A dynamic reconstruction method for farmland gas sensor networks

Abstract

This invention discloses a method for dynamic reconfiguration of a farmland gas sensor network, comprising: acquiring raw sensor array data, preprocessing it, and generating a standardized sensor dataset and a network topology state matrix; calculating a dynamic weight matrix based on the standardized sensor dataset by quantifying the dynamic changes in information entropy, and evaluating information complementarity between sensors in combination with the network topology state matrix to obtain redundant information identification results; applying multi-timescale convergence control to the dynamic weight matrix to generate a convergent and stable weight configuration; and fusing the convergent and stable weight configuration with the redundant information identification results to generate a final network reconfiguration decision. This invention deeply couples weight allocation with the dynamic changes in information value and introduces a multi-scale convergence control mechanism, improving the monitoring accuracy, stability, and intelligence level of the sensor network.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural Internet of Things, and in particular to a method for dynamically reconfiguring a farmland gas sensor network. Background Art

[0002] Precision agriculture is a core development direction for modern agriculture. Its goal is to finely manage agricultural production processes through information technology and intelligent equipment, thereby achieving cost reduction, increased efficiency, energy conservation, and environmental protection. In this context, real-time, accurate monitoring of key gases in farmland environments is of paramount importance. When crops are exposed to pest and disease stress or nutritional imbalances, their metabolic processes change, releasing specific volatile or semi-volatile gases such as ethylene, ammonia, and sulfides. Deploying wireless sensor networks in the field to sensitively capture the concentration, distribution, and dynamic changes of these signal gases provides critical data support for extremely early warning of pests and diseases, accurate diagnosis of crop growth, and on-demand decision-making for irrigation and fertilization. Therefore, building an efficient, stable, and adaptive farmland gas sensor network is a core technological cornerstone for advancing precision agriculture from concept to large-scale application, ensuring the coordinated development of food security and the ecological environment.

[0003] Extensive research has been conducted on the application of wireless sensor networks in agricultural environmental monitoring. Regarding network deployment, existing solutions often employ static, pre-planned physical layout strategies, with regular grid layouts being the most common. In terms of data processing and fusion, to comprehensively utilize multi-sensor information, researchers typically employ weighted fusion algorithms to reconstruct gas concentration fields. Weight coefficients are often assigned based on static physical or geometric properties, such as the spatial distance between sensor nodes (e.g., the inverse distance method) or the stability of signal transmission. To address information redundancy in networks, existing research has also incorporated tools from information theory. For example, by calculating mutual information between sensor nodes, this method can identify and assess high correlations in data, thus providing a basis for network optimization. Regarding network maintenance and adjustment, relatively simple single-scale temporal models are often employed. Short-term data statistical characteristics (e.g., packet loss rate, node energy consumption) or manual inspections are used to trigger node sleep, wakeup, or replacement to maintain basic network operations.

[0004] However, these existing technologies still face challenges when dealing with the highly dynamic, heterogeneous, and complex environment of farmland. For example, existing farmland sensor network approaches primarily rely on static weights based on physical location. This prevents them from identifying and focusing on truly valuable early warning information that dynamically emerges in the data, resulting in blind spots in information value. Simple dynamic adjustment schemes introduced to address this issue often overreact to transient disturbances because they cannot distinguish between real events and environmental noise. This leads to a stability dilemma in which the network struggles to balance responsiveness and stability, resulting in frequent oscillations and low reliability in monitoring results. Summary of the Invention

[0005] The purpose of the invention is to provide a method for dynamic reconstruction of a farmland gas sensor network to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: A method for dynamically reconfiguring a farmland gas sensor network, including:

[0007] Obtain raw sensor array data, preprocess it, and generate standardized sensor data sets and network topology state matrices;

[0008] Based on the standardized sensor data set, the dynamic weight matrix is ​​calculated by quantifying the dynamic change of information entropy, and the information complementarity between sensors is evaluated in combination with the network topology state matrix to obtain the redundant information identification result.

[0009] Apply multi-time-scale convergence control to the dynamic weight matrix to generate convergent and stable weight configurations;

[0010] The convergent and stable weight configuration and redundant information identification results are integrated to generate the final network reconstruction decision.

[0011] Beneficial effect: The present invention deeply couples weight distribution with the dynamic changes of information value and introduces a multi-scale convergence control mechanism, which improves the detection success rate and positioning accuracy of early and weak disease sources, while improving the monitoring accuracy, stability and intelligence level of the sensor network. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart of the steps of a method for dynamically reconfiguring a farmland gas sensor network provided in an embodiment of the present application.

[0013] Figure 2 A flowchart of the steps for calculating the dynamic weight matrix provided in an embodiment of the present application.

[0014] Figure 3 A flowchart of the steps for performing nonlinear mapping on the information entropy change rate matrix provided in an embodiment of the present application.

[0015] Figure 4A flowchart of the steps for generating a convergent and stable weight configuration provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] In order to solve the problems of low monitoring accuracy and poor adaptability of existing static networks in dynamic environments, such as Figure 1 As shown in the figure, a dynamic reconstruction method of farmland gas sensor network is proposed, which includes the following steps:

[0019] Obtain raw sensor array data, preprocess it, and generate standardized sensor data sets and network topology state matrix.

[0020] In this embodiment, the original sensor array data covering the operation area is read by the multi-source sensor data synchronous acquisition module {R i (t), i=1, 2, ..., N} and related environmental parameter data {E i(t)}, and through operations such as timestamp alignment, a synchronized raw data matrix is ​​formed. The raw sensor array data includes ambient gas concentration data, sensor physical state information, time series acquisition information, soil moisture and temperature data, wind speed, and light intensity. Environmental parameter data includes meteorological data, light and radiation indicators, and soil surface and vegetation information. Based on this, the synchronized data is quality assessed and anomalies are marked. For example, an improved 3σ criterion combined with a sliding window can be used to detect anomalies. The data credibility of each sensor is calculated to obtain a data quality assessment matrix Q(t). The marked anomaly data is adaptively cleaned and compensated, for example, using an adaptive interpolation algorithm based on neighborhood correlation to correct it, obtaining a cleaned data matrix D(t). Combining the preset sensor location information P with the cleaned data, network connectivity is constructed through spatial distance and signal correlation calculations, ultimately generating a standardized sensor dataset S(t) and a network topology state matrix T(t). The improved 3σ criterion for data quality assessment improves on the sliding window-based local standard deviation rather than the global data standard deviation. Specifically, when determining whether a data point is an outlier, the system calculates the mean and standard deviation within a fixed-size sliding window (preferably, the sliding window size can be set to 10 to 20 consecutive data sampling points) within which the data point resides. The use of local statistics allows anomaly detection to better adapt to the local fluctuations of the data, avoiding the technical problem of normal data being misclassified as anomalies due to sudden changes in the global environment. This embodiment provides a high-quality, uniformly formatted data foundation for subsequent information value assessment and network reconstruction. It addresses technical issues such as the diverse sources of raw sensor data, the presence of noise and outliers, and time asynchrony, and is a prerequisite for ensuring the accuracy of the entire method.

[0021] Based on the standardized sensor dataset, the dynamic weight matrix is ​​calculated by quantifying the dynamic change of information entropy, and the information complementarity between sensors is evaluated in combination with the network topology state matrix to obtain the redundant information identification result.

[0022] In this embodiment, based on a standardized sensor dataset S(t), the information contribution of each sensor is calculated using the information entropy-time derivative weight model to obtain a dynamic weight matrix W(t). Essentially, this matrix represents the value of information using its rate of change. Combining the dynamic weight matrix W(t) with the network topology state matrix T(t), a complementarity evaluation algorithm is used to construct information relationships between sensors, resulting in an information complementarity matrix M(t). Based on this matrix, redundant sensor pairs with overlapping information are identified, forming the redundant information identification result R(t). This step analyzes the network from two dimensions: single-point information value and multi-point information relationships. The goal is to identify sensors with high information value but overlapping information, providing a basis for subsequent optimization decisions.

[0023] Multi-timescale convergence control is applied to the dynamic weight matrix to generate convergent and stable weight configurations.

[0024] In this embodiment, since environmental changes may cause the dynamic weight matrix W(t) to fluctuate violently and affect the stability of the system, a multi-time-scale progressive adjustment mechanism is introduced. Specifically, the system will construct at least two time windows of different scales, short-term and long-term, to track the historical evolution of the dynamic weight matrix W(t) and form a multi-scale weight history matrix WH(t). The convergence of the weights is evaluated by analyzing the data characteristics in the history matrix, and the unstable weights are adjusted with damping according to the evaluation results, and finally a converged and stable weight configuration W is output. stable (t). In order to solve the technical problem that dynamic weights may fluctuate violently due to sudden changes in the environment, this embodiment introduces time dimension control to ensure the system's sensitivity to environmental changes while improving the long-term stability and reliability of the weight configuration.

[0025] The convergent and stable weight configuration and redundant information identification results are integrated to generate the final network reconstruction decision.

[0026] In this embodiment, the stable weight configuration W stable (t) and the redundant information identification result R(t) are used as inputs, and the final network optimization solution is generated by reconstructing the decision algorithm. The decision can be at the logical level, such as adjusting the logical weight when fusion data is used; it can also be at the physical level. For example, when the decision involves physical position adjustment, the optimal new position of the sensor is determined by solving a multi-objective optimization problem and generating a physical position adjustment suggestion P. new This embodiment converts the analysis results of the previous steps into specific, executable network optimization actions, closing the complete technical loop of state perception-value assessment-relationship analysis-stable control-optimization decision-making. It solves the problem of large positioning errors and low detection success rates of traditional regular grid layout sensor arrays caused by plant obstruction, terrain undulations, and airflow disturbances in farmland.

[0027] Research has revealed that existing technologies commonly suffer from static blind spots in information value. Traditional weighting mechanisms based on fixed spatial topology essentially equate a sensor's physical location with its information value. This assumption is fragile in farmland environments, where gases diffuse freely and are heavily influenced by airflow and plant shading. For example, a sensor located near a newly infected site may have low absolute values ​​for its monitoring data, but its changing trend (i.e., a large time derivative of information entropy) suggests high early warning value. Meanwhile, a sensor located further away, even if its signal is stable, may maintain a steady-state background for extended periods, resulting in lower information value. Static weighting methods fail to capture this event-driven dynamic fluctuation in information value, resulting in valuable early warning signals at key nodes being buried amidst a flood of low-value data. This results in a significant waste of network sensing resources and is a key contributor to the large positioning errors and low detection success rates of existing solutions.

[0028] like Figure 2 As shown, according to one aspect of the present application, calculating a dynamic weight matrix includes:

[0029] Calculate the Shannon information entropy of the time series of each sensor in the standardized sensor dataset;

[0030] Perform time series difference processing on the Shannon information entropy of the time series to obtain the time derivative of the Shannon information entropy of the time series and form the information entropy change rate matrix; the time derivative is used to characterize the dynamic change of the sensor information value;

[0031] Nonlinear mapping is performed on the information entropy change rate matrix to convert the information entropy change rate of each sensor into a dynamic weight, thereby obtaining a dynamic weight matrix.

[0032] In one embodiment of the present application, for each sensor i in the standardized sensor data set S(t), the Shannon information entropy H of its time series is calculated using a sliding time window. i (t). Shannon information entropy is a classic indicator for measuring uncertainty in information theory. The larger its value, the higher the uncertainty of the data sequence observed by the sensor and the richer the information. i (t) Perform first-order time difference to approximate its time derivative dH i / dt = [H i (t)-H i (t-Δt)] / Δt, where t is time and Δt is the calculation step size. dH i / dt is the rate of change of information entropy, which represents the degree to which the amount of information carried by the sensor changes over time. i / dt means that the sensor is capturing new, unexpected events, and its information value is highlighted at this moment. i / dt together form the information entropy change rate matrix dH / dt. Through the nonlinear mapping function f(·), the information entropy change rate dH of each sensor is converted to i / dt is converted into its dynamic weight W i (t), that is, W i (t) = f(dH i / dt).

[0033] In an optional embodiment, the characterization of the rate of change of information entropy may not be limited to the first-order derivative. i / dt (the speed of change of information entropy), and its second-order time derivative d can also be calculated 2 H i / dt 2 , this value can be understood as the acceleration of information entropy change, which can reflect the acceleration or deceleration trend of information volume change. The first-order derivative and the second-order derivative can be combined into a eigenvector [dH i / dt,d 2 H i / dt 2 ] and uses this vector as the input of the nonlinear mapping function f(·), which enables the system to understand the information dynamics more comprehensively and deeply and make more accurate weight judgments.

[0034] like Figure 3 As shown, according to one aspect of the present application, the step of performing nonlinear mapping on the information entropy change rate matrix includes:

[0035] Based on the original sensor array data, the data credibility of each sensor is calculated to obtain a data quality assessment matrix; the data quality assessment matrix is ​​combined to calculate the sensor characteristic coefficient vector used to characterize the individual response characteristics of each sensor;

[0036] Based on the sensor characteristic coefficient vector and the information entropy change rate matrix, for each sensor, in the nonlinear mapping process of each sensor, the dynamic weight is calculated jointly by fusing the information entropy change rate of the sensor, the corresponding evaluation result, and the corresponding characteristic coefficient.

[0037] In one embodiment of the present application, the nonlinear mapping process f(·) does not depend only on dH i / dt, but rather integrates information from more dimensions to improve the accuracy and robustness of the weights. Specifically, W i (t) = f(dH i / dt,Q i (t), β i ), where Q i (t) is the quality score of the corresponding sensor i at time t in the data quality assessment matrix Q(t), which is used to downgrade the low-quality data source in real time; βi It is the characteristic coefficient calibrated for each sensor i, which is used to realize individualized weight calculation.

[0038] This embodiment improves the sensitivity and accuracy of network perception by changing the core basis of weight calculation from static physical position to dynamic information entropy time derivative dH / dt. According to information theory, the time derivative of information entropy dH / dt is proportional to the intensity of the change in environmental state (i.e., Fisher information amount). Therefore, by calculating and applying dH in real time, i / dt to determine the sensor weight W i (t), can dynamically and automatically shift the focus of data fusion to those areas that are experiencing significant gas concentration changes (i.e., dH i In the farmland gas monitoring scenario, the release of characteristic gases caused by a new pest and disease site or abnormal crop area will inevitably cause the dH value of nearby sensors to increase. i / dt increases instantaneously. This example can immediately capture this signal and increase the weight of the sensor, giving it a greater voice. This solves the information value blind spot problem caused by the use of static weights, where key warning signals are averaged and drowned in massive background data. It also improves the detection success rate and positioning accuracy of early-stage, subtle disease sources.

[0039] According to one aspect of the present application, each sensor characteristic coefficient in the sensor characteristic coefficient vector is calculated by fusing at least two of the following indicators:

[0040] The first indicator is used to measure the activity of the sensor's historical information, which is derived from the variance of the sensor's historical information entropy sequence;

[0041] The second indicator, used to evaluate the long-term reliability of the sensor, is derived from the average quality level of the sensor recorded in the data quality assessment matrix.

[0042] In one embodiment of the present application, for example, the characteristic coefficient β i It can be calculated according to the following formula: i =α base ×(1+γ×Var(H i_history ))×Q avg_i ; where α base is the basic coefficient, which can usually be set to 1.0 as the calculation benchmark; γ is the variance sensitivity factor, which reflects the importance the system attaches to the activity of sensor historical information. Its recommended value range is [0.1, 0.5]. The optimal value of γ can be determined by offline grid search or optimization testing on a set of benchmark data sets containing various typical farmland environmental change modes (such as normal, disease occurrence, and weather mutation); Var(H i_history) is the variance of the sensor's historical information entropy sequence, which is used to characterize its historical information activity. Sensors with frequent historical fluctuations are considered to have greater potential; Q avg_i is the long-term average quality score of the sensor, representing its long-term reliability. i ,A unique personality profile is established for each sensor, so that the weight allocation can take into account the historical performance and inherent quality of the sensor, avoiding the drawbacks of homogenization.

[0043] Furthermore, the nonlinear mapping process is specifically as follows: judging the relationship between the information entropy change rate θ and the first threshold θ1 and the second threshold θ2; when θ≤θ1, it is determined to be in a stable change range, and a quadratic function is used; when θ1<θ≤θ2, it is determined to be in a transition change range, and a logarithmic function is used; when θ>θ2, it is determined to be in a saturated change range, and exponential decay is used; wherein θ1<θ2, and θ1 and θ2 are dynamically set according to the real-time statistical distribution of the values ​​in the information entropy change rate matrix.

[0044] In one embodiment of the present application, the nonlinear mapping function f(·) is designed as a piecewise function to adapt to different information change patterns. Let the input information entropy change rate be θ (ie, dH i / dt): When θ≤θ1, the dynamic weight W i (t) is calculated by a quadratic function with an upward opening. The information in this interval changes steadily, and the weight increases slowly as the rate of change increases. When θ1<θ≤θ2, W i (t) is calculated by the logarithmic function. This interval corresponds to the information transition change. The weight growth rate slows down as the change rate increases, reflecting the diminishing marginal utility of information. When θ>θ2, W i (t) is calculated using an exponential decay function. This interval corresponds to information saturation or abnormal mutations. To prevent excessive weighting of individual sensors from affecting the stability of the entire network, the growth of their weights is suppressed or even attenuated. The thresholds θ1 and θ2 are not fixed but are dynamically set based on the real-time statistical distribution of the information entropy change rate dH / dt of all sensors in the entire network (for example, based on the 25th and 75th percentiles). This allows the weight model to automatically adapt to the overall activity level of the farmland environment. For example, the model can find an appropriate mapping interval in both the morning when airflow is stable and the afternoon when airflow is highly turbulent, demonstrating strong environmental generalization capabilities.

[0045] In an optional embodiment, the adaptive update mechanism for the dynamic thresholds θ1 and θ2 can be specifically configured as follows: the system periodically (e.g., every 30 minutes) performs a complete statistical analysis of the dH / dt matrix constructed by all sensors in the current network and updates the values ​​of θ2 and θ1 based on the latest 75th and 25th percentiles. Furthermore, to prevent jerky weight transitions near the thresholds θ1 and θ2, a narrow transition interval can be introduced at the boundary between the two piecewise functions. Within this interval, a smoothing function such as a sigmoid function or cubic spline interpolation can be used to ensure the continuity and derivative continuity of the entire nonlinear mapping function, thereby enhancing the smoothness of weight changes.

[0046] This embodiment achieves both robustness and a high dynamic range for weight calculation by designing a piecewise nonlinear mapping function with adaptive thresholds. Specifically, for small dH / dt changes (stationary period), a quadratic function achieves a smooth response and effectively filters background noise. For moderate dH / dt (transition period), a logarithmic function reflects the law of diminishing marginal utility of information value, preventing unconstrained linear growth of weights. For extremely large dH / dt (saturation period), an exponential decay function actively suppresses outliers caused by sensor failures or extreme airflow disturbances, preventing anomalous data from a single node from contaminating the entire network. Crucially, the thresholds θ1 and θ2 are dynamically adjusted based on the real-time statistical distribution of network-wide data. This means that the mapping relationship can automatically adapt to the overall activity level of the farmland environment, ensuring that each function segment operates within the optimal range, whether in the early morning when airflow is stable or in the afternoon when airflow is intense. This enables the weight model to not only accurately quantify information value but also possesses strong environmental adaptability and anti-interference capabilities, ensuring the rationality of weight allocation and the accuracy of the final concentration field reconstruction.

[0047] The research also found that, in an effort to address the drawbacks of static methods, some dynamic adjustment schemes have fallen prey to stability challenges. These schemes attempt to dynamically adjust weights based on short-term statistical characteristics of the data (such as variance), but they often lack the ability to discern dynamic characteristics across different timescales. Changes in farmland environments occur at multiple scales: from rapid, transient fluctuations on the order of seconds and minutes caused by airflow disturbances to slower, hourly and daily trends caused by crop growth and the alternation of day and night. Simple dynamic models treat all changes equally and are prone to over-responding to short-term, irregular noise (such as a gust of wind). This can lead to rapid and frequent fluctuations in the network weight configuration within a short period of time. This not only increases unnecessary computational overhead but also undermines the stability of the entire monitoring system and the credibility of the results. Existing technologies lack an intelligent control mechanism that can rapidly respond to real-world events while effectively suppressing transient noise interference. Specifically, they lack a multi-scale convergent control strategy that balances responsiveness and stability.

[0048] like Figure 4 As shown, according to one aspect of the present application, generating a convergent and stable weight configuration includes:

[0049] For the dynamic weight matrix, construct at least two time windows of different scales, short-term and long-term, track the historical evolution of the time windows, and obtain a multi-scale weight history matrix;

[0050] Analyze the data characteristics in the multi-scale weight history matrix to evaluate the weight convergence, and select or adjust the weights based on the evaluation results to output a convergent and stable weight configuration.

[0051] In one embodiment of the present application, for the dynamic weight matrix W(t), multiple time windows with different lengths are established to track each weight W i (t) historical evolution. For example, a short-term window WS (e.g., 5 minutes), a medium-term window WM (e.g., 30 minutes), and a long-term window WL (e.g., 2 hours) can be established. The lengths of these windows can even be adaptively determined based on the average response time of the network. For each time window, the historical trend of the weight is calculated using the exponentially decaying weighted average (EMA), resulting in a WS ema (t), WM ema (t) and other time series. These historical weight sequences at different scales are assembled into a multi-scale weight history matrix WH(t). This fully characterizes the historical dynamic characteristics of weights at different time granularities and serves as the data foundation for subsequent convergence assessment.

[0052] According to one aspect of the present application, analyzing the data characteristics within the multi-scale weight history matrix to evaluate the weight convergence is achieved through a multi-dimensional comprehensive criterion; the multi-dimensional comprehensive criterion integrates considerations of at least two dimensions: the statistical fluctuation stability of the weight within a first time window (short-term time window), and the convergence of the weight change trend within a second time window (long-term time window) that is larger than the first time window.

[0053] In this embodiment, the convergence of weights is not judged by a single indicator, but is comprehensively evaluated by a multi-dimensional comprehensive criterion. This criterion integrates at least the following two dimensions:

[0054] Short-term statistical fluctuation stability: This dimension focuses on the small fluctuations of weights in the short term. Specifically, it can be measured by calculating the coefficient of variation (CV) of the weights in the short-term window WS, that is, σ(WS i ) / μ(WS i ), where σ(WS i ) and μ(WSi ) are the standard deviation and mean of the weights in WS respectively. When the value is less than the preset threshold θ1, that is, σ(WS i ) / μ(WS i ) < θ1, the weights are considered stable in the short term.

[0055] Convergence of medium- to long-term change trends: This dimension focuses on whether the change trend of the weight on the medium- to long-term time scale tends to be flat. Specifically, it can be measured by calculating the absolute value of the time derivative (i.e., the rate of change) of the weight on the medium-term window WM or the long-term window WL. When the absolute value of the rate of change is less than the preset threshold θ2, that is, |dWM i / dt| <θ2, then the weight change trend is considered to be convergent.

[0056] Optionally, a third dimension can be introduced, such as long-term consistency, that is, to judge the long-term window. The specific criterion can be whether the difference between the current value in WL and the value of an earlier period is less than a threshold θ3: |WL i (t) - WL i (tT long )| <θ3, where WL i (t) is the weighted EMA value in the long-term time window, T long is a longer comparison period (e.g. 24 hours), and θ3 is the corresponding threshold. This criterion is used to evaluate whether there is a long-term, systematic drift in the weights. Finally, by weighted summing the evaluation results of multiple dimensions, a comprehensive convergence evaluation index Conv can be obtained. total_i = w c1 ×Conv 1_i +w c2 ×Conv 2_i + ...; where Conv 1_i 、Conv 2_i etc. are the Boolean results or quantitative scores of each fractal criterion. Correspondingly, the comprehensive convergence index Conv total_i The weight w of each fractal dimension in c1 , w c2 , w c3 It can also be adjusted dynamically. For example, if the system detects that the recent weights are frequently adjusted due to short-term noise interference, it can appropriately increase w c1 to enhance consideration of short-term volatility stability.

[0057] According to one aspect of the present application, the step of selecting or adjusting the weights based on the evaluation results includes:

[0058] When the evaluation results indicate that the weights do not converge, the adaptive damping coefficient is determined through a dual-factor adjustment mechanism, and the determined adaptive damping coefficient is used to constrain the amplitude of the weight adjustment; the dual-factor adjustment mechanism includes: determining the basic damping strength based on the degree of convergence assessed by a multi-dimensional comprehensive criterion, and dynamically modulating the basic damping strength using an adjustment factor that is positively correlated with the rate of change of the weight itself.

[0059] In this embodiment, when the comprehensive criterion Conv total_i Indicates a weight W i (t) When in a non-convergent state, the system does not directly apply the new weight value, but first calculates the adaptive damping coefficient Δ i (t), is used to constrain and buffer the adjustment step of the weight. i The determination of (t) is achieved through a two-factor adjustment mechanism. Specifically, Δ i (t) can be calculated according to the following formula: i (t) =Δ min + (Δ max -Δ min )×exp(-Conv total_i ×ξ)×(1+η× |V i (t)| / V max ). The adjustment mechanism of this formula includes two core factors: The first factor: the basic damping strength based on the degree of convergence. exp(-Conv total_i ×ξ) is the source of basic damping; Conv total_i It is a comprehensive convergence evaluation indicator; when the weight convergence is good (Conv total_i When the value is large, the exponential term approaches 0, making the overall damping Δ i (t) approaches the minimum value Δ min , allowing the weights to be adjusted significantly; on the contrary, when the convergence is poor (Conv total_i When the value is small, the exponential term approaches 1, which enhances the foundation damping. ξ is the convergence sensitivity coefficient, Δ max is the maximum damping coefficient, Δ min is the minimum damping coefficient. The second factor: dynamic modulation based on the rate of change. (1+η×|V i (t)| / V max ) This term is the dynamic modulation factor. V i (t) is the weight W i (t) Current rate of change (i.e., time derivative), V max is the historical maximum rate. The physical meaning of this term is that no matter how the current convergence is, as long as the change rate of the weight itself |V i(t)| is too fast, additional damping is also applied. η is the rate sensitivity coefficient. Finally, the damping coefficient Δ calculated by this two-factor mechanism is i (t), will be used for the progressive update of weights, such as W new_i (t) = W old_i (t) +Δ i (t)×ΔW i This achieves an intelligent control effect where better convergence and slower changes result in smaller damping, while worse convergence and faster changes result in larger damping, improving the stability and robustness of the entire network in complex dynamic environments.

[0060] This embodiment effectively addresses the stability dilemma during dynamic adjustment by constructing a dual-factor adaptive damping mechanism that simultaneously considers both historical convergence states and current rates of change, achieving both improved responsiveness and stability. Specifically, the first factor, based on a multi-scale comprehensive criterion, assesses weight stability from a historical and global perspective. An unstable historical state triggers strong basic damping, contributing to overall stability. The second factor monitors the current rate of change of weights in real time, applying additional dynamic braking to any excessively rapid transient changes. The synergistic effect of these two factors enables the system to intelligently distinguish between dynamic changes of different natures. In farmland environments, the system can smoothly follow sustained weight trend changes caused by the development of real disease sources with minimal damping; however, the system can suppress dramatic but brief weight pulses caused by strong winds with strong damping. This reduces response latency while also minimizing weight oscillations, enabling the network to operate stably and reliably over the long term in complex and changing environments.

[0061] According to one aspect of the present application, obtaining a redundant information identification result includes:

[0062] For each pair of sensors, the information complementarity of each pair of sensors is quantified to construct an information complementarity matrix, wherein the quantization process fuses the standard mutual information calculation results with the time-varying spatial correlation coefficient derived from the network topology state matrix;

[0063] Based on the quantized values ​​in the information complementarity matrix, redundant sensor pairs with information overlap are identified and redundant information recognition results are obtained.

[0064] Specifically, in order to quantify the information complementarity between any pair of sensors (e.g., sensors i and j), this embodiment constructs an information complementarity matrix M(t). The elements M in the matrix ij The higher the (t) value, the stronger the complementarity of the information provided by the two sensors, and vice versa. ijThe calculation of (t) combines the mutual information in information theory and the spatial correlation in the physical world. For example, its calculation can be based on the following formula: ij (t) = H(S i ) +H(S j ) - H(S i , S j ) +α ij ×C ij (t). Here H(S i ) + H(S j ) - H(S i , S j ) is the standard mutual information between sensors i and j, and α ij ×C ij (t) is the compensation term introduced and related to the time-space correlation. Among them, the time-varying spatial correlation coefficient α ij It is an important component, derived from the network topology state matrix T(t) and historical data, used to couple the correlation characteristics of the physical space to the calculation of information theory. For example, α ij Can be achieved through α ij =(1 / dist ij κ )×exp(Corr ij )×C ij (t) for dynamic calibration. ij is the physical distance between sensors i and j; κ is the distance attenuation exponent, which reflects the physical intuition that the closer the distance, the stronger the correlation; Corr ij is the signal correlation coefficient calculated based on historical data; and C ij (t) is the result of time-varying correlation evaluation. This makes the quantification of complementarity no longer a pure data statistic, but fully considers the layout and mutual influence of sensors in the physical world. After constructing the information complementarity matrix M(t), the system will identify redundant sensor pairs based on the quantified values ​​therein. For example, the redundancy R can be calculated ij (t) = 1-M ij (t) / Max(M(t)). When R ij If the value of (t) exceeds the preset redundancy threshold (for example, 0.7), the sensor pair (i, j) is marked as a redundant pair with serious information overlap, and the final redundant information identification result R(t) and the corresponding redundant grading matrix L(t) are obtained.

[0065] In an alternative embodiment, the time-varying spatial correlation coefficient α ij The calculation of Corr ij(Historical data correlation coefficient) In a specific implementation, the Pearson correlation coefficient can be used for calculation. The data source for the calculation is the synchronized and cleaned reading sequence of sensors i and j in the past long historical data window (for example, the past 24 hours). To adapt to the slow changes in the environment, the historical correlation coefficient Corr ij It is not fixed but can be updated periodically, for example, every 30 minutes or when the network topology changes significantly (such as adding or removing nodes).

[0066] According to one aspect of the present application, the calculation of mutual information in the quantization process includes solving the conditional entropy, and the solution of the conditional entropy is performed by applying a mean difference correction factor to the standard conditional entropy result; the mean difference correction factor is determined based on the normalized difference in the mean of the measurement data between the sensor pairs, making the evaluation of information complementarity more sensitive to changes in the measurement range of the sensor.

[0067] In a preferred embodiment, in order to solve the limitation that the standard mutual information algorithm may ignore the data distribution characteristics, the calculation process of mutual information is improved. i ;S j ) can be expressed as H(S i ) - H(S i |S j ), where H(S i |S j ) is the conditional entropy. The improvement of this embodiment focuses on the calculation of conditional entropy. Specifically, the improved conditional entropy H improved (S i |S j ) is calculated by the following formula: improved (S i |S j ) = H(S i |S j )×(1 +ψ×|μ(S i ) -μ(S j )| / σ(S i , S j )). Among them, H(S i |S j ) is the standard conditional entropy calculation result; ψ is the introduced mean difference correction factor, which is a positive coefficient used to adjust the sensitivity to the difference in sensor measurement range. Its recommended value range is [0.5, 1.5]. The larger the value, the stronger the penalty for the difference in measurement range; μ(S i ) and μ(S j ) are the mean values ​​of the measurement data of sensors i and j over a period of time, |μ(S i )-μ(S j)| reflects the difference in the measurement range of the two sensors; σ(S i , S j ) is the joint standard deviation of the two sensor measurements, which is used to normalize the mean difference to eliminate the impact of dimension. The motivation for introducing this mean difference correction factor is that in reality, there may be two sensors whose reading trends are highly correlated (resulting in a high standard mutual information), but their measurement domains are completely different (for example, one sensor reading fluctuates in the range of 10-20ppm, and the other fluctuates in the range of 80-90ppm). They are actually monitoring events at different concentration levels, and the information is complementary rather than redundant. The standard mutual information algorithm cannot effectively distinguish this situation. By introducing this correction factor, when the mean difference of the two sensor measurements is large, the correction term ψ×|μ(S i )-μ(S j )| / σ(S i , S j ) will increase, making the improved conditional entropy H improved The increase in the calculated mutual information eventually leads to a decrease, which more accurately reflects the lower redundancy and higher complementarity between them. This makes the evaluation of information complementarity more sensitive to the actual measurement range changes of the sensors.

[0068] This embodiment introduces a mean difference correction factor ψ into the calculation of standard conditional entropy, achieving a more accurate and physically realistic assessment of inter-sensor information redundancy. This improves network resource utilization efficiency and addresses the misjudgment problem of traditional mutual information algorithms in specific scenarios. Traditional methods focus solely on the correlation of data variation patterns, while this embodiment incorporates a normalized measurement mean difference to additionally consider the sensor's operating range. In agricultural gas monitoring, there may be two sensors: one monitoring the source closely (e.g., reading 80 ± 10 ppm) and the other monitoring diffused gas at a distance (e.g., reading 5 ± 1 ppm). Their variation trends may be highly correlated, but they monitor completely different concentration ranges, effectively providing complementary information. Traditional methods may misclassify them as redundant, while this embodiment reduces their redundancy assessment due to the significant difference in their mean values. This creates a synergistic effect, preventing high-weight sensors from being suppressed by being mistakenly identified as redundant, preserving the network's fine resolution across different concentration ranges, and ultimately improving the overall accuracy and detail of the gas concentration field reconstruction.

[0069] According to one aspect of the present application, the step of generating a final network reconstruction decision, when the decision involves physical location adjustment, includes:

[0070] The optimal new sensor locations are determined by solving a multi-objective optimization problem, where the optimization objective is to maximize a fitness function that balances the expected information acquisition gain with the corresponding physical deployment cost.

[0071] Information collection gain, at least taking into account the improvement of spatial coverage and the increase in the overall information volume of the network;

[0072] The physical deployment cost at least takes into account the physical cost of moving the sensor and the possible signal interference in the new location.

[0073] Specifically, the physical location adjustment is not always performed, but is determined by a network reconstruction trigger condition. For example, the system can be based on the size of the coverage blind spot |B(t)|, the number of severely redundant sensor pairs |L severe (t)| and the variance of the weight after stabilization Var(W stable (t)) Construct a comprehensive evaluation index η(t) = w1×|B(t)| + w2×|L severe (t)| + w3×Var(W stable (t)). Among them, w1, w2, and w3 are network reconstruction trigger weights, reflecting the priority of attention to different issues when triggering physical reconstruction. Typical initial values ​​can be w1=0.5 (focus on coverage blind spots), w2=0.3 (focus on severe redundancy), and w3=0.2 (focus on network imbalance). In actual applications, users can appropriately adjust these weights according to the main contradictions of a specific farmland (for example, a farmland with complex terrain and prominent blind spot problems). When the comprehensive evaluation index η(t) exceeds the preset threshold, the physical location adjustment decision D is triggered. trigger (t). Once triggered, the process of determining the optimal new position of the sensor is modeled as a multi-objective optimization problem, which can usually be solved by using intelligent optimization algorithms such as improved genetic algorithms. The core of this optimization algorithm is to construct a comprehensive fitness function Fitness i , used to evaluate the comprehensive benefits of moving sensor i to a candidate location. A preferred fitness function Fitness i Can be designed as: Fitness i = w f1 ×Coverage gain_i -w f2 ×Move cost_i + w f3 ×Info gain_i -w f4 ×Interference penalty_i Among them, Information Gain Coverage gain_iIt represents the improvement in spatial coverage brought about by moving the sensor. It can be quantified by analyzing the reduction in the original coverage blind area B(t) after the movement, which directly corresponds to the improvement in spatial coverage. gain_i It represents the increase in the overall network information volume brought about by the mobile sensor. It can be calculated based on the overall information entropy of the network after the movement, or the degree of optimization of the information complementarity matrix M(t), reflecting the increase in the overall network information volume. cost_i Represents the physical cost of moving the sensor, which can be a quantitative value related to the moving distance, terrain complexity, required manpower, etc., reflecting the physical cost of moving the sensor. penalty_i The penalty term representing the signal interference or conflict that may occur after the sensor is deployed to a new location can be evaluated by the minimum distance between the new location and other sensors or known interference sources. For example, the constraint min j≠i |P new_i -P j |≥D min To avoid physical collision or signal crosstalk, it reflects the signal interference that may occur in the new location. new_i is the candidate new position coordinate of sensor i, P j is the current physical deployment location of other sensors j in the network, D min is the minimum safety distance threshold. The weight coefficient w in the function f1 , w f2 , w f3 , w f4 It is not fixed, but can be dynamically adjusted based on the specific status of the current network (for example, whether the coverage problem is more prominent or the deployment cost is more sensitive) to achieve a balance between different optimization goals. By maximizing this fitness function, the algorithm can search for a series of optimal candidate locations that take into account both information benefits and deployment costs, and form the final physical location adjustment recommendation P new This ensures the scientific and practical nature of network reconstruction decisions and avoids the drawback of only pursuing theoretical information gain while ignoring actual engineering constraints.

[0074] In an optional embodiment, the specific implementation of the improved genetic algorithm includes: To accelerate algorithm convergence, its population initialization is not completely random. A portion of individuals (e.g., 30%) are strategically initialized at the center of the coverage blind spot identified in the previous step and its adjacent area, while the remaining individuals are randomly generated within the feasible deployment area, ensuring that the initial population has a tendency to solve key problems. A crossover operator based on sensor geographic location clustering is used. Before performing the crossover operation, the algorithm first performs a rapid K-means clustering of the sensor locations in the current population. The crossover operation preferentially selects parent individuals from different clusters for exchanging gene segments (i.e., location coordinates). This encourages the exchange of excellent genes from different regions, helps promote global search, and prevents the algorithm from prematurely converging to a local optimal solution. An adaptive mutation rate mechanism is employed. The algorithm monitors changes in the population's fitness. If it detects no significant improvement in the population's average fitness over multiple generations (e.g., 10 generations), the system automatically multiplies a base mutation probability (e.g., 0.05) by a dynamic gain factor (e.g., 1.5) to temporarily increase the probability of mutation and help the population escape any local optimal point it may be trapped in. In a typical configuration, the genetic algorithm's population size can be set to 100, the maximum number of iterations to 200 generations, and the crossover probability to 0.8.

[0075] According to another aspect of the present application, a method for dynamically reconfiguring a farmland gas sensor network may also include:

[0076] S1. Read the real-time detection data and network status parameters of the sensor array, perform data cleaning and preliminary processing, and obtain the standardized sensor data set and network topology status matrix.

[0077] In this embodiment, the raw sensor array data {R i (t), i=1, 2, ..., N} and environmental parameter data {E i (t)}, through timestamp alignment and data format unification, we get the synchronized raw data matrix R(t) and the environmental parameter matrix E(t). Read the synchronized raw data matrix R(t), and calculate the data credibility score of each sensor by combining the improved 3σ criterion with the sliding window anomaly detection. i (t), obtain the data quality assessment matrix Q(t) and the abnormal sensor tag list A(t). Read the synchronized original data matrix R(t), the data quality assessment matrix Q(t) and the abnormal sensor tag list A(t), and correct the abnormal data through the adaptive interpolation algorithm based on neighborhood correlation to obtain the cleaned data matrix D(t). Specifically, read the cleaned data matrix D(t) and the sensor position information P, and use the distance decay function Corr ij = exp(-dist ij / λ)×ρ ij Calculate the spatial correlation between sensors, where ρ ij is the correlation coefficient of historical data, and the spatial correlation matrix Spatial is obtained Corr (t) and the valid neighborhood index table Neighbor index (t). Read the abnormal sensor tag list A(t) and the spatial correlation matrix Spatial Corr (t) and data quality assessment matrix Q(t), and calculate W through dynamic weight repair_ij =Spatial Corr_ij ×Q j (t)×(1-|A j (t)|) Assign a neighborhood repair weight to each abnormal sensor and obtain the repair weight matrix W repair (t) and the trusted neighbor list Trusted Neighbors (t). Read the repair weight matrix W repair (t), Trusted Neighborhood List Neighbors (t) and synchronized original data matrix R(t), through the weighted interpolation algorithm D repaired_i (t) =ΣW repair_ij ×D j (t) / Σ W repair_ij Repair the abnormal data and fuse it with the original data according to the quality weight to obtain the repaired data matrix D repaired (t), and this data is used as the cleaned data matrix D(t). The cleaned data matrix D(t) and sensor location information P are read, and the network connection relationship is constructed through spatial distance and signal correlation calculation to obtain the network topology state matrix T(t) and the standardized sensor data set S(t).

[0078] S2. Read the standardized sensor data set, calculate the information contribution of each sensor through the information entropy-time derivative weight model, and obtain the sensor information entropy matrix and time-varying correlation evaluation results.

[0079] In this embodiment, the standardized sensor data set S(t) is read and the information entropy H of each sensor is calculated by sliding window Shannon entropy. i (t), and then calculate the time derivative dH i / dt, and obtain the sensor information entropy sequence H(t) and the information entropy change rate matrix dH / dt. Read the information entropy change rate matrix dH / dt and the data quality assessment matrix Q(t), and use the nonlinear mapping function W i (t) = f(dH i / dt,Q i (t), β i ) calculates the dynamic weight, where β iThe sensor characteristic coefficient is obtained, and the dynamic weight matrix W(t) and the weight change trend ΔW(t) are obtained. Specifically, the information entropy change rate matrix dH / dt and the data quality assessment matrix Q(t) are read, and β is calculated by analyzing the sensor response characteristics. i =α base ×(1 +γ×Var(H i_history ))×Q avg_i , where α base is the basic coefficient, γ is the variance sensitive factor, and the sensor characteristic coefficient vector β(t) and characteristic stability evaluation Stability are obtained. Score (t). Read the information entropy change rate matrix dH / dt, sensor characteristic coefficient vector β(t) and data quality assessment matrix Q(t), and use the piecewise nonlinear function f(x) = {a1x 2 + b1x + c1 (x≤θ1), a2ln(x) + b2 (θ1 <x≤θ2),a3e (-x / c3) + b3 (x>θ2)} for mapping calculation, where a, b, c are different parameters, and the thresholds θ1 and θ2 are adaptively determined based on the data distribution to obtain the segmented mapping parameter set Params mapping (t) and the mapping interval partition result Zone division (t). Read segment mapping parameter set Params mapping (t), Mapping interval division result Zone division (t) and characteristic stability evaluation Score (t), through W i (t) = f(dH i / dt,Q i (t), β i )×Stability Scorei (t) Calculate the original weight and then perform group normalization W norm_i (t) = W i (t) / ΣW j (t), get the original dynamic weight vector W raw (t) and the normalized weight vector W norm (t). Read the normalized weight vector W norm (t) and the original dynamic weight vector W raw (t), through the first-order difference ΔW i (t) = W norm_i (t) - W norm_i (t-1) and moving average filter W smooth_i (t) =Σ(k=0 to n-1)W norm_i(tk)×exp(-k / τ) / Σ exp(-k / τ) is smoothed, where τ is the time scale factor, to obtain the dynamic weight matrix W(t) and the weight change trend ΔW(t). Read the dynamic weight matrix W(t) and the standardized sensor data set S(t), and calculate IG through weighted information gain i (t) = W i (t) × log(P(S i |Context) / P(S i )), and obtain the sensor information entropy matrix IE(t) and the time-varying correlation evaluation result C(t).

[0080] In an optional embodiment, the construction of the information entropy time derivative weight model can also be: performing multi-order calculation and feature extraction of the information entropy time derivative: calculating the first-order derivative: dH i / dt = [H i (t) - H i (t-Δt)] / Δt; calculate the second-order derivative: d 2 H i / dt 2 = [dH i / dt(t) - dH i / dt(t-Δt)] / Δt; construct the derivative eigenvector: D i (t) =[dH i / dt,d 2 H i / dt 2 ,sign(dH i / dt),|dH i / dt| / max(dH / dt)]. Frequency domain analysis and pattern recognition of time derivative: i / dt sequence short-time Fourier transform: STFT{dH i / dt}; extract the main frequency component: f dominant_i = argmax(|STFT{dH i / dt}|); Identify the change pattern: Pattern i= {periodic, sudden, gradual, chaotic}. Adaptive boundary determination for nonlinear mapping: Real-time clustering based on information entropy derivative distribution: Use the DBSCAN algorithm to cluster dH / dt; Cluster boundary as piecewise function threshold: θ1 = boundary(cluster1, cluster2); Boundary fuzzification: Introduce the transition interval [θ1-ε, θ1+ε] and use sigmoid smoothing. Cluster1 is a low-variance group, cluster2 is a high-variance group, and ε is the transition width parameter. Online learning and parameter update of mapping function: Establish mapping error feedback: ε map (t) = |W predicted (t) - W optimal (t) | Update the mapping parameters using reinforcement learning: a new = a old +α·▽ε map / ▽a; introduce the experience pool to save the historical optimal mapping parameter combination. predicted (t) is the previous weighted prediction value, W optimal (t) is the optimal weight of the current target, a old is the current parameter of the mapping function, α is the learning rate, and ▽ is the gradient.

[0081] S3. Read the sensor information entropy matrix and time-varying correlation evaluation results, build the information relationship between sensors through the complementarity evaluation algorithm, and obtain the information complementarity matrix and redundant information identification results.

[0082] In this embodiment, the sensor information entropy matrix IE(t) and the time-varying correlation evaluation result C(t) are read, and the complementary index M is calculated by the mutual information algorithm. ij (t) = H(S i ) + H(S j ) - H(S i , S j ) +α ij ×C ij (t), where α ij Specifically, the sensor information entropy matrix IE(t), the time-varying correlation evaluation result C(t) and the sensor position information P are read, and α is calculated comprehensively by spatial distance and signal correlation. ij = (1 / dist ij κ )×exp(Corr ij )×C ij (t), where κ is the distance decay exponent, and the spatial correlation coefficient matrix α(t) and the correlation intensity grading result Level are obtained. α(t). Read the sensor information entropy matrix IE(t) and the standardized sensor data set S(t), and calculate H through conditional entropy improved (S i |S j ) = H(S i |S j )×(1 +ψ×|μ(S i ) -μ(S j )| / σ(S i , S j )), where ψ is the mean difference correction factor, and the improved conditional entropy matrix H is obtained condimproved (t) and information dependency matrix Dependence(t). Read the improved conditional entropy matrix H condimproved (t), spatial correlation coefficient matrix α(t) and information dependency matrix Dependence(t), and calculate M through multi-dimensional complementarity ij (t)=ω1×[H(S i )+H(S j )-H improved (S i , S j )]+ω2×[α ij ×C ij (t)]+ ω3×[1-Dependence ij (t)], where ω1, ω2, and ω3 are dynamic weight coefficients, and the multidimensional complementary component M is obtained. components (t) and fusion weight coefficient ω(t). Read the multidimensional complementary component M components (t) and fusion weight coefficient ω(t), corrected by symmetry M sym_ij = (M ij + M ji ) / 2 and stability filter M stableij (t) = Φ×M sym_ij (t) + (1-Φ)×M stableij (t-1) is processed, where Φ is the update rate parameter, and the information complementarity matrix M(t) and the matrix stability index Stability are obtained. M (t). Read the information complementarity matrix M(t) and the network topology state matrix T(t), and determine R by the redundancy threshold ij (t) = 1-M ij(t) / Max(M(t)) identifies redundant sensor pairs and classifies them into three levels according to the degree of redundancy: mild (0.3-0.5), moderate (0.5-0.7), and severe (>0.7). The redundant information identification result R(t) and the redundant classification matrix L(t) are obtained. Specifically, the information complementarity matrix M(t) and the network topology state matrix T(t) are read, the quantile distribution of the M matrix is ​​calculated through statistical analysis, and the threshold θ is adaptively set. light = P75(M), θ medium = P50(M),θ severe = P25(M), and get the adaptive threshold set Θ adaptive (t) and distribution statistical parameters Stats M (t). Read the information complementarity matrix M(t) and the adaptive threshold set Θ adaptive (t), calculate R by redundancy ij (t) = 1-M ij (t) / Max(M(t)) and grading judgment Level ij = {1, ifR ij <θ light ; 2, ifθ light ≤R ij <θ medium ; 3, ifθ medium ≤R ij <θ severe ; 4, if R ij ≥θ severe}, get the redundancy matrix R degree (t) and the redundancy level matrix Level redundancy (t). Read the redundancy matrix R degree (t), Redundancy level matrix Level redundancy (t) and spatial correlation coefficient matrix α(t), the high redundant sensor pairs are grouped by hierarchical clustering algorithm, and the comprehensive score Score ij = R ij (t) ×α ij (t)×Priority factor_ij Sorting, get the redundant information identification result R(t) and redundant grading matrix L(t). Read the information complementarity matrix M(t), sensor position information P and redundant grading matrix L(t), and calculate the spatial coverage efficiency Cov by combining the Voronoi diagram with the information weight i =Area i ×W i (t) / Total Area , and obtain the spatial coverage efficiency matrix Cov(t) and coverage blind area identification result B(t).

[0083] Optionally, the information complementarity matrix can also be constructed by: performing spatiotemporal coupling modeling of spatial correlation coefficients: constructing spatiotemporal correlation tensor: α tensor [i, j, t] = spatial factor × temporal factor × coupling factor ; spatial factor spatial factor The refinement takes into account the influence of obstacles and wind direction; the time factor temporal factor The refinement takes into account diurnal and seasonal variations; the coupling factor factor The time-varying coupling adjustment term α in coupling = exp(-|Ξα spatial / Ξt|), where α spatial is the spatial factor, and Ξ is the partial derivative. Uncertainty quantification of complementarity index: Calculation of complementarity confidence interval: M ij ±σ M × t critical , where σ M is the complementarity index M ij The standard deviation of critical is the critical value of the t distribution corresponding to the confidence interval; introduce fuzzy complementarity: use triangular fuzzy numbers to represent M ij = (M low , M mid , M high ); Confidence-based decision-making: only when confidence(M ij Complementarity is considered only when the complementarity matrix is ​​> 0.95. Topological structure analysis of the complementarity matrix: Construct a complementary network graph: node = sensor, edge weight = M ij ; Calculate network characteristics: clustering coefficient, betweenness centrality, community structure; Identify key nodes: Evaluate sensor importance based on PageRank algorithm.

[0084] S4. Read the information complementarity matrix and redundant information identification results, optimize the weights through a multi-time-scale progressive adjustment mechanism, and obtain a convergent and stable weight configuration and weight adjustment strategy.

[0085] In this embodiment, the dynamic weight matrix W(t) is read, and the weight history sequences of three time scales, namely the short-term window WS (5min), the medium-term window WM (30min), and the long-term window WL (2hour), are established. The weight trend of each scale is calculated by exponential decay weighted average to obtain the multi-scale weight history matrix WH(t) and the weight change rate matrix V(t). Specifically, the dynamic weight matrix W(t) is read, and the window parameters are determined by data change rate analysis: the short-term window WS length = max(5min, 2×Avg response_time ), where Avg response_time is the average response time, mid-term window WM length = 6×WS length , long-term window WL length = 4×WM length , get the time window parameter set Window params (t) and window length adaptation result Adaptive length (t). Read the dynamic weight matrix W(t) and the time window parameter set Window params (t), through the attenuation coefficient λS = 2 / (WS length +1), λM = 2 / (WM length +1), λL = 2 / (WL length +1) and weighted average calculation EMA scale_i (t) = λ scale ×W i (t) + (1-λ scale )×EMA scale_i (t-1), get the attenuation parameter vector λ vector (t) and three-scale EMA weight WS ema (t), WM ema (t), WL ema (t). Read the three-scale EMA weight WS ema (t), WM ema (t), WL ema (t), calculate the rate of change of each scale VS by difference i (t) = dWS ema_i / dt,VM i (t) = dWM ema_i / dt,VL i (t) = dWL ema_i / dt, and perform trend decomposition V trend_i (t) = w s ×VS i + w m ×VM i+ w l ×VL i , get the multi-scale change rate matrix V multiscale (t) and weight trend vector W trend (t). Read the three-scale EMA weight WS ema (t), WM ema (t), WL ema (t) and the multi-scale rate of change matrix V multiscale (t), by matrix assembly WH(t) = [WS ema (t); WM ema (t); WL ema (t)] and inter-scale correlation calculation Corr scales_ij = Corr(WS ema_i , WM emaj , WL ema_k ), get the multi-scale weight history matrix WH(t) and weight change rate matrix V(t). Read the multi-scale weight history matrix WH(t) and weight change rate matrix V(t), and use the convergence criterion Conv i (t) = σ(WS i ) / μ(WS i ) <θ1 AND |dWM i / dt| < θ2 to judge the weight stability and apply a damping coefficient Δ to the weight that does not converge i (t) = exp(-|V i (t)| / V max ), and obtain the convergence state evaluation C state (t) and the damping coefficient matrix Δ(t). Specifically, read the multi-scale weight history matrix WH(t) and the weight change rate matrix V(t), and calculate the coefficient of variation benchmark CV in the stable state through statistical analysis of historical data baseline and rate of change reference dW baseline , set θ1 = 1.5×CV baseline , θ2 = 2.0×dW baseline , get the baseline statistical parameter Baseline stats (t) and the adaptive threshold Θ convergence (t). Read the multi-scale weight history matrix WH(t) and the adaptive threshold pair Θ convergence (t), through the multi-dimensional criterion Conv 1_i = σ(WS i ) / μ(WS i ) <θ1,Conv 2_i = |dWM i / dt| <θ2,Conv 3_i = |WLi (t) -WL i (tT)| <θ3Calculate the comprehensive convergence index Conv total_i = w c1 ×Conv 1_i + w c2 ×Conv 2_i + w c3 ×Conv 3_i , get the fractal dimension convergence state Conv individual (t) and comprehensive convergence evaluation Conv comprehensive (t). Read the weight change rate matrix V(t) and the comprehensive convergence evaluation Conv comprehensive (t), calculate Δ by damping strength i (t) = Δ min +(Δ max - Δ min ) × exp(-Conv total_i ×ξ) × (1 + η× |V i (t)| / V max ), where ξ is the convergence sensitivity coefficient, η is the rate sensitivity coefficient, and the basic damping coefficient Δ base (t) and rate adjustment factor Rate factor (t). Read the basic damping coefficient Δ base (t), rate adjustment factor Rate factor (t) and fractal dimension convergence state Conv individual (t), through the stability test Stability check_i = |Δ i (t) - Δ i (t-1)| <Δ threshold and outlier correction Δ final_i =median filter (Δ base_i , window=3), and get the convergence state evaluation C state (t) and damping coefficient matrix Δ(t). Read the damping coefficient matrix Δ(t), weight change rate matrix V(t) and dynamic weight matrix W(t), and gradually adjust the formula W new_i (t) = W old_i (t) + Δ i (t) × ΔW i (t) × λ(t), where λ(t) is the global learning rate, and the convergent and stable weight configuration W is obtained. stable(t) and weight adjustment strategy Strategy(t). To ensure that the weight adjustment converges quickly in the early stage and tends to be stable in the later stage, the global learning rate λ(t) can adopt a strategy of decaying over time. For example, using exponential decay λ(t) = λ0 × 0.999 t , where t is the number of iterations and the initial learning rate λ0 can be set to 0.1.

[0086] Optionally, the convergence criterion of the three-scale time window can also be: Cross-validation mechanism of time scale: Define the consistency index between scales: Consistency SM = corr(WS trend , WM trend ); Calculation scale conflict: Conflict ij =|decision scale_i - decision scale_j |; Establish scale weight adjustment rules: When Conflict > threshold, reduce the conflict scale weight. Dynamic weight allocation of convergence criteria: Dynamically adjust the criterion weight based on historical convergence results: w ci (t+1) = w ci (t) × Performance i (t); Introduce the criterion credibility: Trust i = exp(-variance(Conv i_history )); Construct adaptive criterion combination: Conv adaptive = Σ(w ci × Trust i × Conv i Identification and processing of abnormal convergence modes: Definition of abnormal convergence modes: oscillation, divergence, pseudo-convergence; Oscillation detection: |W(t) - W(t-2T)| <ε AND |W(t) - W(tT)| > ε; Divergence detection: monotonic increase (variance(W history )); Abnormal mode special processing strategy: oscillation → increase damping, divergence → reset parameters. Nonlinear scheduling strategy of damping coefficient: establish damping scheduling function: Δ schedule (t) = Δ base × (1 + sin(2π·t / T cycle ) × modulation(Conv state )); Introduce learning rate decay: λ(t) = λ0 / (1 + decay rate × t); Design emergency braking mechanism: when |dW / dt| > criticalthreshold When Δ = Δ emergency WS trend For a convergent trend data series on a small scale (such as a short-term window), WM trend For the convergence trend data series on the mesoscale (e.g. medium-term window), decision scale_i and decision scale_j are the convergence criteria or decision outputs at the i-th and j-th scales respectively; w ci (t) is the weight of the i-th scale criterion at time t; Performance i (t) is the actual convergence effect evaluation value of the i-th scale criterion at time t; Trust i is the credibility of the i-th scale criterion, W(t) is the weight or optimization variable value at the current time point; ε is the tolerance threshold for oscillation detection; T cycle is the cycle length, modulation(Conv state ) is the modulation function of the current convergence state, decay rate is the learning rate attenuation parameter, critical threshold To determine the threshold value of drastic changes, Δ emergency It is the damping value during emergency braking.

[0087] S5. Read the converged and stable weight configuration and weight adjustment strategy, generate a network optimization plan through the reconstruction decision algorithm, and obtain physical location adjustment suggestions and logical weight reconstruction configuration.

[0088] In this embodiment, the converged and stable weight configuration W is read. stable (t), coverage blind spot identification result B(t) and redundancy classification matrix L(t), and the comprehensive evaluation index η(t) = w1×|B(t)| + w2×|L severe (t)| + w3×Var(W stable (t)) Determine whether reconstruction is needed and obtain reconstruction trigger decision D trigger (t) and reconstruction urgency P level (t). Read the reconstruction trigger decision D trigger (t), coverage blind spot identification result B(t) and sensor position information P, optimize the sensor position through the improved genetic algorithm to minimize the coverage blind spot and maximize the information gain, and calculate the position adjustment benefit Cost move_i =Gain info_i - Cost physical_i , where Gain info_i is the information gain brought by the adjustment of the i-th candidate position, Cost physical_iThe actual physical cost required to move the sensor to the candidate position; the physical position adjustment suggestion P new Specifically, read the coverage blind spot identification result B(t), sensor position information P and information complementarity matrix M(t), and use the multi-objective fitness function Fitness i =w f1 ×Coverage gain_i -w f2 ×Move cost_i + w f3 ×Info gain_i -w f4 ×Interference penalty_i Design evaluation indicators, where each weight is dynamically adjusted based on the current network state to obtain the fitness weight vector w fitness (t) and multi-objective evaluation matrix Objectives(t). Read the sensor position information P, coverage blind spot identification result B(t) and multi-objective evaluation matrix Objectives(t), through geographic constraints Geo constraint : |P new_i -P old_i | ≤ D max and interference constraints constraint : min j≠i |P new_i -P j | ≥ D min Establish a search space and use the crossover mutation operator to optimize the search to obtain the constrained search space Search space (t) and candidate position set Candidates(t). Read the candidate position set Candidates(t), fitness weight vector w fitness (t) and the information complementarity matrix M(t), through detailed benefit analysis Cost move_i =(Coverage improvement_i + Info enhancement_i ) - (Physical cost_i + Deployment risk_i ) Calculate the net benefit of each adjustment plan, where Coverage improvement_i is the coverage blind area improvement value brought about by the i-th position change, Info enhancement_i The information complementarity improved by adjusting the i-th position, Physical cost_i Deployment risk_iThe system risk assessment value caused by the deployment adjustment is obtained by prioritizing and obtaining the physical location adjustment suggestion P new And adjust the priority list Priority(t). Read the converged and stable weight configuration W stable (t), weight adjustment strategy Strategy(t) and physical location adjustment suggestion P new , logical weight compensation W for sensors that cannot be physically adjusted logic_i (t) = W stablei (t) × Comp factori (t), where the compensation factor is calculated based on the position change of the neighboring sensors, and the logical weight reconstruction configuration W is obtained. logic (t) and reconstruct the execution plan Plan(t).

[0089] S6. Read the physical location adjustment suggestions, logical weight reconstruction configuration, and standardized sensor data set, calculate the gas distribution through the reconstructed network configuration, and obtain the optimized gas concentration distribution map and network performance evaluation report.

[0090] In this embodiment, the standardized sensor data set S(t) and the physical position adjustment suggestion P are read. new and logical weight reconstruction configuration W logic (t), respectively calculate the information coverage, detection accuracy and response time before and after reconstruction, and obtain the performance improvement evaluation report Performance(t) and reconstruction effect quantitative indicators Metrics(t). Read the logical weight reconstruction configuration W logic (t), standardized sensor dataset S(t) and spatial coverage efficiency matrix Cov(t), and spatial interpolation C(x, y, t) = Σ W is performed by weighted Kriging interpolation algorithm combined with optimized weights logic_i (t)×S i (t)×K(xx i ,yy i ), and obtain the optimized gas concentration distribution map C field (t). K( ) is the kernel function in Kriging interpolation, which represents the spatial point (x i ,y i ) to the target point (x, y). Specifically, read the logical weight reconstruction configuration W logic (t), the standardized sensor dataset S(t) and the spatial coverage efficiency matrix Cov(t), the covariance function γ modified by the weight weighted (h) = γ standard (h) × (1 +ρ × W relative (h) Calculate the spatial correlation, where W relative is the relative weight at distance h, ρ is the weight modulation coefficient, γstandard (h) is the standard covariance function or semivariogram, and the weighted covariance parameter Cov is obtained. paramsweighted (t) and weighted influence factor W influence (t). Read the weighted covariance parameter Cov paramsweighted (t), weighted impact factor W influence (t) and sensor position information P, through multi-layer weight fusion W total_i (x, y) = W logic_i × W spatial_i (x, y) × W reliability_i Calculate the overall weight, where W spatial_i is the spatial matching weight, W reliability_i is the data reliability weight; then perform Kriging interpolation C(x, y, t) =ΣW total_i (x, y) × S i (t), and get the overall weight field W totalfield (t) and the interpolation intermediate result C intermediate (t). Read the interpolation intermediate result C intermediate (t), the overall weight field W totalfield (t) and the coverage blind area identification result B(t), through boundary constraint processing and Gaussian smoothing filter C smooth (x, y, t) = G σ * C intermediate (x, y, t) is post-processed, where G σ is a Gaussian smoothing filter operator, and the adaptive standard deviation σ is adaptively adjusted based on the local weight density to obtain the optimized gas concentration distribution map C field (t) and concentration field quality assessment Quality field (t). Read the reconstruction effect quantitative index Metrics(t) and the optimized gas concentration distribution map C field (t) and the reconstruction execution plan Plan(t), and update the reconstruction parameter θ through the feedback learning mechanism new = θ old + α × ▽Performance(θ), and generate optimization suggestions for the next cycle, obtain the network performance evaluation report Report(t) and continuous optimization parameter update Params update (t).

[0091] In a specific embodiment of the present application, the specific execution process of the weight convergence control strategy of the multi-scale time window is as follows: Assume that for a certain sensor i in the network, the dynamic weight W i(t) The sequence in the last 10 time steps (e.g., one data point per minute) is: [0.25, 0.26, 0.28, 0.35, 0.32, 0.30, 0.38, 0.45, 0.42, 0.40]. According to the settings of this application, the system uses the following parameters for convergence control: Window parameter: short-term window WS length = 5 time steps, medium-term window WM length = 10 time steps; convergence criterion threshold: short-term volatility threshold θ1 = 0.15, medium-term trend threshold θ2 = 0.02; damping adjustment parameter: Δ min =0.2, Δ max =1.0, convergence sensitivity coefficient ξ=2.0, rate sensitivity coefficient η=0.5. Status to be updated: current value of weight W old_i (t) = 0.40, an external event causes the algorithm to give a larger weight adjustment recommendation ΔW i (t)=+0.10. Current weight change rate V i (t) can be approximated as (0.40 -0.42) = -0.02, and the historical maximum change rate V max is 0.08 (occurs during the jump from 0.30 to 0.38).

[0092] Analyze the data in the multi-scale weight history matrix WH(t). For the short-term window WS (i.e., the last 5 data points: [0.30, 0.38, 0.45, 0.42, 0.40]): calculate the mean μ(WS i ) = (0.30 + 0.38 + 0.45 + 0.42 +0.40) / 5 = 0.39. Calculate the standard deviation σ(WS i ) ≈ 0.055. For the medium-term window WM (all 10 data points), calculate its overall change trend dWM i / dt, which can be approximated as the linear regression slope of the sequence, is approximately 0.0167. The system evaluates based on a multi-dimensional comprehensive criterion: Dimension 1 (short-term volatility): Calculate the coefficient of variation CV = σ(WS i ) / μ(WS i ) ≈ 0.055 / 0.39 ≈ 0.141. Since 0.141 < θ1(0.15), this dimension is considered stable. Conv 1_i = 1. Dimension 2 (Medium-term Trend): Evaluate the trend of change | dWM i / dt| ≈ 0.0167. Since 0.0167 <θ2(0.02), this dimension is considered converged. Conv 2_i = 1. Assume that the weights of the two dimensions are equal (w c1 =0.5, wc2 =0.5), then the comprehensive convergence index Conv total_i = 0.5* 1 + 0.5 * 1 = 1.0. The evaluation results show that the current state of the weight is convergent. According to the evaluation results Conv total_i = 1.0, the system calculates the damping coefficient Δ through a dual-factor adjustment mechanism i (t): First factor (basic damping): exp(-Conv total_i ×ξ) = exp(-1.0 × 2.0) = exp(-2) ≈ 0.135. Because of the good convergence (Conv total_i =1), so the smaller the factor, the weaker the basic damping. The second factor (rate modulation): (1 + η× |V i (t)| / V max ) = (1 + 0.5 × |-0.02| / 0.08) = 1 + 0.5 × 0.25 = 1.125. Final damping coefficient: Δ i (t) = Δ min + (Δ max -Δ min ) × [first factor] × [second factor] ≈ 0.2 + (1.0 - 0.2) × 0.135 × 1.125 = 0.2 + 0.8 × 0.152 ≈ 0.2 + 0.122 = 0.322. The calculated result is Δ i (t) ≈ 0.322, which is between 0 and 1 and is an effective damping coefficient. The system uses the calculated damping coefficient Δ i (t) to constrain the adjustment range of the weight: the original adjustment amount: ΔW i (t) = +0.10; Adjustment after damping: Δ i (t) ×ΔW i (t) ≈ 0.322 × 0.10 = 0.0322; Final new weight: W new_i (t) = W old_i (t) + Damped Adjustment ≈ 0.40 + 0.0322 = 0.4322. As can be seen, although the algorithm suggests a significant adjustment of +0.10, the convergence control mechanism constrains this adjustment with a damping factor of approximately 0.322, ultimately smoothing the actual adjustment to approximately +0.0322. This example demonstrates how quantitative evaluation and adaptive damping can effectively prevent drastic weight jumps, ensuring the stability and reliability of the entire sensor network dynamic reconstruction process.

[0093] In a specific deployment process of this application, when the system is first started or cold-started, a series of key parameters and variables need to be initialized: Dynamic weight matrix W (t=0): In the absence of any prior knowledge, the initial weights of all sensors should be set equal, that is, for a network containing N sensors, the initial weight of each sensor W i (0) are all 1 / N. Information complementarity matrix M(t=0): can be initialized to a zero matrix or an identity matrix. Multi-scale weight history matrix WH(t=0): when historical data is insufficient, it can be filled with the initial weight value of each sensor.

[0094] In one embodiment of the present application, the system has emergency response capabilities for large-scale sensor failures. When the number of sensors in the network that fail (e.g., due to physical damage or communication interruption exceeding a preset long period, such as one hour) exceeds a preset threshold (e.g., 30% of the total number), the system automatically triggers emergency mode. In this mode, the dynamic reconstruction algorithm is suspended to prevent erroneous optimization decisions based on severely incomplete data. Accordingly, the network's data fusion weights automatically fall back to a static, preconfigured weight model based on the sensor's physical location (e.g., the inverse distance or a preset Voronoi diagram weight). Simultaneously, the system increases its reliance on the neighborhood interpolation compensation algorithm to minimize data blind spots and sends a high-level alert to the user management platform, indicating that network integrity has been severely compromised and manual intervention is required. Furthermore, the present application has the capability to operate in a distributed manner to address situations where sensor network partitioning, potentially caused by communication failures (e.g., relay node failure), occurs—that is, the network is divided into two or more isolated islands that cannot communicate with each other. Specifically, the present application, deployed on edge computing nodes or gateways, will run independently within each isolated island. In other words, each sub-network will only perform local dynamic weight optimization and logic reconstruction based on the sensor data that can still communicate within it. This ensures that local failures will not lead to the paralysis of the entire monitoring system, and the rest of the network can still maintain the optimal operating state, demonstrating high fault tolerance. For the more common case of data communication interruption of a single sensor, the system has also designed a corresponding processing mechanism. When the data of a sensor fails to be successfully received within a preset short period of time (for example, 5 consecutive sampling cycles), the system will temporarily mark it as offline. In the weight calculation, its dynamic weight W i (t) will be temporarily reset to zero, removing it from subsequent data fusion calculations. The data for its original monitoring area will be completely estimated and covered by its neighboring sensors using a weighted interpolation compensation algorithm. Once the sensor resumes communication and uploads valid data, the system automatically restores it to an online state and reintegrates it into the dynamically reconstructed calculation process. This ensures that the system can smoothly handle temporary and intermittent failures of individual nodes without significantly impacting the overall operation of the network.

[0095] To validate the effectiveness of this application, a simulation-based performance evaluation case study is provided. Experimental design: A comparative simulation experiment was conducted in a 1-hectare (100 m x 100 m) simulated cornfield environment, where a simulated disease source continuously releasing a specific gas was located. The control group employed a conventional 36-sensor grid layout (6 x 6 array) and used static weights based on the inverse of spatial distance for data fusion. The experimental group employed the same number of 36 sensors as the control group, initially randomly distributed, but subsequently employed the dynamic reconstruction method described in this application, including dynamic weight allocation, redundancy analysis, and physical position adjustment based on a genetic algorithm. Performance metrics included: average source location error (meters): the Euclidean distance between the predicted source location output by the algorithm and the true source location; and anomaly detection time (seconds): the time from the start of gas release from the simulated source to the network's first identification of an abnormal concentration and the issuance of an alarm. After 50 repeated experiments with simulated sources placed at different random locations and taking the average, the experimental results showed that the control group achieved an average location error of 5.1 meters and a detection time of 85 seconds. The experimental group using this application, by dynamically optimizing the network layout and data fusion weights, ultimately reduced the average positioning error to 1.9 meters and the detection time to 32 seconds. This example demonstrates its beneficial effect in improving the accuracy and timeliness of farmland gas monitoring.

[0096] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A method for dynamic reconstruction of a farmland gas sensor network, characterized in that: include: Obtain raw sensor array data, preprocess it, and generate standardized sensor data sets and network topology state matrices; Based on the standardized sensor data set, the dynamic weight matrix is ​​calculated by quantifying the dynamic change of information entropy, and the information complementarity between sensors is evaluated in combination with the network topology state matrix to obtain the redundant information identification result. Apply multi-time-scale convergence control to the dynamic weight matrix to generate convergent and stable weight configurations; The converged and stable weight configuration and redundant information identification results are integrated to generate the final network reconstruction decision; When the decision involves physical relocation, this includes: The optimal new sensor locations are determined by solving a multi-objective optimization problem, where the optimization objective is to maximize a fitness function that balances the expected information acquisition gain with the corresponding physical deployment cost. Information collection gain, at least taking into account the improvement of spatial coverage and the increase in the overall information volume of the network; The physical deployment cost at least takes into account the physical cost of moving the sensor and the possible signal interference in the new location; Calculate the dynamic weight matrix, including: Calculate the Shannon information entropy of the time series of each sensor in the standardized sensor dataset; Perform time series difference processing on the Shannon information entropy of the time series to obtain its time derivative and form the information entropy change rate matrix; among them, the time derivative is used to characterize the dynamic change of sensor information value; Perform nonlinear mapping on the information entropy change rate matrix to convert the information entropy change rate of each sensor into a dynamic weight to obtain a dynamic weight matrix; Generate convergent and stable weight configurations, including: For the dynamic weight matrix, construct at least two time windows of different scales, short-term and long-term, track its historical evolution, and obtain a multi-scale weight history matrix; Analyze the data characteristics in the multi-scale weight history matrix to evaluate the weight convergence, and select or adjust the weights based on the evaluation results to output a convergent and stable weight configuration; The steps for selecting or adjusting weights based on the evaluation results include: When the evaluation results indicate that the weights do not converge, the adaptive damping coefficient is determined through a dual-factor adjustment mechanism, and the amplitude of the weight adjustment is constrained accordingly; The dual-factor adjustment mechanism includes: determining the basic damping strength based on the degree of convergence assessed by multi-dimensional comprehensive criteria, and dynamically adjusting the basic damping strength using an adjustment factor that is positively correlated with the rate of change of the weight itself; Obtain redundant information identification results, including: For each pair of sensors, their information complementarity is quantified to construct an information complementarity matrix. The quantization process combines the standard mutual information calculation results with the time-varying spatial correlation coefficient derived from the network topology state matrix. Based on the quantized values ​​in the information complementarity matrix, redundant sensor pairs with information overlap are identified and redundant information recognition results are obtained.

2. The method according to claim 1, characterized in that The step of performing nonlinear mapping on the information entropy change rate matrix includes: Based on the original sensor array data, the data credibility of each sensor is calculated to obtain a data quality assessment matrix; based on this, the sensor characteristic coefficient vector used to characterize the individual response characteristics of each sensor is calculated; Based on the sensor characteristic coefficient vector and the information entropy change rate matrix, for each sensor, in its nonlinear mapping process, the dynamic weight is calculated by fusing the sensor's information entropy change rate, the corresponding evaluation result, and the corresponding characteristic coefficient.

3. The method according to claim 2, characterized in that The nonlinear mapping process is as follows: Determine the relationship between the information entropy change rate θ and the first threshold θ1 and the second threshold θ2; When θ≤θ1, it is determined to be in a stable change range and the quadratic function is used; When θ1<θ≤θ2, it is determined to be in the transition change range and the logarithmic function is used; When θ>θ2, it is determined to be in the saturation change range and exponential decay is used; Wherein θ1<θ2, and θ1 and θ2 are both dynamically set according to the real-time statistical distribution of the values ​​in the information entropy change rate matrix.

4. The method according to claim 2, characterized in that Each sensor characteristic coefficient in the sensor characteristic coefficient vector is calculated by fusing at least two of the following indicators: The first indicator is used to measure the activity of the sensor's historical information, which is derived from the variance of the sensor's historical information entropy sequence; The second indicator, used to evaluate the long-term reliability of the sensor, is derived from the average quality level of the sensor recorded in the data quality assessment matrix.

5. The method according to claim 1, wherein: Analyzing the data characteristics within the multi-scale weight history matrix to evaluate weight convergence is achieved through a multi-dimensional comprehensive criterion; The multi-dimensional comprehensive criterion integrates considerations of at least the following two dimensions: the stability of statistical fluctuations of weights in the first time window, and the convergence of the changing trend of weights in a second time window that is larger than the first time window.

6. The method according to claim 1, wherein: The quantization process includes solving the conditional entropy in its mutual information calculation. The conditional entropy is solved by applying a mean difference correction factor to the standard conditional entropy result. The mean difference correction factor is determined based on the normalized difference in the mean of the measurement data between the sensor pair.

Images