An artificial intelligence-based real-time monitoring and processing system for subsurface temperature and humidity

Abstract

The present application relates to the technical field of intelligent monitoring of agricultural environment, in particular to a real-time monitoring and processing system for soil temperature and humidity based on artificial intelligence, comprising: a data acquisition module, which periodically acquires multi-dimensional original data under the mulch through a sensor array; a data modeling module, which aligns and cleans the data in time and space, and constructs a three-dimensional environmental field model; a feature extraction module, which extracts key environmental features by calculating the local difference of environmental parameters in the time and space dimensions and filtering the areas exceeding the sensitive threshold; a prediction control module, which inputs the key features into a pre-trained dynamic evolution neural network, predicts the future evolution trajectory of the environmental field and generates a high-dimensional environmental state prediction tensor, and based on this, solves the optimal environmental regulation instruction combination that meets the crop growth constraints. The present application can accurately identify local environmental anomalies and make forward-looking predictions of the dynamic evolution of the entire environmental field, thereby realizing intelligent and refined agricultural environmental regulation.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for agricultural environments, and in particular to an artificial intelligence-based real-time monitoring and processing system for temperature and humidity under plastic film. Background Technology

[0002] Currently, in facility agriculture and mulching cultivation, monitoring the microenvironment under the mulch film often relies on deploying sensor arrays to periodically collect point data such as temperature and humidity. Existing technical solutions mostly treat each sensor as an independent data source or perform simple regional averaging calculations, using the overall average or readings from specific monitoring points as the basis for judging the environmental state. This approach simplifies the three-dimensional spatial environmental field into discrete, isolated point information, ignoring the continuous spatial distribution differences of environmental parameters and their dynamic correlations over time.

[0003] The shortcomings of existing technical solutions lie in their crude data utilization methods, making it difficult to effectively identify localized, rapidly changing environmental anomalies. Due to the lack of refined analysis of spatiotemporal gradient changes, the system cannot detect and locate potential crop stress areas in advance. Furthermore, conventional prediction methods are typically based on single-point historical data or two-dimensional planar models, predicting only a single parameter value or static distribution at a future moment, failing to provide a continuous and dynamic picture of the entire three-dimensional environmental field over a future period. This results in the generation of subsequent environmental control commands lacking a precise and forward-looking global state blueprint as support, leading to lagging or globally uniform control actions, making it difficult to achieve refined and predictive management that meets the needs of crop growth. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an artificial intelligence-based real-time monitoring and processing system for temperature and humidity under plastic film.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence, comprising: The data acquisition module periodically collects multi-dimensional raw data of the environment under the mulch film through a sensor array deployed under the mulch film; The data modeling module performs spatiotemporal alignment and outlier cleaning on the multidimensional raw data to form a regular spatiotemporal sequence data, and constructs a three-dimensional environmental field model based on the spatiotemporal sequence data. The feature extraction module extracts key environmental features with significant spatiotemporal variation gradients from the three-dimensional environmental field model. These key environmental features are obtained by calculating the local differences of environmental parameters in the spatial and temporal dimensions and filtering out regions that exceed a preset sensitivity threshold. The predictive control module inputs the key environmental features into a pre-trained dynamic evolutionary neural network. The dynamic evolutionary neural network predicts the evolution trajectory of the three-dimensional environmental field model within a set time period in the future based on the spatiotemporal pattern of the key environmental features, and generates a high-dimensional environmental state prediction tensor. Based on the high-dimensional environmental state prediction tensor, it solves for the optimal environmental regulation instructions that satisfy the preset crop growth constraints. The optimal environmental regulation instructions include a combination of irrigation amount, ventilation intensity and shading ratio.

[0006] As a further aspect of the present invention, the step of performing spatiotemporal alignment and outlier cleaning on the multidimensional raw data to form a regular spatiotemporal sequence data includes: The multidimensional raw data includes soil temperature, soil moisture, near-surface air temperature, near-surface air humidity, and light intensity; Receive multidimensional raw data streams from sensor arrays, and perform spatiotemporal grid interpolation on the data from different sensors based on the unique geographic identifier and timestamp of each sensor, unifying them to the same time step and spatial grid. For each environmental parameter time series at each spatial grid point, outliers are identified using a statistical distribution-based outlier detection method, and the environmental parameters are replaced by linear interpolation using the values ​​at adjacent spatiotemporal grid points. The cleaned and interpolated environmental parameters are arranged in chronological order and combined to form a multidimensional environmental parameter vector for each spatial grid point. The vector set of all grid points constitutes a regular spatiotemporal sequence data.

[0007] As a further aspect of the present invention, the construction of a three-dimensional environmental field model based on the spatiotemporal sequence data includes: The three-dimensional environmental field model uses two-dimensional geographic coordinates as one dimension, time as the second dimension, and multiple environmental parameters as the third dimension. A two-dimensional plane coordinate grid is defined based on the geographical extent of the farmland area, serving as the spatial base surface for the three-dimensional environmental field model; The continuous monitoring time is divided into equal-length time intervals, which serve as the time axis scale of the three-dimensional environmental field model. For each environmental parameter, the parameter value from the spatiotemporal sequence data corresponding to the spatiotemporal unit is filled in each spatiotemporal unit consisting of a spatial datum and a time axis. The values ​​of all environmental parameters in all spatiotemporal units together constitute the three-dimensional environmental field model.

[0008] As a further aspect of the present invention, the extraction of key environmental features with significant spatiotemporal variation gradients from the three-dimensional environmental field model includes: For each environmental parameter in the three-dimensional environmental field model, the numerical difference between adjacent grids in the spatial dimension is calculated to obtain the spatial gradient field. Simultaneously, the numerical differences between adjacent time periods in the time dimension are calculated to obtain the temporal gradient field; Spatiotemporal units whose gradient magnitudes exceed a preset sensitivity threshold in the spatial and temporal gradient fields are marked as sensitive units. The environmental parameter types, spatial locations, time points, and gradient directions corresponding to all sensitive units are extracted and packaged into a set of key environmental features.

[0009] As a further aspect of the present invention, the key environmental features are input into a pre-trained dynamic evolutionary neural network, which predicts the evolution trajectory of a three-dimensional environmental field model within a future set time period based on the spatiotemporal patterns of the key environmental features, generating a high-dimensional environmental state prediction tensor, including: The key environmental features are organized into a fixed-format input tensor, which includes feature type encoding, spatial location encoding, timestamp, and gradient information. The input tensor is fed into a dynamic evolutionary neural network, which includes spatiotemporal convolutional layers and recurrent memory units to capture spatiotemporal dependencies. The high-dimensional tensor is output by performing nonlinear transformations through multiple hidden layers of a dynamic evolutionary neural network. This high-dimensional tensor predicts the change sequence of all environmental parameters at all spatial grid points within a set future time period, i.e., the high-dimensional environmental state prediction tensor.

[0010] As a further aspect of the present invention, the step of solving for the optimal environmental regulation command that satisfies the preset crop growth constraints based on the high-dimensional environmental state prediction tensor includes: Construct an optimization model with a high-dimensional environmental state prediction tensor as input and the actuator adjustment variable as decision variable; In the optimization model, crop growth constraints are set, including suitable soil temperature range, soil moisture range, and air humidity range for each growth stage of the crop. Define an optimization objective function that aims to minimize the combined deviation between the predicted environmental state and the crop's optimal environmental state, while minimizing the total energy consumption of the actuators; The optimization model is solved using the gradient descent algorithm to obtain the optimal combination of actuator adjustment quantities that optimizes the objective function under crop growth constraints, i.e., the optimal environmental adjustment command.

[0011] As a further aspect of the present invention, the system further includes: The instruction execution module encodes the optimal environment adjustment instruction into a low-latency control signal and sends it to the actuator array; The closed-loop learning module receives feedback status data from the actuator array in real time. The feedback status data includes actual irrigation volume, actual ventilation status, and actual shading status. The module fuses and corrects the feedback status data with the concurrent three-dimensional environmental field model data, updates the current state of the three-dimensional environmental field model, and uses the updated three-dimensional environmental field model data to perform online incremental learning on the dynamic evolution neural network, adjusting its network weight parameters. Based on the incrementally learned dynamic evolution neural network, the closed-loop process from extracting key environmental features to generating low-latency control signals is re-executed to achieve adaptive real-time monitoring and processing of temperature and humidity under the mulch film. As a further aspect of the present invention, the step of encoding the optimal environment adjustment command into a low-latency control signal and sending it to the actuator array includes: The irrigation amount, ventilation intensity and shading ratio in the optimal environmental adjustment command are respectively converted into pulse width modulation duty cycle command, relay switching sequence command or motor speed command that can be recognized by the corresponding actuator. Each instruction is accompanied by a geographic identifier of the target executing agency and an execution priority tag; Control command packets with geographic identifiers and priority tags are broadcast to the actuator array via a wireless communication network.

[0012] As a further aspect of the present invention, the step of fusing and correcting the feedback state data with the concurrent three-dimensional environmental field model data to update the current state of the three-dimensional environmental field model includes: Receive feedback status data returned by the actuator array, the feedback status data including the type, magnitude, time and location of the actual executed action; The theoretical estimate of the environmental impact of the actual environmental regulation in the feedback state data is compared with the original predicted value of the corresponding spatiotemporal unit in the three-dimensional environmental field model. Based on the residuals generated by the comparison results, the Kalman filter algorithm is used to correct and update the parameter values ​​of the affected region in the three-dimensional environmental field model in real time.

[0013] As a further aspect of the present invention, the online incremental learning of the dynamic evolutionary neural network using the updated three-dimensional environmental field model data includes: From the updated 3D environmental field model, extract the complete environmental evolution sequence data within the most recent time window as new training samples; Calculate the loss function between the predicted output of the dynamic evolutionary neural network for the new training sample and the actual updated value; Using the backpropagation algorithm, the loss function gradient calculated based only on the new training samples is used to fine-tune and update some or all of the network weight parameters of the dynamically evolving neural network.

[0014] The process of re-executing the closed-loop flow from extracting key environmental features to generating low-latency control signals based on the dynamically evolving neural network after incremental learning includes: After completing the online incremental learning of the dynamic evolutionary neural network, the latest 3D environmental field model data is immediately used to re-extract key environmental features; The newly extracted key environmental features are input into the dynamically evolving neural network after incremental learning to predict the new environmental evolution. Based on the new environmental evolution prediction, the optimal environmental regulation command is solved again and a low-latency control signal is generated, thereby starting the next monitoring and processing cycle.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By calculating the local differences of environmental parameters in the spatial and temporal dimensions and filtering out areas exceeding preset sensitivity thresholds, the system directly achieves automatic identification and feature focusing of significantly changing areas in the environmental field under the plastic film. This approach enables the system to accurately extract spatiotemporal information characterizing local adverse conditions or key change processes from massive monitoring data, effectively overcoming the shortcomings of traditional methods where local anomaly signals are overwhelmed due to global averaging or single-point monitoring. This improves the detection sensitivity and positioning accuracy of the monitoring system for early, localized environmental stresses, while reducing the data dimensionality and noise interference in subsequent calculations.

[0016] By inputting the aforementioned key environmental features into a pre-trained dynamic evolutionary neural network, the system predicts the evolution trajectory of the complete three-dimensional environmental field in future time periods and generates a high-dimensional environmental state prediction tensor. This scheme constructs a unified prediction model that integrates information on continuous changes in multiple future timeframes, the entire field, and multiple parameters. It provides a dynamic and quantitative state blueprint covering the entire spatial region and future time windows for environmental regulation decisions, enabling the solution of irrigation, ventilation, and shading commands to be globally optimized based on accurate simulations of the future global landscape. This achieves a shift from delayed responses based on current or past states to forward-looking and precise control based on predicted states. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the artificial intelligence-based real-time monitoring and processing system for temperature and humidity under plastic film as described in this invention. Figure 2 Flowchart for spatiotemporal alignment and outlier cleaning; Figure 3 A flowchart for extracting key environmental features. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 The data acquisition module periodically collects multidimensional raw data of the environment under the mulch film using a sensor array deployed under the film. The data modeling module performs spatiotemporal alignment and outlier cleaning on the multidimensional raw data to form a regular spatiotemporal sequence data, and constructs a three-dimensional environmental field model based on the spatiotemporal sequence data. The feature extraction module extracts key environmental features with significant spatiotemporal variation gradients from the three-dimensional environmental field model. These key environmental features are obtained by calculating the local differences of environmental parameters in the spatial and temporal dimensions and filtering regions exceeding a preset sensitivity threshold. The prediction and control module inputs the key environmental features into a pre-trained dynamic evolutionary neural network. The dynamic evolutionary neural network predicts the evolution trajectory of the three-dimensional environmental field model within a set future time period based on the spatiotemporal patterns of the key environmental features, generating a high-dimensional environmental state prediction tensor. Based on the high-dimensional environmental state prediction tensor, it solves for the optimal environmental regulation instructions that satisfy preset crop growth constraints. The optimal environmental regulation instructions include a combination of irrigation amount, ventilation intensity, and shading ratio.

[0021] See Figure 2In one embodiment of the present invention, a rectangular farmland area of ​​100 square meters is considered. A sensor array containing 25 sensing nodes is deployed under the mulch film. The sensor array periodically collects multi-dimensional raw data such as soil temperature, soil moisture, near-ground air temperature, near-ground air humidity, and light intensity. The system receives the multi-dimensional raw data stream from the sensor array. Each data point in the multi-dimensional raw data stream is associated with a unique geographic identifier of the sensor and a timestamp accurate to the second. In some embodiments, the system performs spatiotemporal gridded interpolation on the data from different sensors based on the unique geographic identifier and timestamp of each sensor, unifying them to the same time step and spatial grid. For example, the farmland area is divided into a 5-meter by 5-meter spatial grid, and the time is unified to one time step per minute. A bilinear interpolation algorithm is used to estimate the sensor data at non-grid points or non-integer minute times to the value of each spatial grid point at each time step. In practical implementation, outliers are identified for each environmental parameter time series at each spatial grid point using a statistical distribution-based outlier detection method. For soil temperature time series, the mean and standard deviation of the series are calculated, and outliers are detected using the z-score method. Optionally, the preset sensitivity threshold is set to three, but it can be adjusted according to the actual application scenario. After identifying outliers, linear interpolation is performed using the values ​​of environmental parameters at adjacent spatiotemporal grid points. For example, for an abnormal soil moisture value at a certain grid point at a specific time, the arithmetic mean of the soil moisture values ​​of four adjacent grid points at the same time is used for replacement.

[0022] In some embodiments, the linear interpolation replacement process specifically involves taking the corresponding environmental parameter values ​​of the four adjacent spatial grid points (upper, lower, left, and right) at the same time for each spatiotemporal cell marked as an outlier, calculating the arithmetic mean of these four values ​​as the replacement value, and if data from adjacent grid points is unavailable, then data from adjacent time points are used for interpolation. In a specific implementation, by comparing the data before and after cleaning, spikes in the original data stream caused by instantaneous sensor malfunctions are corrected. For example, if the soil temperature reading of a sensor deviates significantly from the normal range at a certain moment, after outlier detection and linear interpolation replacement, the reading is adjusted to a reasonable value that is compatible with the surrounding environment.

[0023] Understandably, the cleaning process eliminates inconsistencies in the data. The cleaned and interpolated environmental parameters are arranged chronologically and combined into a multidimensional environmental parameter vector for each spatial grid point. Each vector contains values ​​for five dimensions: soil temperature, soil moisture, near-surface air temperature, near-surface air humidity, and light intensity. Optionally, the multidimensional environmental parameter vectors are stored as an array. The set of vectors from all grid points constitutes a well-organized spatiotemporal sequence data, where each data point has a consistent time and spatial index.

[0024] See Figure 3 In one embodiment of the present invention, taking a square farmland with an area of ​​one hectare as an example, the three-dimensional environmental field model uses two-dimensional geographic coordinates as one dimension, time as the second dimension, and multiple environmental parameters as the third dimension. These environmental parameters include soil temperature, soil moisture, near-surface air temperature, near-surface air humidity, and light intensity. A two-dimensional planar coordinate grid is defined based on the geographical extent of the farmland area, dividing the one-hectare farmland into one hundred regular square grids of ten meters by ten meters, serving as the spatial base of the three-dimensional environmental field model. The continuous monitoring time is divided into equal-length time intervals, for example, ten minutes as a basic time interval, serving as the time axis scale of the three-dimensional environmental field model. The time axis scale is continuously numbered starting from the monitoring start time. For each environmental parameter, within each spatiotemporal cell formed by the spatial datum and the time axis, the parameter value from the corresponding spatiotemporal sequence data is filled. The spatiotemporal cell is uniquely determined by a specific grid number and time period number. For example, the soil temperature value of the grid numbered (5,5) within the time period numbered 100 is read from the corresponding position in the regular spatiotemporal sequence data and filled. The values ​​of all environmental parameters in all spatiotemporal cells together constitute the three-dimensional environmental field model. The three-dimensional environmental field model is represented in computer memory as a four-dimensional tensor with dimensions (10,10,T,5), where the first two dimensions represent the rows and columns of the spatial grid, the third dimension represents the time step T, and the fourth dimension represents the five environmental parameters.

[0025] In some embodiments, key environmental features with significant spatiotemporal gradients are extracted from the constructed three-dimensional environmental field model. For each environmental parameter in the three-dimensional environmental field model, the numerical difference between adjacent grids in the spatial dimension is calculated. The numerical difference between adjacent grids in the east-west direction for the soil temperature parameter is denoted as... The numerical difference between adjacent grids in the north-south direction is denoted as Spatial gradient magnitude The calculation formula is: This calculation is performed on each environmental parameter at each grid point at each time step to obtain a spatial gradient field covering the entire spatial datum. The spatial gradient field quantifies the horizontal inhomogeneity of the environmental parameters and calculates their numerical differences between adjacent time intervals. For each environmental parameter at each grid point, the numerical change between two adjacent ten-minute intervals is calculated to obtain the temporal gradient field, which reflects the rate of change of the environmental parameters over time.

[0026] Optionally, a preset sensitivity threshold is set based on the crop's tolerance to environmental abrupt changes. Spatiotemporal units in the spatial and temporal gradient fields whose gradient amplitudes exceed the preset sensitivity threshold are marked as sensitive units. For example, the preset sensitivity threshold for soil moisture gradient is set to 5% per meter per hour. When the spatial gradient amplitude of soil moisture at a certain grid point at a specific time exceeds the preset sensitivity threshold, the sensitive unit is identified. A spatiotemporal unit is marked as a sensitive unit when the absolute value of the change in soil moisture over two consecutive time periods exceeds the corresponding temporal change. The environmental parameter type, spatial location, time point, and gradient direction corresponding to all sensitive units are extracted. The gradient direction includes the direction of the spatial gradient and whether the change is increasing or decreasing. These are packaged into a key environmental feature set. In some embodiments, data comparison shows that within one hour after irrigation, the spatial gradient amplitude of soil moisture at grid points near the water source is significantly higher than that at grid points far from the water source, and the temporal gradient value also shows a rapid increase. These spatiotemporal units are successfully marked and extracted. It can be understood that the key environmental feature set focuses on areas and times of drastic change in the environmental field, providing high-information-density input for subsequent predictions. Optionally, the key environmental feature set is stored in list form, where each entry records the complete feature information of a sensitive unit.

[0027] In one embodiment of the present invention, the key environmental feature set is organized into a fixed-format input tensor. The input tensor is a four-dimensional array, whose dimensions represent the batch size, feature sequence length, number of feature items, and encoding dimension of each feature item, respectively. Each feature item includes feature type encoding, spatial location encoding, timestamp, and gradient information. For example, for a key environmental feature set containing ten sensitive units, the feature type encoding uses one-hot vectors to represent different types such as soil temperature and soil moisture; the spatial location encoding uses grid row and column coordinates; the timestamp is the number of minutes since the start time; and the gradient information includes spatial gradient components and temporal difference values. In some embodiments, the input tensor is input into a dynamic evolutionary neural network. The dynamic evolutionary neural network includes a spatiotemporal convolutional layer and a recurrent memory unit to capture spatiotemporal dependencies. The spatiotemporal convolutional layer uses a three-dimensional convolutional kernel to extract features from the input tensor in both time and spatial dimensions, and the recurrent memory unit uses a long short-term memory network structure to process long-term dependencies in the time series. The dynamic evolutionary neural network performs nonlinear transformations through multiple hidden layers, ultimately outputting a high-dimensional tensor. This high-dimensional tensor predicts the sequence of changes in all environmental parameters at all spatial grid points within a set future time period. This is the high-dimensional environmental state prediction tensor. For example, it predicts the state of five environmental parameters per minute at one hundred spatial grid points within the next hour. The dimensions of the output high-dimensional environmental state prediction tensor are (100, 60, 5), corresponding to the number of spatial grid points, the future prediction time step, and the number of environmental parameters, respectively.

[0028] In practical implementation, an optimization model is constructed with a high-dimensional environmental state prediction tensor as input and actuator adjustment quantities as decision variables. The decision variables include the planned irrigation amount, ventilation intensity, and shading ratio for the future period. Crop growth constraints are set in the optimization model, based on the physiological needs of the target crop during the flowering period, including suitable soil temperature range, soil moisture range, and air humidity range for each growth stage. For example, the soil moisture range constraint is 70% to 80% of field capacity. An optimization objective function is defined, aiming to minimize the comprehensive deviation between the predicted environmental state and the optimal environmental state for the crop, while minimizing the total energy consumption of the actuators. The comprehensive deviation is calculated using a weighted mean square error, and the total energy consumption is the sum of the products of the power consumption and action time of each actuator. The gradient descent algorithm is used to solve the optimization model. The gradient descent algorithm iteratively adjusts the values ​​of the decision variables, calculates the partial derivative of the objective function with respect to each decision variable, and updates the variables along the negative gradient direction to obtain the optimal combination of actuator adjustment quantities that optimizes the objective function while satisfying the crop growth constraints; this is the optimal environmental regulation command. An example formula for the optimization objective function is:

[0029] in: Represents the objective function value. This represents the predicted value in the high-dimensional environment state prediction tensor. This represents the optimal environmental condition value for crops. This represents the power consumption per unit time of the i-th type of actuator. This represents the planned action time of the i-th type of executive agency. and It is a weighting coefficient that balances environmental deviations and energy consumption.

[0030] It is understandable that the optimization process seeks a balance among multiple competing objectives. In some embodiments, data comparisons show that when predictions indicate excessively high local humidity, the optimization model may tend to increase ventilation intensity rather than reduce irrigation, because increasing ventilation can more effectively reduce air humidity and meet constraints within acceptable energy consumption. Optionally, the gradient descent algorithm sets a maximum number of iterations and a loss change convergence threshold. The computation stops when the maximum number of iterations is reached or the loss change is less than the threshold, and the current optimal environmental adjustment command is output.

[0031] In one embodiment of the invention, the instruction execution module receives an optimal environmental regulation instruction from the predictive control module. This instruction includes specific values ​​for irrigation volume, ventilation intensity, and shading ratio for a specific area within the next ten minutes. In a specific implementation, the instruction execution module encodes the optimal environmental regulation instruction into a low-latency control signal and sends it to an actuator array, which includes water valves, fans, and shading curtain motors deployed in the field. In some embodiments, the irrigation volume, ventilation intensity, and shading ratio in the optimal environmental regulation instruction are converted into pulse width modulation duty cycle instructions, relay switching sequence instructions, or motor speed instructions that the corresponding actuators can recognize. For example, for a drip irrigation instruction requiring five minutes of operation at a flow rate of two liters per second, it is converted into sending a pulse width modulation signal with a 100% duty cycle and a duration of 300 seconds to a specific electromagnetic water valve. The instruction execution module attaches a geographic identifier and an execution priority tag to each instruction. The geographic identifier corresponds to the coordinates of the sensor grid, and the priority tag is set according to the urgency of the environmental regulation; for example, an emergency irrigation instruction to maintain the minimum survival level of crops is given the highest priority. Control command packets with geographic identifiers and priority tags are broadcast to the actuator array via a wireless communication network. The wireless communication network uses low-power wide-area network technology. The control command packets contain command type, target address, action parameters, and checksum. Optionally, the conversion relationship between different command types can be implemented through a preset mapping table. The mapping table defines the mathematical relationship or direct correspondence between adjustment command values ​​and specific control signals, as shown in Table 1.

[0032] Table 1: Relationship between Command Type and Control Signal Conversion

[0033] End-to-end delay of control command packets in the wireless channel It consists of transmission delay, propagation delay, and processing delay, and their relationship can be expressed as:

[0034] in: It is the time required for a data packet to travel from the transmission queue to being fully delivered into the channel, and it depends on the data packet size and link bandwidth. It is the time required for a signal to travel from the transmitter to the receiver in the physical medium. This refers to the time consumed by network nodes in processing data packets, such as verification and routing. It's understandable that low-latency control signals require... The numerical values ​​are kept as small as possible to ensure real-time control. The closed-loop learning module receives feedback status data from the actuator array in real time. The feedback status data is generated by the status sensors or drive feedback circuits on the actuators. The feedback status data includes actual irrigation volume, actual ventilation status, and actual shading status. Actual irrigation volume may be fed back as actual opening time and flow meter reading, actual ventilation status as actual fan speed, and actual shading status as actual opening and closing angle of the shading curtain. In specific implementation, the feedback status data is fused and corrected with the concurrent three-dimensional environmental field model data to update the current state of the three-dimensional environmental field model. The updated three-dimensional environmental field model data is used to perform online incremental learning on the dynamic evolution neural network, adjusting its network weight parameters. It can be understood that the state of the three-dimensional environmental field model after the command execution needs to be corrected according to the actual execution effect. Based on the incrementally learned dynamic evolution neural network, the closed-loop process from extracting key environmental features to generating low-latency control signals is re-executed to achieve adaptive real-time monitoring and processing of temperature and humidity under the mulch film. For example, after an irrigation command is executed, the system retrains the neural network according to the actual soil moisture changes and immediately starts a new round of prediction and decision-making based on the new state.

[0035] In one embodiment of the present invention, the closed-loop learning module receives feedback state data returned by the actuator array. The feedback state data includes the type, magnitude, time, and location of the actual executed action. In a specific implementation, the theoretical estimate of the environmental impact of the actual environmental regulation in the feedback state data is compared with the original predicted value of the corresponding spatiotemporal unit in the three-dimensional environmental field model. The theoretical estimate is calculated based on a known physical or empirical model, such as the increase in soil moisture estimated based on actual irrigation volume and soil type. In some embodiments, based on the residual generated by the comparison result, a Kalman filter algorithm is used to correct and update the parameter values ​​of the regulated area in the three-dimensional environmental field model in real time. The Kalman filter algorithm, through two steps—state prediction and measurement update—integrates the theoretical estimate with new observation data from actual sensors to obtain the optimal estimate of the state of the three-dimensional environmental field model. The formula for the state update of a soil moisture parameter is expressed as:

[0036] in: It is the soil moisture state estimation vector updated at time k. It is the prior state estimation vector obtained at time k based on the state at the previous time and the theoretical model. It is the Kalman gain matrix at time k. It is the actual measurement vector fed back by the sensor at time k. The observation matrix is ​​used to map the state space to the measurement space. It can be understood that the Kalman filter algorithm calculates the residuals... And multiply by the Kalman gain matrix This is used to correct prior estimates, thereby reducing the error between model predictions and actual environmental responses.

[0037] In specific implementation, the updated 3D environmental field model data is used to perform online incremental learning on the dynamic evolutionary neural network. Complete environmental evolution sequence data within the most recent time window is extracted from the updated 3D environmental field model as new training samples. The time window length is set, for example, to the past 60 minutes, containing data from twelve consecutive 5-minute intervals. In some embodiments, a loss function is calculated between the predicted output of the dynamic evolutionary neural network for the new training samples and the actual updated values. The loss function is typically in the form of mean squared error, measuring the difference between the environmental parameter change sequence predicted by the neural network and the actual change sequence after Kalman filtering correction. Using the backpropagation algorithm, based only on the gradient of the loss function calculated from the new training samples, some or all of the network weight parameters of the dynamic evolutionary neural network are fine-tuned and updated. The backpropagation algorithm starts from the output layer, calculating the contribution of each network weight to the loss layer by layer according to the loss function value, and making small adjustments to the weights along the gradient descent direction. Optionally, the online incremental learning process uses a small learning rate, for example, set to one-tenth of the initial training phase learning rate, to prevent excessive forgetting of long-term historical patterns.

[0038] After completing online incremental learning of the dynamic evolutionary neural network, the latest 3D environmental field model data is immediately used to re-extract key environmental features. This latest 3D environmental field model data already includes results corrected by feedback state data. The re-extracted key environmental features are input into the incrementally learned dynamic evolutionary neural network to predict new environmental evolution. Based on this new environmental evolution prediction, the optimal environmental regulation command is solved again, and a low-latency control signal is generated, thus initiating the next monitoring and processing cycle. In some embodiments, data comparison shows that after over-irrigation due to model prediction bias, feedback state data triggered the 3D environmental field model to correct soil moisture. After online incremental learning, the dynamic evolutionary neural network more accurately predicts the rate of soil moisture increase under similar weather conditions in subsequent predictions, thereby generating more water-saving irrigation commands.

[0039] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence, characterized in that, include: The data acquisition module periodically collects multi-dimensional raw data of the environment under the mulch film through a sensor array deployed under the mulch film; The data modeling module performs spatiotemporal alignment and outlier cleaning on the multidimensional raw data to form a regular spatiotemporal sequence data, and constructs a three-dimensional environmental field model based on the spatiotemporal sequence data. The feature extraction module extracts key environmental features with significant spatiotemporal variation gradients from the three-dimensional environmental field model. These key environmental features are obtained by calculating the local differences of environmental parameters in the spatial and temporal dimensions and filtering out regions that exceed a preset sensitivity threshold. The predictive control module inputs the key environmental features into a pre-trained dynamic evolutionary neural network. The dynamic evolutionary neural network predicts the evolution trajectory of the three-dimensional environmental field model within a set time period in the future based on the spatiotemporal pattern of the key environmental features, and generates a high-dimensional environmental state prediction tensor. Based on the high-dimensional environmental state prediction tensor, it solves for the optimal environmental regulation instructions that satisfy the preset crop growth constraints. The optimal environmental regulation instructions include a combination of irrigation amount, ventilation intensity and shading ratio.

2. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 1, characterized in that, The process of performing spatiotemporal alignment and outlier cleaning on the multidimensional raw data to form a regular spatiotemporal sequence data includes: The multidimensional raw data includes soil temperature, soil moisture, near-surface air temperature, near-surface air humidity, and light intensity; Receive multidimensional raw data streams from sensor arrays, and perform spatiotemporal grid interpolation on the data from different sensors based on the unique geographic identifier and timestamp of each sensor, unifying them to the same time step and spatial grid. For each environmental parameter time series at each spatial grid point, outliers are identified using a statistical distribution-based outlier detection method, and the environmental parameters are replaced by linear interpolation using the values ​​at adjacent spatiotemporal grid points. The cleaned and interpolated environmental parameters are arranged in chronological order and combined to form a multidimensional environmental parameter vector for each spatial grid point. The vector set of all grid points constitutes a regular spatiotemporal sequence data.

3. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 1, characterized in that, The construction of a three-dimensional environmental field model based on the spatiotemporal sequence data includes: The three-dimensional environmental field model uses two-dimensional geographic coordinates as one dimension, time as the second dimension, and multiple environmental parameters as the third dimension. A two-dimensional plane coordinate grid is defined based on the geographical extent of the farmland area, serving as the spatial base surface for the three-dimensional environmental field model; The continuous monitoring time is divided into equal-length time intervals, which serve as the time axis scale of the three-dimensional environmental field model. For each environmental parameter, the parameter value from the spatiotemporal sequence data corresponding to the spatiotemporal unit is filled in each spatiotemporal unit consisting of a spatial datum and a time axis. The values ​​of all environmental parameters in all spatiotemporal units together constitute the three-dimensional environmental field model.

4. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 3, characterized in that, The extraction of key environmental features with significant spatiotemporal gradients from the three-dimensional environmental field model includes: For each environmental parameter in the three-dimensional environmental field model, the numerical difference between adjacent grids in the spatial dimension is calculated to obtain the spatial gradient field. Simultaneously, the numerical differences between adjacent time periods in the time dimension are calculated to obtain the temporal gradient field; Spatiotemporal units whose gradient magnitudes exceed a preset sensitivity threshold in the spatial and temporal gradient fields are marked as sensitive units. The environmental parameter types, spatial locations, time points, and gradient directions corresponding to all sensitive units are extracted and packaged into a set of key environmental features.

5. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 1, characterized in that, The process involves inputting the key environmental features into a pre-trained dynamic evolutionary neural network. The dynamic evolutionary neural network predicts the evolution trajectory of a three-dimensional environmental field model within a predetermined future time period based on the spatiotemporal patterns of the key environmental features, generating a high-dimensional environmental state prediction tensor, including: The key environmental features are organized into a fixed-format input tensor, which includes feature type encoding, spatial location encoding, timestamp, and gradient information. The input tensor is fed into a dynamic evolutionary neural network, which includes spatiotemporal convolutional layers and recurrent memory units to capture spatiotemporal dependencies. The high-dimensional tensor is output by performing nonlinear transformations through multiple hidden layers of a dynamic evolutionary neural network. This high-dimensional tensor predicts the sequence of changes in all environmental parameters at all spatial grid points within a set future time period.

6. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 5, characterized in that, The step of solving for the optimal environmental regulation command that satisfies the preset crop growth constraints based on the high-dimensional environmental state prediction tensor includes: Construct an optimization model with a high-dimensional environmental state prediction tensor as input and the actuator adjustment variable as decision variable; In the optimization model, crop growth constraints are set, including suitable soil temperature range, soil moisture range, and air humidity range for each growth stage of the crop. Define an optimization objective function that aims to minimize the combined deviation between the predicted environmental state and the crop's optimal environmental state, while minimizing the total energy consumption of the actuators; The optimization model is solved using the gradient descent algorithm to obtain the optimal combination of actuator adjustment quantities that optimizes the objective function under crop growth constraints, i.e., the optimal environmental adjustment command.

7. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 1, characterized in that, The system also includes: The instruction execution module encodes the optimal environment adjustment instruction into a low-latency control signal and sends it to the actuator array; The closed-loop learning module receives real-time feedback status data from the actuator array, including actual irrigation volume, actual ventilation status, and actual shading status. This feedback status data is then fused and corrected with concurrent 3D environmental field model data to update the current state of the 3D environmental field model. The updated 3D environmental field model data is then used to perform online incremental learning on the dynamic evolutionary neural network, adjusting its network weight parameters. Based on the incrementally learned dynamic evolutionary neural network, the closed-loop process from extracting key environmental features to generating low-latency control signals is re-executed, achieving adaptive real-time monitoring and processing of temperature and humidity under the mulch film.

8. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 7, characterized in that, The step of encoding the optimal environment adjustment command into a low-latency control signal and sending it to the actuator array includes: The irrigation amount, ventilation intensity and shading ratio in the optimal environmental adjustment command are respectively converted into pulse width modulation duty cycle command, relay switching sequence command or motor speed command that can be recognized by the corresponding actuator. Each instruction is accompanied by a geographic identifier of the target executing agency and an execution priority tag; Control command packets with geographic identifiers and priority tags are broadcast to the actuator array via a wireless communication network.

9. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 7, characterized in that, The step of fusing and correcting the feedback state data with the concurrent 3D environmental field model data to update the current state of the 3D environmental field model includes: Receive feedback status data returned by the actuator array, the feedback status data including the type, magnitude, time and location of the actual executed action; The theoretical estimate of the environmental impact of the actual environmental regulation in the feedback state data is compared with the original predicted value of the corresponding spatiotemporal unit in the three-dimensional environmental field model. Based on the residuals generated by the comparison results, the Kalman filter algorithm is used to correct and update the parameter values ​​of the affected region in the three-dimensional environmental field model in real time.

10. The real-time monitoring and processing system for temperature and humidity under plastic film based on artificial intelligence as described in claim 7, characterized in that, The method of using updated 3D environmental field model data to perform online incremental learning of the dynamic evolutionary neural network includes: From the updated 3D environmental field model, extract the complete environmental evolution sequence data within the most recent time window as new training samples; Calculate the loss function between the predicted output of the dynamic evolutionary neural network for the new training sample and the actual updated value; Using the backpropagation algorithm, the loss function gradient calculated based only on the new training samples is used to fine-tune and update some or all of the network weight parameters of the dynamically evolving neural network. The process of re-executing the closed-loop flow from extracting key environmental features to generating low-latency control signals based on the dynamically evolving neural network after incremental learning includes: After completing the online incremental learning of the dynamic evolutionary neural network, the latest 3D environmental field model data is immediately used to re-extract key environmental features; The newly extracted key environmental features are input into the dynamically evolving neural network after incremental learning to predict the new environmental evolution. Based on the new environmental evolution prediction, the optimal environmental regulation command is solved again and a low-latency control signal is generated, thereby starting the next monitoring and processing cycle.

Images