An end-cloud cooperative processing method for farmland monitoring data

Abstract

The application relates to the technical field of distributed network communication and edge computing, and discloses an end-cloud cooperative processing method for farmland monitoring data, which comprises the following steps: a sensing monitoring terminal calculates a time domain residual error vector and a disturbance energy index according to a state prediction vector, sends state heartbeat data packets when the index does not exceed a preset gating threshold, and adjusts a quantization bit width and encodes and sends to a field edge computing gateway when the index exceeds the threshold; the field edge computing gateway constructs a multi-node space topology graph to extract regional space correlation features, and determines a data distribution ratio in combination with a battery state of charge and a channel signal-to-noise ratio; a cloud computing cluster solves a physical change gradient according to feature data, and issues updated parameters and a scheduling strategy to complete full-link closed-loop calibration; the application eliminates source communication redundancy and reduces node power consumption through time-varying elastic sensing, and realizes dynamic optimal configuration of edge computing power and transmission energy consumption.

Description

Technical Field

[0001] This invention relates to the field of distributed network communication and edge computing technology, and more specifically, to a method for edge-cloud collaborative processing of farmland monitoring data. Background Technology

[0002] Large-scale farmland IoT monitoring networks typically deploy sensor monitoring terminals in the fields, powered by batteries or photovoltaic energy storage systems. These terminals periodically collect multi-dimensional physical parameters such as soil moisture, ambient temperature and humidity, and light intensity. The collected data is aggregated to the field edge computing gateway via wireless transmission links, and then forwarded by the field edge computing gateway to the cloud computing cluster for centralized storage and large-scale data analysis. As the number of farmland monitoring nodes continues to increase, this approach can meet the needs of remote collection and monitoring of farmland environmental data.

[0003] Farmland is typically distributed across large, open areas, where wireless communication is easily affected by factors such as terrain undulations, weather changes, and vegetation canopy obstruction. Link quality can fluctuate frequently and even experience deep fading. Furthermore, sensor monitoring terminals and field edge computing gateways mostly rely on limited batteries or energy storage devices, significantly restricting their power availability. Existing systems typically poll and collect data at fixed intervals, reporting all collected data. When the monitored object remains in a stable state for an extended period, a large amount of minimally changing sampled data can still trigger continuous transmission from the wireless RF module, causing unnecessary communication overhead and accelerating the power consumption of the sensor monitoring terminal. During data aggregation and task offloading, existing offloading methods for field edge computing gateways often employ fixed latency and energy consumption tradeoffs. When wireless channel quality deteriorates or the gateway's remaining power is insufficient, these fixed settings fail to reflect the rapid increase in wireless transmission energy consumption in a timely manner, easily leading to continuous high-power wireless retransmissions. In severe cases, this can even cause the field edge computing gateway to over-discharge and fail. Stopping operations and simply increasing wireless transmission power or adding edge computing resources will not only increase system deployment costs but also further burden the energy consumption of field nodes. For example, Chinese invention patent application CN116132438A discloses a cloud-edge collaborative computing offloading method for smart agriculture, which constructs a response time and power consumption model and solves the computational offloading result through particle swarm optimization. However, this solution highly relies on periodic smoothing data at the historical annual and monthly macro-statistical scales for task volume prediction. Such prediction mechanisms are based on the pre-set condition that agricultural operations have periodic regularity. When facing the sudden increase in non-stationary high-frequency sensor data caused by local microclimate changes in the field Internet of Things, such as irrigation disturbances, instantaneous rain infiltration, or end-side hardware anomalies, this existing technology cannot effectively suppress redundant information and adaptively encode waveforms at the source on the end side. It is also difficult to achieve real-time closed-loop control of the entire link parameters according to the dynamic gradient of micro-physical parameters, which can easily lead to uplink transmission congestion and gateway discharge failure under weak network and low power conditions.

[0004] Therefore, given the limited energy storage of sensor nodes and the constantly changing quality of wireless channels, the technical problems to be solved by this invention are: how to reduce the wireless transmission of redundant data at the edge, how to enable the field edge computing gateway to dynamically arrange local computing and cloud offloading based on channel status and remaining power, and how to use the cloud's analysis results of environmental change trends to adjust the parameters of the edge and the end side, thereby balancing the overall network communication overhead and computing load. Summary of the Invention

[0005] To address the problems in the background art, this invention provides a method for edge-cloud collaborative processing of farmland monitoring data, comprising the following steps:

[0006] Step S1: The sensor monitoring terminal acquires the sampled data and calculates the temporal residual vector and disturbance energy index based on the state prediction vector of the previous cycle. When the disturbance energy index does not exceed the preset threshold, a 1-byte state heartbeat data packet is sent to the field edge computing gateway. When the disturbance energy index exceeds the preset threshold, the quantization bit width is adjusted according to the disturbance energy index, the temporal residual vector is adaptively quantized and encoded, and then sent to the field edge computing gateway.

[0007] Step S2: The field edge computing gateway receives data sent by each sensor monitoring terminal, constructs a multi-node spatial topology map and extracts regional spatial feature vectors; adjusts the latency and energy consumption trade-off factor according to the gateway battery state of charge and wireless channel signal-to-noise ratio, solves the objective function to minimize the weighted cost, and determines the data allocation ratio between edge local computing and data offloaded to the cloud computing cluster.

[0008] In step S3, the cloud computing cluster receives the regional spatial feature vectors, calls the spatiotemporal evolution model to calculate the physical change gradient, and generates the smoothing coefficient of the state prediction vector based on the physical change gradient, the adjustment amount of the preset gate threshold, and the uplink scheduling strategy for the field edge computing gateway. This is then broadcast to the field edge computing gateway and the sensor monitoring terminal through the downlink control link to complete the closed-loop calibration of the collaborative parameters.

[0009] Preferably, in step S1, the state prediction vector is obtained by linearly weighting and summing the actual monitoring state vector of the previous period and the state prediction vector of the previous period according to the smoothing coefficient; the disturbance energy index is determined by the ratio of the sum of squares of each dimension of the time domain residual vector to the preset benchmark variance; when the disturbance energy index does not exceed the preset threshold, the radio frequency communication unit of the sensor monitoring terminal remains in sleep mode and only reports the node's network status through single-byte short frames.

[0010] Preferably, in step S1, the data structure of the status heartbeat data packet includes a 7-bit node address identifier and a 1-bit steady-state flag. When the number of times the sensing and monitoring terminal continuously generates status heartbeat data packets reaches the preset heartbeat upper limit threshold, the current sampling data of the multidimensional physical parameters of the farmland is fully sampled in the current sampling period and encapsulated into a baseline status data packet and reported to the field edge computing gateway to reset the status baseline of the terminal side and the edge side.

[0011] Preferably, in step S2, constructing a multi-node spatial topology graph and extracting regional spatially relevant feature vectors includes the following steps: Step S21, constructing a node topology graph based on the spatial geographic coordinates of each sensor monitoring terminal deployed in the same survey area, calculating the Euclidean distance between any two sensor monitoring terminals, calculating the topological connection weights between nodes based on the Gaussian kernel function, and generating a topology graph adjacency matrix; Step S22, calculating the degree matrix and normalized Laplacian matrix based on the topology graph adjacency matrix, inputting the temporal residual quantized feature vectors reported by each sensor monitoring terminal as graph node signals into a two-layer Chebyshev graph convolutional network, and extracting regional spatially relevant feature vectors after eliminating spatial Euclidean distance-related redundancy.

[0012] Preferably, in step S2, determining the data allocation ratio between edge local computing and data offloading / uploading to the cloud computing cluster includes the following steps: Step S23, obtaining the current battery state of charge and the current uplink wireless channel signal-to-noise ratio of the field edge computing gateway, and establishing a joint cost objective function that includes local computing energy consumption, wireless transmission energy consumption, local queue processing latency, and network transmission latency; Step S24, determining the energy consumption penalty weight based on the negative correlation between the current battery state of charge and the preset rated capacity, determining the transmission revenue weight based on the positive correlation between the current uplink wireless channel signal-to-noise ratio, using the subgradient descent method to solve the optimal solution that minimizes the joint cost objective function, and outputting the data allocation ratio indicating the local feature dimensionality reduction task and the remote direct offloading task.

[0013] Preferably, in step S3, the smoothing coefficient of the state prediction vector generated based on the physical change gradient, the adjustment amount of the preset gating threshold, and the uplink scheduling strategy for the field edge computing gateway includes the following steps: Step S31, the cloud computing cluster inputs the regional spatial correlation feature vectors of multiple consecutive scheduling cycles in time into the bidirectional long short-term memory network to calculate the spatial diffusion rate and temporal evolution acceleration of the multidimensional physical parameters of the farmland environment in the two-dimensional grid space, and obtains the physical change gradient; Step S32, when the magnitude of the physical change gradient is less than the preset lower limit gradient threshold, the smoothing coefficient is increased and the adjustment amount of the preset gating threshold is increased to reduce the reporting frequency of the edge; when the magnitude of the physical change gradient is greater than the preset upper limit gradient threshold, the smoothing coefficient is decreased and the adjustment amount of the preset gating threshold is reduced to improve the sensitivity of edge mutation sampling, and the uplink scheduling strategy to shorten the feature upload cycle is issued to the field edge computing gateway.

[0014] Preferably, in step S3, the cloud computing cluster also calculates the dynamic covariance matrix of multidimensional physical parameters based on regional macro-meteorological forecast data and farmland water and heat balance model, reconstructs the segmented non-uniform quantization level interval of the quantizer in the sensor monitoring terminal, generates a dynamic correction codebook parameter matrix and broadcasts it into the register of the sensor monitoring terminal through the downlink control link, so that the sensor monitoring terminal reduces the quantization step in the physical parameter transition interval and increases the quantization step in the smooth interval.

[0015] Preferably, in step S2, when the field edge computing gateway detects that the uplink wireless channel transmission bit error rate between the field and the cloud computing cluster is continuously higher than the preset bit error rate threshold, a local degradation storage mechanism is triggered: Step S25, stop sending high-frequency feature data to the cloud computing cluster and start the non-volatile circular storage buffer inside the gateway; Step S26, perform principal component analysis to reduce the dimensionality of the extracted regional spatial related feature vectors, and write the core principal component features after dimensionality reduction along with the corresponding timestamps into the non-volatile circular storage buffer in sequence; when the uplink wireless channel transmission bit error rate is detected to recover to below the preset bit error rate threshold, the cached data is resent upwards in burst transmission mode according to the timestamp order.

[0016] Preferably, in step S2, after receiving the data sent by each sensor monitoring terminal, the field edge computing gateway also performs local outlier detection on the temporal residual quantization feature vector of each sensor monitoring terminal: Step S27, calculate the Mahalanobis distance between the temporal residual quantization feature vector of the target sensor monitoring terminal and the mean of the feature vectors of its physically adjacent sensor monitoring terminals; Step S28, when the Mahalanobis distance exceeds the preset spatial consistency deviation threshold and the duration reaches the preset number of abnormal cycles, determine that the target sensor monitoring terminal is in a hardware sensor fault state; reset the topology connection weight corresponding to the target sensor monitoring terminal to zero in the multi-node spatial topology graph, add a fault alarm flag bit to the generated feature data packet, and upload it to the cloud computing cluster.

[0017] The embodiments of the present invention have at least the following beneficial effects:

[0018] 1. In the edge-cloud collaborative processing of farmland monitoring data, the sensor monitoring terminal determines whether the current monitoring data needs to be fully reported based on the time-domain residual vector and disturbance energy index. When the environment is stable, only the status heartbeat data packet is sent to reduce wireless transmission caused by repeated sampling data. When sudden changes occur, the quantization bit width is adjusted according to the degree of disturbance and the corresponding data is sent. This can reduce unnecessary radio frequency communication and repeated data transmission in stable conditions, while taking into account the transmission needs of sudden change data, reducing the communication power consumption of the sensor monitoring terminal, and helping to extend the continuous working time of battery-limited nodes.

[0019] 2. Field edge computing gateways reduce duplicate data by utilizing the spatial correlation between adjacent sensor monitoring terminals. They also dynamically schedule local computing and cloud offloading based on the gateway's battery state of charge and the wireless channel signal-to-noise ratio. When the gateway's remaining power is low or the wireless channel quality is poor, unsuitable real-time wireless transmissions can be reduced, and more local processing and delayed batch uploads can be used. This avoids repeated retransmissions under weak network conditions that continuously consume the gateway's energy storage. This reduces the amount of data that needs to be uploaded and allows local computing and cloud offloading to be adjusted according to the gateway's power supply status and communication conditions, improving the field edge computing gateway's continuous operation capability under weak network and energy-limited conditions.

[0020] 3. The cloud computing cluster adjusts the smoothing coefficient of the sensor monitoring terminal, the preset threshold, and the uplink scheduling strategy of the field edge computing gateway based on the physical change gradient of the regional environment. This allows the data reporting and gateway upload rhythm at the edge to adjust with changes in the farmland environment. When the environment is stable, the sensitivity of the edge to minor fluctuations can be reduced, and redundant communication can be decreased. When there are significant changes in the environment, the monitoring sensitivity can be improved and the feature upload cycle can be shortened, reducing the possibility of missing information about sudden changes. As a result, the data transmission volume and quantification accuracy of the entire network can be dynamically adjusted with environmental changes, making more rational use of wireless communication bandwidth and edge and cloud computing resources. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein:

[0022] Figure 1 This is a flowchart of the edge-cloud closed-loop collaborative processing of farmland sensing data according to the present invention.

[0023] Figure 2 This is a diagram showing the edge unloading decision and parameter adaptive control of the present invention. Detailed Implementation

[0024] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0025] A method for edge-cloud collaborative processing of farmland monitoring data includes the following steps:

[0026] Step S1: The sensor monitoring terminal acquires the sampled data and calculates the temporal residual vector and disturbance energy index based on the state prediction vector of the previous cycle. When the disturbance energy index does not exceed the preset threshold, a 1-byte state heartbeat data packet is sent to the field edge computing gateway. When the disturbance energy index exceeds the preset threshold, the quantization bit width is adjusted according to the disturbance energy index, the temporal residual vector is adaptively quantized and encoded, and then sent to the field edge computing gateway.

[0027] Step S2: The field edge computing gateway receives data sent by each sensor monitoring terminal, constructs a multi-node spatial topology map and extracts regional spatial feature vectors; adjusts the latency and energy consumption trade-off factor according to the gateway battery state of charge and wireless channel signal-to-noise ratio, solves the objective function to minimize the weighted cost, and determines the data allocation ratio between edge local computing and data offloaded to the cloud computing cluster.

[0028] In step S3, the cloud computing cluster receives the regional spatial feature vectors, calls the spatiotemporal evolution model to calculate the physical change gradient, and generates the smoothing coefficient of the state prediction vector based on the physical change gradient, the adjustment amount of the preset gate threshold, and the uplink scheduling strategy for the field edge computing gateway. This is then broadcast to the field edge computing gateway and the sensor monitoring terminal through the downlink control link to complete the closed-loop calibration of the collaborative parameters.

[0029] Preferably, in step S1, the state prediction vector is obtained by linearly weighting and summing the actual monitoring state vector of the previous period and the state prediction vector of the previous period according to the smoothing coefficient; the disturbance energy index is determined by the ratio of the sum of squares of each dimension of the time domain residual vector to the preset benchmark variance; when the disturbance energy index does not exceed the preset threshold, the radio frequency communication unit of the sensor monitoring terminal remains in sleep mode and only reports the node's network status through single-byte short frames.

[0030] Preferably, in step S1, the data structure of the status heartbeat data packet includes a 7-bit node address identifier and a 1-bit steady-state flag. When the number of times the sensing and monitoring terminal continuously generates status heartbeat data packets reaches the preset heartbeat upper limit threshold, the current sampling data of the multidimensional physical parameters of the farmland is fully sampled in the current sampling period and encapsulated into a baseline status data packet and reported to the field edge computing gateway to reset the status baseline of the terminal side and the edge side.

[0031] Preferably, in step S2, constructing a multi-node spatial topology graph and extracting regional spatially relevant feature vectors includes the following steps: Step S21, constructing a node topology graph based on the spatial geographic coordinates of each sensor monitoring terminal deployed in the same survey area, calculating the Euclidean distance between any two sensor monitoring terminals, calculating the topological connection weights between nodes based on the Gaussian kernel function, and generating a topology graph adjacency matrix; Step S22, calculating the degree matrix and normalized Laplacian matrix based on the topology graph adjacency matrix, inputting the temporal residual quantized feature vectors reported by each sensor monitoring terminal as graph node signals into a two-layer Chebyshev graph convolutional network, and extracting regional spatially relevant feature vectors after eliminating spatial Euclidean distance-related redundancy.

[0032] Preferably, in step S2, determining the data allocation ratio between edge local computing and data offloading / uploading to the cloud computing cluster includes the following steps: Step S23, obtaining the current battery state of charge and the current uplink wireless channel signal-to-noise ratio of the field edge computing gateway, and establishing a joint cost objective function that includes local computing energy consumption, wireless transmission energy consumption, local queue processing latency, and network transmission latency; Step S24, determining the energy consumption penalty weight based on the negative correlation between the current battery state of charge and the preset rated capacity, determining the transmission revenue weight based on the positive correlation between the current uplink wireless channel signal-to-noise ratio, using the subgradient descent method to solve the optimal solution that minimizes the joint cost objective function, and outputting the data allocation ratio indicating the local feature dimensionality reduction task and the remote direct offloading task.

[0033] Preferably, in step S3, the smoothing coefficient of the state prediction vector generated based on the physical change gradient, the adjustment amount of the preset gating threshold, and the uplink scheduling strategy for the field edge computing gateway includes the following steps: Step S31, the cloud computing cluster inputs the regional spatial correlation feature vectors of multiple consecutive scheduling cycles in time into the bidirectional long short-term memory network to calculate the spatial diffusion rate and temporal evolution acceleration of the multidimensional physical parameters of the farmland environment in the two-dimensional grid space, and obtains the physical change gradient; Step S32, when the magnitude of the physical change gradient is less than the preset lower limit gradient threshold, the smoothing coefficient is increased and the adjustment amount of the preset gating threshold is increased to reduce the reporting frequency of the edge; when the magnitude of the physical change gradient is greater than the preset upper limit gradient threshold, the smoothing coefficient is decreased and the adjustment amount of the preset gating threshold is reduced to improve the sensitivity of edge mutation sampling, and the uplink scheduling strategy to shorten the feature upload cycle is issued to the field edge computing gateway.

[0034] Preferably, in step S3, the cloud computing cluster also calculates the dynamic covariance matrix of multidimensional physical parameters based on regional macro-meteorological forecast data and farmland water and heat balance model, reconstructs the segmented non-uniform quantization level interval of the quantizer in the sensor monitoring terminal, generates a dynamic correction codebook parameter matrix and broadcasts it into the register of the sensor monitoring terminal through the downlink control link, so that the sensor monitoring terminal reduces the quantization step in the physical parameter transition interval and increases the quantization step in the smooth interval.

[0035] Preferably, in step S2, when the field edge computing gateway detects that the uplink wireless channel transmission bit error rate between the field and the cloud computing cluster is continuously higher than the preset bit error rate threshold, a local degradation storage mechanism is triggered: Step S25, stop sending high-frequency feature data to the cloud computing cluster and start the non-volatile circular storage buffer inside the gateway; Step S26, perform principal component analysis to reduce the dimensionality of the extracted regional spatial related feature vectors, and write the core principal component features after dimensionality reduction along with the corresponding timestamps into the non-volatile circular storage buffer in sequence; when the uplink wireless channel transmission bit error rate is detected to recover to below the preset bit error rate threshold, the cached data is resent upwards in burst transmission mode according to the timestamp order.

[0036] Preferably, in step S2, after receiving the data sent by each sensor monitoring terminal, the field edge computing gateway also performs local outlier detection on the temporal residual quantization feature vector of each sensor monitoring terminal: Step S27, calculate the Mahalanobis distance between the temporal residual quantization feature vector of the target sensor monitoring terminal and the mean of the feature vectors of its physically adjacent sensor monitoring terminals; Step S28, when the Mahalanobis distance exceeds the preset spatial consistency deviation threshold and the duration reaches the preset number of abnormal cycles, determine that the target sensor monitoring terminal is in a hardware sensor fault state; reset the topology connection weight corresponding to the target sensor monitoring terminal to zero in the multi-node spatial topology graph, add a fault alarm flag bit to the generated feature data packet, and upload it to the cloud computing cluster.

[0037] Example 1: The current field crop growing season environmental monitoring network includes sensor monitoring terminals deployed within the farmland grid, field edge computing gateways equipped with solar energy storage systems, and remote cloud computing clusters. The sensor monitoring terminals integrate soil moisture sensors, ground temperature sensors, and air humidity sensors, collecting multi-dimensional physical parameters of the farmland environment at a preset sampling period of 60 seconds. A three-dimensional monitoring state vector is constructed through dimensionless normalization processing. The sensor monitoring terminal maintains the actual monitoring state vector of the previous sampling period in its memory. and the state prediction vector of the previous sampling period The state prediction vector for the current sampling period is generated using a first-order autoregressive time-series filtering model deployed on the edge. Calculate the time-domain residual vector between the monitored state vector and the state prediction vector. ; ,in, This is the state prediction vector for the current sampling period. For the end-side first-order autoregressive time-series filtering model, are the smoothing coefficients. This is the actual monitoring state vector from the previous sampling period. This is the state prediction vector for the previous sampling period; ,in, The time-domain residual vector, This is the actual monitoring state vector for the current sampling period.

[0038] The sensing and monitoring terminal uses the time-domain residual vector variance relative to preset benchmark Calculate the disturbance energy index for the current sampling period. The disturbance energy index With preset gate threshold Numerical comparisons were performed when farmland was under steady-state conditions at night or with gradual changes in temperature and humidity, and the following conditions were met. When the sensor monitoring terminal determines that the environment is in a stable state and enters a steady-state maintenance mode, it stops the continuous data encoding and transmission operation of the wireless radio frequency module, assembles a single-byte status heartbeat data packet containing a 7-bit node address identifier and a 1-bit steady-state maintenance flag, and sends it to the field edge computing gateway, putting the end-side radio frequency hardware into a low duty cycle sleep state; when sudden rainfall infiltration or irrigation disturbance causes parameter transitions and meets the requirements... At that time, the sensor monitoring terminal determines that a sudden disturbance has occurred in the environment, based on the disturbance energy index. Calculate adaptive quantization bit width Adaptive quantization bit width For time-domain residual vectors Implement non-uniform scalar quantization coding, complete adaptive quantization coding, encapsulate and generate dynamically compressed monitoring data packets, and upload them to the field edge computing gateway via a wireless radio frequency link; ,in, This is an indicator of disturbance energy. Let L be the L2 norm of the time-domain residual vector. The preset baseline variance for each sensing dimension; ,in, To adapt to the quantization bit width, The preset base quantization bit width, The preset maximum quantization bit width, This is the gain adjustment coefficient. For the interval truncation and rounding operator, the gain adjustment coefficient is set to 1.50 in actual engineering configuration. This value is determined based on the variance distribution dispersion of 5000 discrete residual samples collected in simulated farmland habitat change tests. When the gain adjustment coefficient is less than 1.00, the quantization bit width is not sensitive enough to the increase of disturbance energy, resulting in a signal reconstruction error higher than 1.50%. When the gain adjustment coefficient is greater than 2.50, the quantization bit width quickly reaches the upper limit, increasing unnecessary wireless transmission overhead. Setting it to 1.50 can ensure the optimal match between transmission bandwidth and sensing accuracy under the premise that the root mean square error of signal reconstruction is less than 1.00%.

[0039] The field edge computing gateway receives data packets uploaded by multiple sensor monitoring terminals within its communication area, extracts a node set containing the spatial geographic coordinates of each terminal, constructs a multi-node spatial topology map, and determines the location of each sensor monitoring terminal based on their physical distance. Spatial association feature length The topological connection weights between nodes are calculated based on the Gaussian kernel function. The topology graph adjacency matrix is ​​composed of the topology connection weights, and the received time-domain residual vectors are weighted and fused using this matrix to eliminate spatial monitoring redundancy caused by the continuous distribution of farmland physical habitats between adjacent sensing nodes. ,in, For nodes relative nodes Topological connection weights, For nodes With nodes The physical distance between them The preset spatial association feature length, For nodes In processing graph node signal inputs, if the gateway receives a single-byte status heartbeat data packet from a sensor monitoring terminal, it indicates that the node is in a physically stable state. The gateway extracts the corresponding feature input cache slot in memory and uniformly writes the corresponding input residual component into a three-dimensional all-zero numerical array. That is, the temporal residual input components of soil moisture content, ground temperature, and air humidity are deterministically set to 0.00. This objectively characterizes the spatial location as a stable baseline state without physical mutations and eliminates the cascading drift of historical truncation errors. If the gateway receives a dynamically compressed monitoring data packet containing an effective quantized payload, the parsed actual quantized temporal residual component is filled into the corresponding node vector slot. This constructs a 32-row, 3-column homogeneous full node feature matrix consisting of 32 nodes. The filled feature matrix is ​​then input into a two-layer Chebyshev graph convolutional network. By performing Chebyshev polynomial multiplication with the Laplacian matrix, the spatially relevant feature vectors of the region after eliminating spatial redundancy are extracted, ensuring that the graph node input dimension corresponding to the dormant node remains complete.

[0040] Field edge computing gateway obtains the local processor's operating frequency Effective capacitance coefficient of the chip The amount of data to be processed Current uplink wireless channel transmission rate from edge to cloud Edge gateway RF transmit power and cloud processing latency Calculate edge local execution latency Local computing power consumption Cloud offloading and transmission latency With transmission energy consumption and using the joint utility cost index Performing binary extreme value adjudication of edge local computing and cloud offloading, the field edge computing gateway reads the remaining battery state of charge percentage in real time. and uplink wireless channel signal-to-noise ratio When rainy weather limits energy storage and multipath obstruction in the field induces channel deep fading, a reference signal-to-noise ratio is introduced. Trade-off between latency and energy consumption By implementing dynamic decay correction, the corrected tradeoff factor is obtained. and will Substitute it into the joint utility cost index as a current trade-off factor ;when When the local computational overall utility is determined to be superior, the extraction of relevant regional spatial feature vectors is completed locally at the edge gateway. When determining that the overall efficiency of cloud offloading and transmission is superior, the relevant feature vectors of the regional space are packaged and offloaded to the cloud computing cluster through the uplink wireless channel. ,in, For edge local computation latency, This represents the total number of computation cycles required for the local feature extraction task. The operating frequency of the edge gateway processor, Energy consumption for edge local computing The effective capacitance coefficient of the chip is defined as follows: Specifically, the total number of calculation cycles is determined by the total number of multiply-accumulate instructions required for the gateway to run the Chebyshev graph convolutional network and principal component analysis algorithm locally.

[0041] In actual operation, based on the amount of data to be processed reported by the 32 sensor nodes in the current communication area, and following the engineering calibration rule that a single node multiply-accumulate operation consumes 4 clock cycles, the gateway calculates that the total number of clock cycles required to extract the spatially relevant feature vectors of the region is 48,000,000. This provides accurate data support for the precise quantification of local computing latency and energy consumption. ,in, To reduce transmission latency by offloading data to the cloud, The amount of data to be processed. For uplink wireless channel transmission rate, To handle latency in the cloud, Offloading transmission energy consumption to the cloud For edge gateway RF transmit power, ,in, As a joint utility cost indicator, For the trade-off between latency and energy consumption; ,in, The revised trade-off factor. This represents the remaining state of charge (SBC) of the battery, corresponding to the ratio of the current SBC to the preset rated capacity. The measured signal-to-noise ratio of the wireless channel. Using the baseline signal-to-noise ratio, when actually solving for the continuous data allocation ratio, the gateway decomposes the spatial feature data to be processed into two continuous variables: local feature dimensionality reduction task volume and remote direct offloading task volume. The local feature dimensionality reduction task volume corresponds to local computing energy consumption and local queue processing latency, while the remote direct offloading task volume corresponds to wireless transmission energy consumption and network transmission latency. The local queue processing latency is calculated based on the queuing and execution time of local tasks under the current processor load, and the network transmission latency is determined based on the remote direct offloading data volume and the current uplink wireless channel transmission rate. Based on this, the gateway establishes a joint cost objective function including local computing energy consumption, wireless transmission energy consumption, local queue processing latency, and network transmission latency. The gateway introduces a preset baseline maximum latency constant of 1.000 seconds and a baseline maximum energy consumption constant of 5.000 joules to perform dimensionless normalized division on the latency difference and energy consumption difference, respectively, to resolve the physical dimension breakage conflict between time seconds and energy joules. In the optimization objective function, the data allocation ratio is defined as an indicator of the local computing task... The continuous optimization variable for the proportion of data volume is limited to the range of 0.000 to 1.000. The proportion of remote unloading tasks is set to 1.000 minus the local allocation ratio. The local queue processing latency is determined by multiplying the total amount of data to be allocated by the local allocation ratio and then dividing by the product of the processor's main frequency and the current queue idle rate. The network transmission latency is determined by multiplying the total amount of data to be allocated by the remote unloading ratio and then dividing by the uplink wireless channel transmission rate and adding the cloud fixed computing latency of 0.050 seconds. The local computing energy consumption is calculated by multiplying the processor's effective capacitance coefficient by the cube of the main frequency and the local execution time. The remote wireless transmission energy consumption is calculated by multiplying the RF transmission power by the upload transmission time. The gateway multiplies the normalized total latency and total energy consumption by the transmission revenue weight and the energy consumption penalty weight, respectively, and then performs a weighted sum to construct a dimensionless joint cost objective function. The energy consumption penalty weight is determined by the negative correlation between the current battery state of charge and the preset rated capacity, and the transmission revenue weight is determined by the positive correlation between the current uplink wireless channel signal-to-noise ratio. The data allocation ratio is used as the variable to be optimized.

[0042] The gateway calculates the subgradient direction of the joint cost objective function with respect to the data allocation ratio in successive iterations with an iteration step size of 0.01. When the derived subgradient value is positive, the unloading ratio is decreased by the step size, and when the subgradient value is negative, the unloading ratio is increased. After a maximum of 20 subgradient descent iterations, the data allocation ratio is converged. Under extreme conditions, when the amount of local feature dimensionality reduction task reaches its maximum value and the joint utility cost index is not greater than 0, the system degenerates into a binary decision mode of full local computation, realizing the unification of continuous data allocation ratio solution and binary extreme value adjudication.

[0043] The cloud computing cluster aggregates the regional spatially relevant feature vectors uploaded by each field edge computing gateway and combines them with the cross-gateway historical spatiotemporal sequence matrix to calculate the spatiotemporal evolution model. During the calculation, the regional spatially relevant feature vectors of multiple consecutive scheduling cycles are arranged according to the two-dimensional grid space and scheduling cycle and input into a bidirectional long short-term memory network. The forward and backward time-series calculation results are used to obtain the physical quantity differential changes on the horizontal and vertical spatial axes of the two-dimensional grid, and the spatial diffusion rate is calculated accordingly. Then, the spatial diffusion rate of adjacent scheduling cycles is temporally differentially analyzed to obtain the temporal evolution acceleration, and the physical change gradient is obtained by combining the results. The magnitude of the gradient is determined by... This indicates that the preset lower gradient threshold and the preset upper gradient threshold are used to divide the gradual stable interval and the transitional abrupt interval, respectively.

[0044] When the magnitude of the physical change gradient is less than a preset lower limit gradient threshold, the smoothing coefficient of the state prediction vector is increased. It also increases the adjustment amount of the preset gate threshold, thereby raising the preset gate threshold. Reduce the frequency of edge reporting; when the magnitude of the physical change gradient exceeds the preset upper limit gradient threshold, reduce the smoothing coefficient. And reduce the adjustment amount of the preset gate threshold, thereby lowering the preset gate threshold. To improve the sensitivity of the edge computing terminal to sudden hydrothermal fluctuations, an uplink scheduling strategy to shorten the feature upload cycle is issued to the field edge computing gateway. The cloud computing cluster also combines regional macro-meteorological forecast data with the multi-scale hydrothermal evolution trend given by the large-scale farmland hydrothermal balance physical evolution model. It calculates the dynamic covariance matrix reflecting the nonlinear changes in multi-dimensional physical parameters, reconstructs the segmented non-uniform quantization level interval of the quantizer in the sensing terminal, and generates a dynamic correction codebook parameter matrix. This allows the sensing terminal to reduce the quantization step within the physical parameter transition interval and increase the quantization step within the smooth interval. Subsequently, the updated smoothing coefficients are... Preset gate threshold The dynamic correction codebook parameter matrix and the uplink scheduling strategy for field edge computing gateways are encapsulated into downlink control signaling, which is broadcast to the registers of each field edge computing gateway and sensor monitoring terminal through the downlink control link to complete the closed-loop calibration of collaborative parameters. This enables the network-wide data transmission rate, edge offloading allocation, and local quantization accuracy to be adaptively adjusted according to the evolution of the large-scale physical environment.

[0045] Example 2: The current farmland multi-source IoT sensing test system is deployed in a gridded test area of ​​500m×500m. It includes 32 evenly distributed sensor monitoring terminals, a field edge computing gateway equipped with a 12V20Ah lithium iron phosphate battery and a 50W solar panel, and a remote cloud computing cluster. The sensor monitoring terminals are equipped with a microcontroller with an operating voltage of 3.3V, and integrate a soil volumetric moisture content sensor with a measurement range of 0 to 100% and a measurement accuracy of 0.1%, a ground temperature sensor with a measurement range of -20 to 60℃ and a measurement accuracy of 0.1℃, and an air humidity sensor with a measurement range of 0 to 100%RH and a measurement accuracy of 0.1%RH. Each sensor monitoring terminal establishes a single-hop star topology communication link with the field edge computing gateway through a wireless radio frequency communication module with an operating frequency of 433MHz. The radio frequency transmission power is set to 10dBm to 20dBm. The field edge computing gateway is equipped with a main frequency... It is a 1.2GHz quad-core processor with a chip effective capacitance coefficient. for W·s³, and maintains uplink and downlink connections with a remote cloud computing cluster via a wireless public network. To simulate non-ideal wireless channels and habitat change interference in a field agricultural IoT environment, additive white Gaussian noise with dynamic fluctuations in signal attenuation and signal-to-noise ratio between 5.0dB and 25.0dB is injected into the wireless communication radio frequency channel. Simultaneously, step transitions of multidimensional physical parameters are simulated through artificial sprinkler irrigation and rain shower infiltration, constructing a test input data stream with dynamic noise perturbation. The preset sampling period is set to 60s. The time constant of the time domain evolution of soil hydrothermal physical quantities is in the range of tens of minutes to several hours. The 60s sampling interval can satisfy the sampling theorem for capturing the step response waveform, while reducing the continuous working duty cycle of the end processor and the number of radio frequency transmissions. The preset baseline variance is... Set to 0.01, this parameter is calibrated by calculating the steady-state measurement noise variance of 1000 sets of normalized sensor state sequences collected under static dry soil conditions. It is used to eliminate the false alarm effect of sensor hardware background white noise on time-domain residual determination; preset threshold. Set to 0.20, its numerical calibration rule is based on the upper boundary of the 95% confidence interval of the static environmental disturbance energy accumulation distribution function, ensuring that the data emissivity under steady state is controlled below 5%, and triggering event-driven compression when hydrothermal parameters undergo sudden changes; spatial correlation feature length The distance is set to 15.0m. This parameter is determined based on the autocorrelation fitting distance of the Gaussian covariance semivariogram of the transverse diffusion of soil water and heat in the experimental field. It is used to measure the spatial physical correlation between adjacent sensing and monitoring terminals within a distance of 10.0m to 30.0m.

[0046] To verify the operational results of edge-side communication energy consumption, edge-side collaborative offloading, and cloud-based closed-loop adjustment under different working conditions, this embodiment sets up five test groups. The test group performs edge-side state prediction vector calculation, disturbance energy index gating, and adaptive quantization coding. A multi-node spatial topology map is constructed at the field edge computing gateway, and regional spatially relevant feature vectors are extracted. The data allocation ratio between local computation and data offloading to the cloud computing cluster is determined based on the gateway's battery state of charge and the wireless channel signal-to-noise ratio. The cloud computing cluster then completes closed-loop calibration of the collaborative parameters based on the physical change gradient. Control group 1 cancels edge-side state prediction and disturbance gating, and uses the aforementioned fixed sampling period and 16-bit original sampling width to report to the field edge computing gateway in full polling. Control group 2 retains edge-side gating processing, cancels the extraction of regional spatially relevant feature vectors from the edge-side multi-node spatial topology map, and uses a fixed weight factor in the task offloading decision. A value of 0.5 determines whether local computation or remote offloading occurs; control group 3 retains both endpoint and edge processing logic, but cuts off the downlink closed-loop feedback of cloud physical change gradients, maintaining the endpoint smoothness coefficient. The threshold was set to 0.75, the preset gate threshold was set to the initial value, and the initial quantization codebook parameters remained constant. The out-of-range control group used two deviation settings: low gate threshold and high gate threshold. The test process was divided into four consecutive typical physical condition test phases.

[0047] During the stable, steady-state operation at night in Phase 1, with no light and a gradual change in ambient temperature and humidity, the channel signal-to-noise ratio of the test group, control group 1, and control group 2 was all 22.5 dB, the battery state of charge was all 88.0%, and the three-dimensional monitoring state vector collected by the sensor monitoring terminal remained stable. Based on the nearby fluctuations, the test group of this invention obtained the time-domain residual vector. The disturbance energy index meets In this invention, both the test group and control group 2 have an adaptive quantization bit width of 0. The sensing and monitoring terminal enters a steady-state sleep state and only sends a 1-byte status heartbeat data packet. The average communication power consumption of a single node on both sides is 0.82mW, the periodic data transmission volume is 0.008kbit, and the root mean square error of signal reconstruction is 0.15%. The overall end-to-end processing latency of the test group is 12.4ms, while that of control group 2 is 12.5ms. Control group 1 still uses 16-bit fixed full upload, with an average communication power consumption of 37.45mW per node on the side, a periodic data transmission volume of 48.000kbit, an overall end-to-end processing latency of 118.6ms, and a root mean square error of signal reconstruction of 0.12%. Based on the communication power consumption of the two groups, the communication energy consumption of the test group of this invention is reduced by 97.81%, but its root mean square error of signal reconstruction is slightly higher than that of control group 1.

[0048] During the sudden shower infiltration scenario in Phase 2, the channel signal-to-noise ratio was 18.2 dB, the battery state of charge was 75.0%, and the three-dimensional monitoring state vector collected by the sensor monitoring terminal became... The state prediction vector is This generates the time-domain residual vector. The disturbance energy index was calculated. The corresponding value in the test record is 0.628. Based on this, the test group of this invention adjusted the adaptive quantization bit width. After adjusting to 5 bits and completing non-uniform scalar quantization encoding, a 24,000-bit compressed data packet was uploaded, corresponding to a periodic data transmission volume of 24.000 kbit. The field edge computing gateway performed edge-cloud collaborative offloading, with an average communication power consumption of 14.36 mW per node on the edge side, a comprehensive end-to-end processing latency of 104.0 ms, and a root mean square error of 0.82% for signal reconstruction. Control group 1 used fixed full upload, with an average communication power consumption of 38.12 mW per node on the edge side and a periodic data transmission volume of 48.000 kbit. The overall end-to-end processing latency was 215.8ms, and the root mean square error of signal reconstruction was 0.78%. Control group 3 used a 4-bit static codebook for uploading, with an average communication power consumption of 11.85mW per node on the end side, a periodic data transmission volume of 19.200kbit, an overall end-to-end processing latency of 96.2ms, and a root mean square error of signal reconstruction of 3.46%. The reconstruction error of the test group of this invention, which uses a dynamically corrected codebook parameter matrix, was lower than that of control group 3, which used a static codebook, but slightly higher than that of control group 1, which used a fixed full upload.

[0049] In the weak network and low power conditions of Phase 3, the channel signal-to-noise ratio dropped to 8.0 dB, the battery state of charge of the field edge computing gateway dropped to 18.0%, and the disturbance energy index was 0.512. Both the test group and control group 2 of this invention had an adaptive quantization bit width of 5 bits. The field edge computing gateway used a reference signal-to-noise ratio... The energy consumption penalty weight is determined based on a negative correlation between the current battery state of charge and the preset rated capacity, and the transmission gain weight is determined based on a positive correlation between the current uplink wireless channel signal-to-noise ratio and the current uplink wireless channel signal-to-noise ratio. Under this operating condition, the corrected trade-off factor is obtained according to the existing correction relationship. The value is 0.0496, the energy consumption penalty weight is 0.9504, and the joint utility cost index meets the requirements. The gateway thus performs local calculations and suppresses remote offloading. At this time, the average communication power consumption of a single node on the end side of the test group of this invention is 13.88mW, the periodic data transmission volume is 0.000kbit, the overall end-to-end processing delay is 100.0ms, and the root mean square error of signal reconstruction is 0.94%. The control group 2 continues to use a fixed weight factor to determine remote offloading. Its average communication power consumption of a single node on the end side is 13.90mW, the periodic data transmission volume is 24.000kbit, the overall end-to-end processing delay is 486.5ms, and the root mean square error of signal reconstruction is 4.82%. Remote offloading under weak network conditions is accompanied by channel congestion and retransmission. The test group of this invention uses the gateway to locally process the regional spatial correlation feature vector, which reduces the uplink wireless transmission in this stage.

[0050] During the cloud-based closed-loop optimization process in Phase 4, the channel signal-to-noise ratio was 19.0 dB, the battery state of charge was 62.0%, and the disturbance energy index was 0.185. The cloud computing cluster aggregated the regional spatial correlation feature vectors uploaded by each field edge computing gateway and calculated the magnitude of the physical change gradient. The value is set to 0.85, and based on this, the smoothing coefficient of the state prediction vector is increased, and the adjustment amount of the preset gate threshold is decreased, thus adjusting the smoothing coefficient. Updated to 0.80, preset gate threshold will be used. The value was adjusted to 0.15, and the dynamic correction codebook parameter matrix was reconstructed. The updated parameters and uplink scheduling strategy were broadcast to the end-side and edge-side registers via downlink control signaling. After the disturbance energy index exceeded the updated preset gating threshold, the adaptive quantization bit width was 5 bits. The test group of this invention entered the adaptive collaborative processing state. The average communication power consumption of a single node on the end side was 9.65mW, the periodic data transmission volume was 12.000kbit, the overall end-to-end processing latency was 88.5ms, and the root mean square error of signal reconstruction was 0.56%. Compared with stage 2, the total network data transmission volume decreased by 50.000% in this stage.

[0051] In addition, out-of-range tests were conducted on the preset gating threshold. In control group 1, the preset gating threshold was set to 0.05. Under conditions of a channel signal-to-noise ratio of 20.0 dB and a battery state of charge of 80.0%, the disturbance energy index was 0.062, exceeding the preset gating threshold. The adaptive quantization bit width was 4 bits. Sensor inherent thermal noise and minor environmental fluctuations caused false triggering of data uploads, with a false triggering transmission rate of 80.000%. The average communication power consumption per node on the end side was 18.60 mW, the periodic data transmission volume was 38.400 kbit, the overall end-to-end processing latency was 165.2 ms, and the root mean square error of signal reconstruction was 0.52%. In control group 2, the preset gating threshold was set to 0.05. With the gate threshold set to 0.80, under the conditions of a channel signal-to-noise ratio of 18.5dB and a battery state of charge of 74.0%, the disturbance energy index is 0.628, which does not exceed the preset gate threshold. The adaptive quantization bit width is 0, and the sensor monitoring terminal does not trigger data upload and remains in sleep mode in response to sudden changes. Its average communication power consumption per node on the end side is 0.85mW, the periodic data transmission volume is 0.008kbit, the overall end-to-end processing latency is 12.6ms, and the root mean square error of signal reconstruction increases to 6.85%. Based on two sets of out-of-range tests, the preset gate threshold in the range of 0.10 to 0.50 is used to balance transmission suppression under stable conditions and data reporting under sudden changes.

[0052] During the aforementioned testing process, the sensor monitoring terminal controlled the state heartbeat data packets and adaptive quantization encoding according to the time-domain residual vector and disturbance energy index. The field edge computing gateway extracted regional spatial feature vectors based on the multi-node spatial topology map and adjusted the data allocation ratio between local computing and data offloading to the cloud computing cluster based on the wireless channel signal-to-noise ratio and battery charge status. The cloud computing cluster updated the smoothing coefficient, preset gating threshold, and uplink scheduling strategy according to the physical change gradient to complete the closed-loop calibration of the collaborative parameters. In the above tests, the source data communication volume changed with the stable state and the sudden state. Under the weak network and low power conditions, the edge processing path switched to local computing, and the cloud parameter update back-end quantization and reporting parameters were adjusted accordingly. This reduced the source data communication redundancy, adjusted the edge computing power and transmission energy consumption configuration, and maintained the reconstruction accuracy of farmland monitoring data.

[0053] Example 3: Current multi-source meteorological forecast data indicates that a strong convective shower is imminent in the survey area. Soil moisture and surface temperature in farmland are transitioning from a gradual evolution to a non-linear step change. The cloud computing cluster aggregates the regional spatial correlation feature vectors uploaded by each field edge computing gateway over multiple consecutive scheduling cycles. Combined with the cross-gateway historical spatiotemporal sequence matrix, a bidirectional long short-term memory (LSTM) network is used as the spatiotemporal evolution model for calculation. This LSM network employs a two-layer recurrent neural network with a hidden layer dimension of 128. The input receives a 50x50 grid time-series matrix composed of regional spatial correlation feature vectors with three-dimensional coordinate attributes reported by each field edge computing gateway. The cloud processor calculates the algebraic difference between the feature values ​​of adjacent grid cells along the two-dimensional horizontal and vertical coordinates of the grid matrix to obtain a spatial difference sequence. Simultaneously, it calculates the difference between the feature values ​​of the corresponding grid cells in the current scheduling cycle and the previous scheduling cycle to obtain the time. The spatial and temporal difference sequences are flattened and concatenated before being input into a bidirectional long short-term memory network. The network extracts spatiotemporal relevant features through two layers of recurrent neural units and outputs a two-dimensional evolutionary state vector via a fully connected layer. The sum of squares of the two components of this two-dimensional evolutionary state vector is calculated and its square root is taken to obtain a dimensionless scalar indicative of the severity of environmental dynamic changes, serving as the magnitude of the physical change gradient. Its value ranges from 0.00 to 2.00. The cloud computing cluster performs branching processing based on the magnitude of the physical change gradient: when the magnitude is less than a preset lower gradient threshold, the smoothing coefficient is increased and the adjustment amount of the preset gating threshold is increased to reduce the reporting frequency at the edge; when the magnitude is greater than a preset upper gradient threshold, the smoothing coefficient is decreased and the adjustment amount of the preset gating threshold is reduced. Simultaneously, an uplink scheduling strategy to shorten the feature upload cycle is issued to the field edge computing gateway. In this embodiment, under the condition of severe convective showers, the magnitude of the physical change gradient... Once the smoothing factor reaches 0.85 and enters the next branch, the cloud computing cluster will reduce the smoothing factor to... And adjust the preset threshold according to the preset threshold adjustment amount. Downgraded to .

[0054] The cloud computing cluster also calculates the dynamic covariance matrix of multidimensional physical parameters based on regional macro-meteorological forecast data and farmland water and heat balance models. In specific calculations, the cumulative precipitation and surface heat radiation flux for the next 3 hours from the regional macro-meteorological forecast data are analyzed and used as boundary conditions input into a farmland water-heat balance model based on the porous media moisture transport equation. The cloud computing cluster performs piecewise linear interpolation of the predicted cumulative precipitation and surface heat radiation flux values ​​for the next 3 hours according to a 60-second discrete time step, solving for a precipitation infiltration intensity sequence and a surface heat flux gradient sequence at 180 discrete time nodes. Combined with the initial soil moisture content and ground temperature data of the current grid nodes, the slope of soil moisture content change, the slope of ground temperature change, and the slope of air humidity change are calculated for each future 60-second sampling step. The variance and cross-covariance of the three-dimensional slope sequence within the time-series sliding window are extracted, constructing a 3-row by 3-column dynamic covariance matrix. Based on the variance components on the diagonal of the dynamic covariance matrix, the values ​​are normalized from -1.000 to +1.000. Discretized time-domain residual probability distribution histograms are generated within the dynamic interval of the uniform residual. Using minimizing the mean square error of reconstruction distortion as the optimization criterion, 32 uniformly quantized intervals are initialized. Through alternating iterations, the decision threshold of each interval is updated to the arithmetic mean of adjacent reconstruction levels, and the reconstruction level of each interval is updated to the conditional centroid value of the probability density within that interval. After 15 fixed iterations and convergence, a structured lookup table containing 32 decision threshold levels and 32 reconstruction level scalar values ​​is output as the dynamic correction codebook parameter matrix. Subsequently, principal component correlation analysis is performed on the predicted distribution to obtain a dynamic covariance matrix reflecting the changes in the three-dimensional parameters of soil moisture content, ground temperature, and air humidity under nonlinear transition states. Based on the time-domain residual probability distribution corresponding to this dynamic covariance matrix, the segmented non-uniform quantization level intervals of the quantizer in the sensing terminal are reconstructed: in the initial stage of sudden rainfall infiltration, the time-domain residual vector... The algorithm exhibits high-frequency clustering, reducing the quantization interval when corresponding to physical parameter transition intervals; increasing the quantization interval within smooth physical parameter intervals; and widening the quantization interval according to the corresponding probability distribution for high-amplitude asymptotic intervals where the residual occurrence probability is low. The cloud computing cluster employs the Lloyd Max non-uniform quantization iterative algorithm, aiming to minimize the mean square error of reconstruction distortion. It divides the residual probability distribution into intervals, sequentially calculates the decision threshold level and reconstruction level scalar value for each quantization interval, and generates a dynamic correction codebook parameter matrix containing 32 sets of discrete level interval values. : ,in, For dynamically correcting the codebook parameter matrix, This represents the total number of segmented quantization intervals. For the first A non-uniform quantization level interval. For continuous time-domain residual components, For the first The reconstructed level scalar value for each interval. For dynamic covariance matrix The obtained time-domain residual probability density function.

[0055] The cloud computing cluster will update the smoothing coefficient. Preset gate threshold Dynamically corrected codebook parameter matrix The uplink scheduling strategy using a dedicated uplink time slot is encapsulated into downlink control signaling data frames, which are broadcast to each field edge computing gateway and sensor monitoring terminal via the wireless downlink. Regional macro-meteorological forecast data and farmland water and heat balance models are used to predict meteorological change trends in the next 15 to 30 minutes. The cloud computing cluster writes the updated parameters into the end-side control register 10 minutes in advance via the downlink control link. When a local sudden shower occurs, the sensor monitoring terminal directly calls the preset gating threshold and dynamic correction codebook parameter matrix that have been written into the register. Within a single 60-second sampling period, the local microcontroller triggers millisecond-level adaptive quantization and data reporting, no longer relying on the cloud control data that arrives again at that moment.

[0056] After receiving the downlink control signaling, the microprocessor of the sensing and monitoring terminal parses the payload and updates the smoothing coefficient. With preset gate threshold Write to the timing filter control register and the dynamic correction codebook parameter matrix In-situ writing to the end-side quantization encoding register replaces the initial static uniform quantization codebook. When heavy rainfall causes a step jump in farmland soil moisture, the three-dimensional monitoring state vector collected by the sensor monitoring terminal jumps to... The time-domain residual vector is obtained. The disturbance energy index meets Adaptive quantization bit width It is 5 bits; then the dynamic correction codebook parameter matrix in the register is called. Perform non-uniform scalar quantization encoding and upload a dynamically compressed monitoring data packet with a data length of 24,000 bits.

[0057] In the edge hardware, the microcontroller allocates a continuous 128-byte memory area with the starting address 0x4000 as a quantization encoding register. After receiving the control signal, the microprocessor writes 32 sets of non-uniform quantization decision thresholds and reconstructed level values ​​into the register sequentially through the direct memory access channel. When the hardware quantizer quantizes the time-domain residual vector, it uses the decision thresholds stored in the register as the comparator input reference and outputs the corresponding binary code through successive approximation logic, so that the quantization level step is updated according to the dynamically corrected codebook parameter matrix. After the field edge computing gateway extracts the regional spatially relevant feature vectors and uploads them to the cloud computing cluster, the root mean square error of signal reconstruction is reduced from 3.46% when using a static uniform codebook and stabilizes at 0.56%, while the data transmission volume remains unchanged.

[0058] Example 4: This example combines Figures 1 to 2 This paper describes a method for edge-cloud collaborative processing of farmland monitoring data, such as... Figure 1 As shown, the sensing and monitoring terminal acquires the sampled data, calculates the time-domain residual vector, and judges the disturbance energy index by comparing it with a preset gate threshold. When the threshold is not exceeded, a heartbeat data packet is sent; when the threshold is exceeded, adaptive quantization encoding is used for transmission. The field edge computing gateway extracts spatially relevant features to construct a multi-node topology map and determines the data allocation ratio, power consumption, and signal-to-noise ratio. The cloud computing cluster calculates the physical change gradient, calls the spatiotemporal evolution model, and then performs closed-loop calibration of the broadcast generation parameters and strategies. The dashed arrow from this box returns to the box that judges the disturbance energy index by comparing it with the preset gate threshold to achieve closed-loop operation.

[0059] like Figure 2 As shown, after acquiring the sampled data, the sensing and monitoring terminal calculates the time-domain residual vector and the disturbance energy index, and then compares them with the preset gating threshold. When the comparison result is within the limit, a status heartbeat data packet is sent; when the comparison result exceeds the limit, adaptive quantization encoding is performed. The field edge computing gateway extracts the regional spatial related feature vector and detects the uplink wireless channel transmission bit error rate. When the detection result is consistently higher than the preset bit error rate threshold, a local degradation storage mechanism is triggered. When the detection result is lower than the preset bit error rate threshold, a binary extreme value decision is executed, followed by joint utility cost index evaluation. When it is determined that the local computing comprehensive utility is superior, edge local computing is performed. When it is determined that the cloud offloading transmission comprehensive utility is superior, the upload is offloaded to the cloud computing cluster and the physical change gradient is calculated. Then, the magnitude of the physical change gradient is determined. When the magnitude determination result is less than the preset lower limit gradient threshold, the smoothing coefficient is increased and the adjustment amount of the preset gating threshold is increased. When the magnitude determination result is greater than the preset upper limit gradient threshold, the smoothing coefficient is decreased and the adjustment amount of the preset gating threshold is decreased.

[0060] Example 5: Before deploying the current field crop growing season environmental monitoring network, an initial baseline for edge sensing is established. A sensing monitoring terminal integrating soil moisture, ground temperature, and air humidity sensors is placed in a controlled constant-temperature testing environment. 1000 sets of static sensing physical parameters are collected at a preset sampling period of 60 seconds and dimensionless normalization is performed. The steady-state measurement variance of the multidimensional sensing physical parameter sequence is calculated as the preset baseline variance. The static environmental mean vector is used as the initial actual monitoring state vector. With the initial state prediction vector The non-volatile memory register of the sensor monitoring terminal is pre-configured. In the first sampling period after deployment, the first-order autoregressive time-series filtering model is directly called to generate the state prediction vector, so as to reduce the impact of the static background noise of the sensor hardware on the calculation of the initial time-domain residual vector and the disturbance energy index. ,in, This is the state prediction vector for the first sampling period. For smoothing coefficients, This is the initial actual monitoring state vector. As the initial state prediction vector, after the initial state is preset, the sensing and monitoring terminal continuously acquires the three-dimensional monitoring state vector at a sampling interval of 60 seconds. The time-domain residual vector is calculated according to the following formula. and disturbance energy index : ,when When the sensor monitoring terminal enters steady-state mode, it only sends a 1-byte status heartbeat data packet to the field edge computing gateway; when At that time, the adaptive quantization bit width is calculated according to the following formula. The time-domain residual vector is then subjected to non-uniform scalar quantization encoding before being uploaded. After the field edge computing gateway completes its on-site hardware deployment, it establishes a multi-node spatial topology map and a wireless channel status benchmark. The gateway reads the spatial geographic coordinates of 32 sensor monitoring terminals within its communication area and calculates the spatial physical distance between any two sensor monitoring terminals. In regions with uniform soil texture, the length of spatial correlation features is determined based on spatial correlation fitting. m, calculates the topology connection weights between sensing and monitoring terminals based on the Gaussian kernel function. This forms a topology graph adjacency matrix and loads it into the edge memory topology table for weighted fusion of multi-node temporal residual quantization feature vectors. ,in, For nodes relative nodes Topological connection weights, For nodes With nodes The physical distance between them For spatial correlation feature length, For nodes The local neighbor node set is used. The field edge computing gateway constructs a degree matrix according to the degree of each node, performs a normalized Laplace transform on the adjacency matrix of the topological graph, and uses the temporal residual quantized feature vector of each sensing and monitoring terminal as the graph node signal input to a two-layer Chebyshev graph convolutional network to extract the regional spatial correlation feature vector after eliminating the spatial redundancy of adjacent nodes.

[0061] The field edge computing gateway simultaneously collects wireless channel quality metrics between itself and the cloud computing cluster, continuously calculating the average signal-to-noise ratio (SNR) over 100 communication time slots, and using this as the baseline SNR. dB, and the reference signal-to-noise ratio Initial delay and energy consumption trade-off factor Write the task to unload the arbitration module; in the event of subsequent energy storage constraints or channel fading, the remaining state of charge percentage of the battery will be considered. With real-time signal-to-noise ratio Dynamically calculated corrected trade-off factors : Based on this, the gateway uses the revised joint utility cost index. An initial decision is made regarding the direction of edge local computing and cloud offloading, and the tasks to be processed are divided into local feature reduction tasks and remote direct offloading tasks. The energy consumption penalty weight is determined by combining the ratio of the current battery state of charge to the preset rated capacity, and the transmission benefit weight is determined by combining the real-time signal-to-noise ratio. A joint cost objective function is established, which includes local computing energy consumption, wireless transmission energy consumption, local queue processing latency, and network transmission latency. The data allocation ratio is used as the variable to be optimized. The minimum value of the joint cost objective function is obtained through subgradient iteration, and the data allocation ratio between the local feature reduction tasks and the remote direct offloading tasks is obtained. Thus, under poor communication or low power conditions, the gateway adjusts the local processing volume and the amount of data offloaded to the cloud computing cluster according to the obtained data allocation ratio.

[0062] A baseline signal-to-noise ratio (SNR) of 10.0 dB corresponds to the lower limit of SNR required for the gateway to ensure a bit error rate below one ten-thousandth. When the actual SNR is much higher than the baseline SNR, the exponential decay term approaches 1, and task allocation is more affected by the remaining battery power. When the actual SNR is lower than the baseline SNR, the exponential correction term reduces the corrected trade-off factor, and the energy consumption penalty weight increases accordingly. The gateway reduces the amount of remote direct offloading tasks under weak channel conditions, thereby reducing battery power consumption caused by high-power retransmissions.

[0063] The cloud computing cluster receives regional spatially relevant feature vectors uploaded by each field edge computing gateway and, combined with historical spatiotemporal evolution data matrices, calculates the physical change gradient using a spatiotemporal evolution model. When regional macro-meteorological forecast data indicates rainfall or severe convective weather, and the magnitude of the physical change gradient exceeds a preset upper limit gradient threshold, the cloud computing cluster adjusts the smoothing coefficient and reduces the adjustment amount of the preset gating threshold. In this embodiment, the smoothing coefficient is updated to... And lower the preset gate threshold to Meanwhile, the cloud computing cluster combines regional macro-meteorological forecast data with farmland water and heat balance models to calculate a dynamic covariance matrix. Based on this dynamic covariance matrix, it reconstructs the segmented non-uniform quantization level intervals of the quantizer within the sensor monitoring terminal. Within the physical parameter transition intervals, the quantization step is reduced; within the smooth physical parameter intervals, the quantization step is increased. Based on this, a dynamic correction codebook parameter matrix is ​​generated. .

[0064] The cloud computing cluster will update the smoothing coefficient. Preset gate threshold Dynamically corrected codebook parameter matrix Furthermore, the uplink scheduling strategy for the field edge computing gateway is encapsulated as downlink control signaling and broadcast to the registers of each field edge computing gateway and sensor monitoring terminal via the downlink to complete the closed-loop calibration of the collaborative parameters. During the subsequent operation of the network, the sensor monitoring terminal performs state prediction, gating judgment and adaptive quantization encoding according to the updated smoothing coefficient, preset gating threshold and dynamic correction codebook parameter matrix. The field edge computing gateway performs local processing and cloud offloading according to the updated uplink scheduling strategy and data allocation ratio, so that the sensor monitoring terminal, the field edge computing gateway and the cloud computing cluster maintain parameter collaborative operation.

[0065] Example 6: The current sensor monitoring terminal has a built-in counter register to record the number of sampling periods in steady-state mode. When the number of consecutive single-byte status heartbeat data packets generated reaches a preset heartbeat upper limit threshold of 120, the sensor monitoring terminal exits steady-state mode in the current sampling period and wakes up the radio frequency communication module. It then fully collects the current physical parameters of soil volumetric water content, ground temperature, and air humidity, encapsulates them with a 16-bit raw sampling width to generate a baseline status data packet, and sends it to the field edge computing gateway. After receiving the baseline status data packet, the gateway updates the locally cached status baseline. Simultaneously, the terminal writes the current actual monitoring status vector into the historical status storage unit of the first-order autoregressive time-series filtering model and resets the actual monitoring status vector. With state prediction vector After resetting the counter register to zero, it re-enters the differential time-varying sensing loop to eliminate the drift effect of the truncated cumulative error caused by long-term micro-amplitude fluctuations on the calculation of the disturbance energy index.

[0066] The field edge computing gateway receives temporal residual quantized feature vectors uploaded by each sensor monitoring terminal within its communication area, and performs local outlier detection based on the multi-node spatial topology graph. The gateway then recovers the residual values ​​of each dimension according to the corresponding quantization codebook, forming the target sensor monitoring terminal temporal residual vector for distance calculation. And the temporal residual sequences of local neighbor nodes, then calculate the spatial Mahalanobis distance between the target sensing and monitoring terminal and the mean vector of the temporal residuals of adjacent sensing and monitoring terminals in the local neighbor node set. and will Deviation from the preset spatial consistency threshold Perform numerical comparison: ,in, For spatial Mahalanobis distance, For target sensing and monitoring terminals The time-domain residual vector obtained by recovering the corresponding time-domain residual quantized feature vector is... This is the mean vector of the temporal residuals of the local neighbor nodes. The time-domain residual covariance matrix of the local neighbor nodes. Let be the inverse matrix of the temporal residual covariance matrix of the local neighbor nodes. When a single sensing and monitoring terminal experiences an abnormal sudden change due to poor sensor probe contact or hardware aging, and satisfies , After five consecutive sampling cycles, the field edge computing gateway determines that the target sensing monitoring terminal is in a hardware sensing failure state. It resets the topological connection weight of the node in the multi-node spatial topology map to zero, so that the abnormal node no longer participates in the spatial fusion of regional spatial feature vectors. When assembling the feature data packet, it writes the fault node number and fault alarm flag and uploads it to the cloud computing cluster. This completes the dynamic adaptive adjustment of the edge topology and the isolation of abnormal nodes. In actual judgment, the preset spatial consistency deviation threshold of 2.80 is taken from the critical quantile of the chi-square distribution with 3 degrees of freedom and a confidence level of 99.50%. When the spatial Mahalanobis distance is greater than 2.80, the probability that the residual distribution of the node deviates from the normal confidence interval of the adjacent node cluster is higher than 99.50%. The preset number of abnormal cycles is set to five consecutive sampling cycles, i.e., 300s, to filter out single instantaneous pulse interference caused by occasional animal trampling probes and avoid misjudging hardware failure.

[0067] Field edge computing gateways measure the uplink wireless channel transmission bit error rate between themselves and cloud computing clusters in real time during field environment monitoring. When severe weather in the wild causes deep fading of the uplink wireless channel, and continuous measurements are taken... At this time, the gateway triggers a local degradation storage mechanism, ceasing the high-frequency transmission of regional spatially relevant feature vector data to the cloud computing cluster and initiating an internal 512MB non-volatile circular storage buffer. Subsequently, principal component analysis is used to reduce the dimensionality of the regional spatially relevant feature vectors, and the extracted core principal component features, along with timestamps, are written into the non-volatile circular storage buffer in the order of generation. When the wireless channel quality recovers and measurements are obtained... At this time, the field edge computing gateway exits the local degradation storage mechanism and retransmits the historical feature data frames stored in the buffer according to the timestamp sequence through the burst transmission mode, so that the monitoring data generated during the period of wireless communication link obstruction continues to be transmitted to the cloud computing cluster in the original time sequence.

[0068] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with, but not limited to, the technical features disclosed in the embodiments of the present invention that have similar functions.

Claims

1. A method for edge-cloud collaborative processing of farmland monitoring data, characterized in that, Includes the following steps: Step S1: The sensor monitoring terminal acquires the sampled data and calculates the temporal residual vector and disturbance energy index based on the state prediction vector of the previous cycle. When the disturbance energy index does not exceed the preset threshold, a 1-byte state heartbeat data packet is sent to the field edge computing gateway. When the disturbance energy index exceeds the preset threshold, the quantization bit width is adjusted according to the disturbance energy index, the temporal residual vector is adaptively quantized and encoded, and then sent to the field edge computing gateway. Step S2: The field edge computing gateway receives data sent by each sensor monitoring terminal, constructs a multi-node spatial topology map, and extracts regional spatially relevant feature vectors. Based on the gateway battery state of charge and the wireless channel signal-to-noise ratio, adjust the latency and energy consumption trade-off factors, solve the weighted cost objective function to minimize the objective function, and determine the data allocation ratio between edge local computing and data offloaded to the cloud computing cluster. In step S3, the cloud computing cluster receives the regional spatial feature vectors, calls the spatiotemporal evolution model to calculate the physical change gradient, and generates the smoothing coefficient of the state prediction vector based on the physical change gradient, the adjustment amount of the preset gate threshold, and the uplink scheduling strategy for the field edge computing gateway. This is then broadcast to the field edge computing gateway and the sensor monitoring terminal through the downlink control link to complete the closed-loop calibration of the collaborative parameters.

2. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S1, the state prediction vector is calculated by linearly weighting the actual monitoring state vector of the previous period and the state prediction vector of the previous period according to the smoothing coefficient; the disturbance energy index is determined by the ratio of the sum of squares of each dimension of the time domain residual vector to the preset benchmark variance; when the disturbance energy index does not exceed the preset threshold, the radio frequency communication unit of the sensor monitoring terminal remains in sleep mode and only reports the node's network status through single-byte short frames.

3. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S1, the data structure of the status heartbeat data packet includes a 7-bit node address identifier and a 1-bit steady-state flag. When the number of times the sensing and monitoring terminal continuously generates status heartbeat data packets reaches the preset heartbeat upper limit threshold, the current sampling data of the multidimensional physical parameters of the farmland is fully sampled in the current sampling period and encapsulated into a baseline status data packet and reported to the field edge computing gateway to reset the status baseline of the terminal side and the edge side.

4. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S2, a multi-node spatial topology graph is constructed and regional spatial correlation feature vectors are extracted, including the following steps: Step S21, construct a node topology graph based on the spatial geographic coordinates of each sensor monitoring terminal deployed in the same survey area, calculate the Euclidean distance between any two sensor monitoring terminals, calculate the topological connection weights between nodes based on the Gaussian kernel function, and generate a topology graph adjacency matrix; Step S22, calculate the degree matrix and normalized Laplacian matrix based on the topology graph adjacency matrix, input the temporal residual quantized feature vectors reported by each sensor monitoring terminal as graph node signals into a two-layer Chebyshev graph convolutional network, and extract regional spatial correlation feature vectors after eliminating spatial Euclidean distance correlation redundancy.

5. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S2, the data allocation ratio between edge local computing and data offloading to the cloud computing cluster is determined, including the following steps: Step S23, obtain the current battery state of charge and the current uplink wireless channel signal-to-noise ratio of the field edge computing gateway, and establish a joint cost objective function that includes local computing energy consumption, wireless transmission energy consumption, local queue processing latency, and network transmission latency; Step S24, determine the energy consumption penalty weight based on the negative correlation between the current battery state of charge and the preset rated capacity, determine the transmission revenue weight based on the positive correlation between the current uplink wireless channel signal-to-noise ratio, use the subgradient descent method to solve the optimal solution that minimizes the joint cost objective function, and output the data allocation ratio indicating the local feature dimensionality reduction task and the remote direct offloading task.

6. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S3, based on the smoothing coefficient of the state prediction vector generated by the physical change gradient, the adjustment amount of the preset gating threshold, and the uplink scheduling strategy for the field edge computing gateway, the following steps are included: Step S31, the cloud computing cluster inputs the regional spatial correlation feature vectors of multiple consecutive scheduling cycles in time into the bidirectional long short-term memory network to calculate the spatial diffusion rate and temporal evolution acceleration of the multidimensional physical parameters of the farmland environment in the two-dimensional grid space, and obtains the physical change gradient; Step S32, when the magnitude of the physical change gradient is less than the preset lower limit gradient threshold, the smoothing coefficient is increased and the adjustment amount of the preset gating threshold is increased to reduce the reporting frequency of the edge; when the magnitude of the physical change gradient is greater than the preset upper limit gradient threshold, the smoothing coefficient is decreased and the adjustment amount of the preset gating threshold is reduced to improve the sensitivity of edge mutation sampling, and the uplink scheduling strategy to shorten the feature upload cycle is issued to the field edge computing gateway.

7. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S3, the cloud computing cluster also calculates the dynamic covariance matrix of multidimensional physical parameters based on regional macro-meteorological forecast data and farmland water and heat balance model, reconstructs the segmented non-uniform quantization level interval of the quantizer in the sensor monitoring terminal, generates a dynamic correction codebook parameter matrix and broadcasts it into the register of the sensor monitoring terminal through the downlink control link, so that the sensor monitoring terminal reduces the quantization step in the physical parameter transition interval and increases the quantization step in the smooth interval.

8. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S2, when the field edge computing gateway detects that the uplink wireless channel transmission bit error rate between the gateway and the cloud computing cluster is continuously higher than the preset bit error rate threshold, a local degradation storage mechanism is triggered: Step S25, stop sending high-frequency feature data to the cloud computing cluster and start the non-volatile circular storage buffer inside the gateway; Step S26, perform principal component analysis to reduce the dimensionality of the extracted regional spatial related feature vectors, and write the core principal component features after dimensionality reduction along with the corresponding timestamps into the non-volatile circular storage buffer in sequence; when the uplink wireless channel transmission bit error rate is detected to recover to below the preset bit error rate threshold, the cached data is resent upwards in burst transmission mode according to the timestamp order.

9. The method for edge-cloud collaborative processing of farmland monitoring data according to claim 1, characterized in that, In step S2, after receiving the data sent by each sensor monitoring terminal, the field edge computing gateway also performs local outlier detection on the temporal residual quantization feature vector of each sensor monitoring terminal: Step S27, calculate the Mahalanobis distance between the temporal residual quantization feature vector of the target sensor monitoring terminal and the mean of the feature vectors of its physically adjacent sensor monitoring terminals; Step S28, when the Mahalanobis distance exceeds the preset spatial consistency deviation threshold and the duration reaches the preset number of abnormal cycles, determine that the target sensor monitoring terminal is in a hardware sensing fault state. In the multi-node spatial topology map, the topology connection weight corresponding to the target sensing and monitoring terminal is reset to zero. After adding the fault alarm flag bit to the generated feature data packet, it is uploaded to the cloud computing cluster.

Images