improvement method of the improved Transformer network structure is: A sparse attention mechanism is added to the traditional Transformer network structure, and a complete encoder-decoder structure is used.

Images