An intelligent agricultural greenhouse environment monitoring system based on an internet of things

Abstract

The application discloses a kind of based on Internet of Things's smart agricultural greenhouse environment monitoring system, comprising: data acquisition module: obtain fan speed value, and greenhouse outside temperature value and greenhouse inside temperature value;Anti-noise time sequence fusion coding module: a plurality of predetermined time point greenhouse outside temperature value, greenhouse inside temperature value and fan speed value are arranged in time sequence and input into a shared anti-noise time sequence encoder, generate a fusion anti-noise greenhouse outside temperature information, anti-noise greenhouse inside temperature information and fan speed information context time sequence characteristic representation;Temperature difference-fan nonlinear correlation modeling module: the context time sequence characteristic representation is input to time sequence cross attention module, and a context-aware transfer feature vector is output;Continuous fan speed control module: the context-aware transfer feature vector is input to regression head, and regression head outputs the continuous value indicating fan speed adjustment or target speed.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a smart agricultural greenhouse environment monitoring system based on the Internet of Things. Background Technology

[0002] Traditional agricultural greenhouses rely on manual experience for environmental control, which suffers from problems such as monitoring lag, insufficient data accuracy, low response efficiency, and serious resource waste. This makes it difficult to meet the real-time, precise sensing and dynamic control requirements of modern precision agriculture for key environmental parameters such as temperature, humidity, light, carbon dioxide concentration, and soil moisture. The aim of this project is to deploy an Internet of Things (IoT) sensor network to achieve real-time collection, transmission, and fusion analysis of multi-dimensional environmental data. Combined with intelligent algorithms, an environment-crop growth model is constructed to achieve precise linkage control of equipment such as automatic irrigation, ventilation, supplemental lighting, and temperature and humidity regulation. This optimizes the crop growth environment, improves yield and quality, and reduces energy consumption and labor costs. Its significance lies not only in promoting the transformation of agricultural production from "experience-based" to "scientific decision-making" through data-driven refined management, but also in promoting the efficient use of agricultural resources, reducing environmental pollution caused by excessive use of chemical fertilizers and pesticides, and enhancing agricultural disaster resistance and sustainable development through the large-scale application of IoT technology. Simultaneously, it provides basic support for agricultural big data accumulation, intelligent equipment development, and collaborative innovation in the agricultural industry chain, ultimately achieving multiple goals such as improved agricultural production efficiency, increased farmer income, and improved rural ecology, thus contributing to the implementation of the rural revitalization strategy and the high-quality development of modern agriculture.

[0003] In existing technologies, the spatial enhanced temperature difference-fan domain transfer feature matrix is ​​processed by a classifier to obtain classification results. The classification results are used to indicate whether the fan speed value at the current time point should be increased, decreased, or maintained. The classifier uses softmax to output three types of decisions: increase / decrease / maintain. However, in agricultural environments, the adjustment of fan speed may have continuous transition intervals, such as fuzzy boundaries of slight increase or significant increase. Hard classification methods cannot quantify the adjustment range, which may lead to an overly discretized control strategy and make it difficult to achieve smooth and continuous speed adjustment. Therefore, a smart agricultural greenhouse environment monitoring system based on the Internet of Things is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an Internet of Things-based smart agricultural greenhouse environment monitoring system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart agricultural greenhouse environmental monitoring system based on the Internet of Things includes: Data acquisition module: acquires fan speed value, and acquires the outside temperature value and inside temperature value of the greenhouse at multiple predetermined time points within a predetermined time period through temperature sensors; Noise-resistant temporal fusion encoding module: The greenhouse outside temperature value, greenhouse inside temperature value and fan speed value at multiple predetermined time points are arranged in time sequence and input into a shared noise-resistant temporal encoder to generate a contextual temporal feature representation that fuses the noise-resistant greenhouse outside temperature information, noise-resistant greenhouse inside temperature information and fan speed information. The shared noise-resistant temporal encoder is configured to learn a temporal pattern that is insensitive to sensor noise. Temperature difference-fan nonlinear correlation modeling module: The context temporal feature representation is input to the temporal cross attention module. The temporal cross attention module dynamically calculates the nonlinear dependency between the fused temperature difference-related information as a query and the historical state information of the fan speed as a key value, and outputs a context-aware transition feature vector. The context-aware transition feature vector encodes the complex dynamic correlation between temperature difference and fan speed. Continuous fan speed control module: The context-aware transfer feature vector is input to the regression head, and the regression head outputs a continuous value representing the fan speed adjustment amount or the target speed. The continuous fan speed control module includes a proportional feedback fine-tuning unit based on the current instantaneous temperature difference, which is used to fine-tune the continuous value output by the regression head in real time and generate the final fan speed control command.

[0006] The above technical solution further includes: Furthermore, the specific steps for the data acquisition module to obtain the fan speed value, the outside temperature value, and the inside temperature value of the greenhouse are as follows: A first temperature sensor and a second temperature sensor are deployed at the shaded and ventilated area outside the greenhouse and at the height of the crop canopy inside the greenhouse, respectively. A speed sensor is installed on the ventilation fan motor of the greenhouse. An edge computing controller is configured to communicate with the first temperature sensor, the second temperature sensor, and the speed sensor. In response to an environmental control cycle command or the arrival of a predetermined time interval, the edge computing controller initiates a new data acquisition task for a predetermined time period. During the predetermined time period, the edge computing controller synchronously reads and records the outside temperature value of the greenhouse measured by the first temperature sensor, the inside temperature value of the greenhouse measured by the second temperature sensor, and the fan speed value measured by the speed sensor at multiple predetermined time points, forming three sets of time-aligned original time-series data sequences. The three sets of original time-series data sequences are cached locally and basic data cleaning, including missing value imputation and obvious outlier removal, is performed to generate a standardized time-series dataset to be processed.

[0007] Furthermore, the shared noise-resistant temporal encoder includes a one-dimensional convolutional layer, a positional encoding layer, and a Transformer encoder layer connected in sequence; The one-dimensional convolutional layer is used to extract local features and perform preliminary smoothing filtering on the input temporal data. The positional encoding layer is used to inject sequential information into the temporal data. The Transformer encoder layer learns long-term temporal dependencies and generates the contextual temporal feature representation through a self-attention mechanism.

[0008] Furthermore, the temporal cross-attention module is a multi-head cross-attention mechanism. It captures diverse nonlinear correlation patterns between the temperature difference-related information and the historical state information of the fan speed from different subspaces through multiple attention heads that are computed in parallel. The outputs of all attention heads are concatenated and linearly transformed to form a unified context-aware transfer feature vector.

[0009] Furthermore, the system includes an interpretable fuzzy inference branch, which receives the context-aware transition feature vector and maps the key dimensions of the context-aware transition feature vector to predefined fuzzy variables, performs inference through a fuzzy rule base, and outputs speed adjustment suggestions.

[0010] Furthermore, the regression head is a fully connected neural network layer, and the input of the regression head is the context-aware transfer feature vector. After the context-aware transfer feature vector undergoes nonlinear transformation and information compression through one or more fully connected layers, it is output through a linear activation function layer to generate continuous scalar values, which are the feedforward control quantities.

[0011] Furthermore, the proportional feedback fine-tuning unit obtains the real-time feedback adjustment amount by multiplying the difference between the instantaneous temperature difference inside and outside the greenhouse at the current moment by a preset proportional coefficient, and then weights and sums the real-time feedback adjustment amount with the continuous value output by the regression head to form the final fan speed control command.

[0012] Furthermore, during the training phase, the joint optimization objective of the shared noise-resistant temporal encoder and the temporal cross-attention module includes a main regression loss function and an auxiliary temporal smoothness constraint loss function, which is applied to the intermediate layer or output layer of the context temporal feature representation.

[0013] Furthermore, it also includes: Visual monitoring and pest identification module: Captures crop images through a camera and performs pest detection and classification based on the ResNet60 deep learning model; Multimodal data fusion module: fuses pest identification results with environmental sensor data for feature fusion; Intelligent decision support module: Based on fusion features and agricultural knowledge graph, it generates suggestions for pest and disease control and environmental regulation.

[0014] The present invention has the following beneficial effects: In this invention, the final decision output is changed from a discrete classification of "increase / decrease / hold" to a continuous speed adjustment amount or target value output by a regression head. This solves the problems of fuzzy control action boundaries and assumed adjustment range caused by traditional hard classification, allowing the system to make smooth speed adjustments of small or large magnitudes according to the actual situation. It avoids temperature oscillations or frequent fan start-stop due to the discretization of control commands, which helps to save energy and extend equipment life. Attached Figure Description

[0015] Figure 1 This is a system block diagram of a smart agricultural greenhouse environment monitoring system based on the Internet of Things proposed in this invention; Figure 2 This is a diagram of the internal structure of the noise-resistant timing encoder proposed in this invention; Figure 3 This is a flowchart of the temporal cross-attention and regression control proposed in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-3 As shown, this invention is a smart agricultural greenhouse environment monitoring system based on the Internet of Things, comprising: Data acquisition module: acquires fan speed value, and acquires the outside temperature value and inside temperature value of the greenhouse at multiple predetermined time points within a predetermined time period through temperature sensors; More specifically, cameras with night vision capabilities are deployed inside the greenhouse to periodically capture images of the crop canopy. The image data is preprocessed by edge computing nodes and then input into a lightweight ResNet60 model for real-time pest detection. The detection results, along with environmental data such as temperature and humidity, are simultaneously uploaded to a cloud-based decision-making platform. Noise-resistant temporal fusion encoding module: The greenhouse outside temperature value, greenhouse inside temperature value and fan speed value at multiple predetermined time points are arranged in time sequence and input into a shared noise-resistant temporal encoder to generate a contextual temporal feature representation that fuses the noise-resistant greenhouse outside temperature information, noise-resistant greenhouse inside temperature information and fan speed information. The shared noise-resistant temporal encoder is configured to learn a temporal pattern that is insensitive to sensor noise. Temperature difference-fan nonlinear correlation modeling module: The context temporal feature representation is input to the temporal cross attention module. The temporal cross attention module dynamically calculates the nonlinear dependency between the fused temperature difference-related information as a query and the historical state information of the fan speed as a key value, and outputs a context-aware transition feature vector. The context-aware transition feature vector encodes the complex dynamic correlation between temperature difference and fan speed. Continuous fan speed control module: The context-aware transfer feature vector is input to the regression head, and the regression head outputs a continuous value representing the fan speed adjustment amount or the target speed. The continuous fan speed control module includes a proportional feedback fine-tuning unit based on the current instantaneous temperature difference, which is used to fine-tune the continuous value output by the regression head in real time to generate the final fan speed control command.

[0018] In one embodiment, the specific steps for the data acquisition module to obtain the fan speed value, the outside temperature value, and the inside temperature value of the greenhouse are as follows: A first temperature sensor and a second temperature sensor are deployed at the shaded and ventilated area outside the greenhouse and at the height of the crop canopy inside the greenhouse, respectively. A speed sensor is installed on the ventilation fan motor of the greenhouse. An edge computing controller is configured to communicate with the first temperature sensor, the second temperature sensor, and the speed sensor. In response to an environmental control cycle command or the arrival of a predetermined time interval, the edge computing controller initiates a new data acquisition task for a predetermined time period. During the predetermined time period, the edge computing controller synchronously reads and records the outside temperature value of the greenhouse measured by the first temperature sensor, the inside temperature value of the greenhouse measured by the second temperature sensor, and the fan speed value measured by the speed sensor at multiple predetermined time points, forming three sets of time-aligned original time-series data sequences. The multiple predetermined time points are determined as follows: within the predetermined time period, a time point sequence is generated according to a fixed and equally spaced sampling frequency, or the sampling density is dynamically adjusted according to the change curve of the light intensity outside the greenhouse, and a higher sampling frequency is used during the stage of drastic light change. More specifically: The edge computing controller simultaneously sends data request commands to the first temperature sensor, the second temperature sensor, and the speed sensor at each predetermined time point. Wait for and receive response data packets from each sensor, and verify the data packets; If the verification is successful, the external temperature value, internal temperature value, and fan speed value of the greenhouse will be extracted and stored respectively. If a sensor fails to verify data or times out without responding, the data at that time point is recorded as missing, and a retry acquisition is triggered or the data is marked as to be interpolated. The three sets of original time-series data sequences are cached locally and basic data cleaning, including missing value imputation and obvious outlier removal, is performed to generate a standardized time-series dataset to be processed.

[0019] The rule for removing obvious outliers is as follows: if the temperature value outside the greenhouse or the temperature value inside the greenhouse at a certain time point changes abruptly beyond the reasonable range of physics relative to the values ​​at adjacent time points before and after it, or if the fan speed value exceeds the rated speed range of the motor, then the corresponding value at that time point is marked as an outlier and removed.

[0020] In one embodiment, the shared noise-resistant temporal encoder includes a one-dimensional convolutional layer, a positional encoding layer, and a Transformer encoder layer connected in sequence. The one-dimensional convolutional layer is used to extract local features and perform preliminary smoothing filtering on the input temporal data. The positional encoding layer is used to inject sequential information into the temporal data. The Transformer encoder layer learns long-term temporal dependencies and generates the contextual temporal feature representation through a self-attention mechanism.

[0021] More specifically: the shared noise-resistant temporal encoder generates the contextual temporal feature representation according to the following steps: First, the one-dimensional convolutional layer performs a convolution operation on the joint input sequence consisting of the greenhouse outside temperature value, greenhouse inside temperature value, and fan speed value arranged in time sequence to perform local feature extraction and generate a primary feature map with a smoothing filtering effect. The one-dimensional convolutional layer uses multiple one-dimensional convolutional kernels with different widths to process the input sequence in parallel to simultaneously capture the local change patterns and short-term trends of the greenhouse outside temperature value, greenhouse inside temperature value, and fan speed value at multiple time scales. Secondly, the position encoding layer encodes absolute or relative position information into vectors and adds and fuses them with the features of the corresponding time step in the primary feature map, thereby injecting the temporal sequence information of the input data into the model. The position encoding layer adopts a position encoding scheme based on sine and cosine functions, and the generated encoding vector matches the dimension of the primary feature map, ensuring that the features of each time step are given a unique identity that can reflect its temporal position. Finally, the Transformer encoder layer receives the feature sequence fused with location information and captures long-distance contextual dependencies within the sequence layer by layer through its multi-layer self-attention mechanism and feedforward neural network. It ultimately outputs the contextual temporal feature representation that characterizes the overall state of the entire predetermined time period. The Transformer encoder layer contains residual connections and layer normalization operations to alleviate the gradient vanishing problem in deep network training and stabilize the training process, thereby ensuring that it can effectively learn and integrate complex temporal dependencies spanning multiple predetermined time points to form a high-quality contextual temporal feature representation.

[0022] In one embodiment, the temporal cross-attention module is a multi-head cross-attention mechanism. It captures diverse nonlinear correlation patterns between the temperature difference-related information and the historical state information of the fan speed from different subspaces through multiple attention heads that are computed in parallel, and concatenates and linearly transforms the outputs of all attention heads to form a unified context-aware transfer feature vector.

[0023] More specifically: the temporal cross-attention module generates the context-aware transition feature vector through the following steps: The contextual temporal features are respectively mapped through independent linear transformation layers to obtain a query vector sequence for representing the current temperature difference state, a key vector sequence for representing the historical fan speed state, and a value vector sequence. The query vector sequence, key vector sequence, and value vector sequence are input into multiple attention heads set in parallel. Each attention head independently calculates the attention weight between the query vector and all key vectors in its own subspace, and performs a weighted summation on the value vector sequence based on the attention weight, thereby generating the output feature sequence of that attention head. Within each attention head, the process of calculating the attention weights further includes: First, calculate the dot product similarity between each query vector in the query vector sequence and all key vectors in the key vector sequence; Secondly, the dot product similarity is scaled by dividing it by a preset scaling factor; Next, the scaled result is normalized using a Softmax function to obtain the final attention weight distribution; All feature sequences output by the multiple attention heads are concatenated along the feature dimension to form a concatenated feature sequence; The concatenated feature sequence is passed through a linear transformation projection layer to fuse information from different subspaces, and finally outputs a unified context-aware transfer feature vector. The output of the temporal cross-attention module is also equipped with a residual connection and layer normalization unit, which is specifically used for: The feature sequence output by the linear transformation projection layer is added to the original input of the temporal cross attention module—that is, the context temporal feature representation—by position to achieve residual connection; The feature sequence obtained after residual connection is subjected to layer normalization, and the result of layer normalization is used as the final context-aware transfer feature vector output to improve the stability and convergence speed of model training.

[0024] In one embodiment, the system includes an interpretable fuzzy inference branch that receives the context-aware transition feature vector, maps the key dimensions of the context-aware transition feature vector to predefined fuzzy variables, performs inference through a learnable or preset fuzzy rule base, and outputs an auxiliary, semantically interpretable speed adjustment suggestion.

[0025] More specifically: the interpretable fuzzy reasoning branch performs the following steps to output the auxiliary, semantically interpretable speed adjustment suggestion: First, the context-aware transfer feature vector is subjected to key dimension screening and mapping, and the values ​​of its specific dimensions are converted into the membership degrees of the corresponding fuzzy linguistic variables. The fuzzy linguistic variables include at least "temperature difference driving force" which represents the intensity of temperature difference driving and "fan response level" which represents the fan response state. The membership function of the "temperature difference driving force" fuzzy linguistic variable is a triangular or trapezoidal function defined on the aggregated value domain of the temperature difference-related dimensions in the context-aware transfer feature vector; the membership function of the "fan response level" fuzzy linguistic variable is a triangular or trapezoidal function defined on the aggregated value domain of the fan speed historical state-related dimensions in the context-aware transfer feature vector. Next, the obtained membership information is input into a preset or learnable fuzzy rule base. The fuzzy rule base contains multiple rules expressed in the form of "if-then" to describe semantic speed adjustment suggestions to be given under different combinations of "temperature difference driving force" and "fan response level". The rules in the preset or learnable fuzzy rule base are constructed in one of the following ways: predefined based on agricultural expert knowledge, or obtained by end-to-end training of the parameters of the premise and conclusion parts of the rules through neural networks or genetic algorithms using historical optimal control data and the corresponding context-aware transfer feature vectors. Then, all applicable rules in the fuzzy rule base are activated and synthesized through a fuzzy inference mechanism to calculate the confidence or membership degree of each possible semantic output suggestion (such as "significantly increase speed", "moderately decrease speed", "maintain current speed range"). The fuzzy inference mechanism adopts the Mamdani-type fuzzy inference model, uses the "AND" operation to take the minimum value of the premise membership degree as the rule activation strength, and uses the maximum value operation to aggregate the fuzzy set of each rule output in the rule synthesis stage. Finally, based on the principle of maximum membership degree or the principle of weighted average, the reasoning results of the semantic output suggestion are defuzzified to produce an auxiliary speed adjustment suggestion with clear semantic description, which is used to explain or verify the logical basis of the control command generated by the regression head.

[0026] In one embodiment, the regression head is a fully connected neural network layer, and the input of the regression head is the context-aware transfer feature vector. After the context-aware transfer feature vector undergoes nonlinear transformation and information compression through one or more fully connected layers, it is output through a linear activation function layer to generate continuous scalar values. The scalar values ​​are the feedforward control quantities, and their physical meaning is defined as fan speed adjustment amount (ΔRPM) or target speed percentage.

[0027] In one embodiment, the proportional feedback fine-tuning unit obtains the real-time feedback adjustment amount by multiplying the difference between the instantaneous temperature difference inside and outside the greenhouse at the current moment by a preset proportional coefficient, and then weights and sums the real-time feedback adjustment amount with the continuous value output by the regression head to form the final fan speed control command.

[0028] The proportional feedback fine-tuning unit obtains the latest temperature values ​​inside and outside the greenhouse at the latest moment (time t) in real time, and calculates their instantaneous difference as the current instantaneous temperature difference; Subsequently, this current instantaneous temperature difference is multiplied by a pre-set or online adaptive proportional coefficient (Kp) to generate a proportional feedback adjustment. The function of this feedback is to immediately provide a control action in the opposite direction to suppress the deviation when the instantaneous temperature difference deviates from the desired range. For example, if the instantaneous internal temperature suddenly exceeds the external temperature, the feedback will generate a command to increase the fan speed.

[0029] In one embodiment, during the training phase, the joint optimization objective of the shared noise-resistant temporal encoder and the temporal cross-attention module includes a main regression loss function and an auxiliary temporal smoothness constraint loss function, which is applied to the intermediate or output layer of the contextual temporal feature representation to further suppress high-frequency noise and enhance the temporal continuity of the feature output.

[0030] In one embodiment, it also includes: Visual monitoring and pest identification module: This module acquires crop images via cameras and performs pest detection and classification based on a ResNet60 deep learning model. More specifically, high-definition cameras are deployed inside the greenhouse to collect crop image data, and pest identification and classification are performed using a ResNet60-based deep learning model. This model employs a residual structure and feature pyramid module, supports multi-scale pest detection, achieves an accuracy rate of over 92%, and supports lightweight deployment using edge computing. Multimodal data fusion module: It fuses the pest identification results with environmental sensing data. More specifically, it fuses environmental data such as temperature, humidity, and wind speed with pest image features. Through the attention mechanism feature fusion module, it performs spatiotemporal alignment and feature integration to form a multidimensional environment-pest fusion feature vector. Intelligent decision support module: Based on fusion features and agricultural knowledge graph, it generates suggestions for pest and disease control and environmental regulation. More specifically, based on large language model and agricultural knowledge graph, combined with environmental data and pest identification results, it generates suggestions for pest and disease control and environmental regulation strategies, and pushes early warning information to farmers through voice or text.

[0031] All data obtained in this invention has been authorized by the user.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart agricultural greenhouse environment monitoring system based on the Internet of Things, characterized in that, include: Data acquisition module: acquires fan speed value, and acquires the outside temperature value and inside temperature value of the greenhouse at multiple predetermined time points within a predetermined time period through temperature sensors; Noise-resistant temporal fusion encoding module: The greenhouse outside temperature value, greenhouse inside temperature value and fan speed value at multiple predetermined time points are arranged in time sequence and input into a shared noise-resistant temporal encoder to generate a contextual temporal feature representation that fuses the noise-resistant greenhouse outside temperature information, noise-resistant greenhouse inside temperature information and fan speed information. The shared noise-resistant temporal encoder is configured to learn a temporal pattern that is insensitive to sensor noise. Temperature difference-fan nonlinear correlation modeling module: The context temporal feature representation is input to the temporal cross attention module. The temporal cross attention module dynamically calculates the nonlinear dependency between the fused temperature difference-related information as a query and the historical state information of the fan speed as a key value, and outputs a context-aware transition feature vector. The context-aware transition feature vector encodes the complex dynamic correlation between temperature difference and fan speed. Continuous fan speed control module: The context-aware transfer feature vector is input to the regression head, and the regression head outputs a continuous value representing the fan speed adjustment amount or the target speed. The continuous fan speed control module includes a proportional feedback fine-tuning unit based on the current instantaneous temperature difference, which is used to fine-tune the continuous value output by the regression head in real time and generate the final fan speed control command.

2. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The specific steps for the data acquisition module to obtain the fan speed value, the outside temperature value, and the inside temperature value of the greenhouse are as follows: A first temperature sensor and a second temperature sensor are deployed at the shaded and ventilated area outside the greenhouse and at the height of the crop canopy inside the greenhouse, respectively. A speed sensor is installed on the ventilation fan motor of the greenhouse. An edge computing controller is configured to communicate with the first temperature sensor, the second temperature sensor, and the speed sensor. In response to an environmental control cycle command or the arrival of a predetermined time interval, the edge computing controller initiates a new data acquisition task for a predetermined time period. During the predetermined time period, the edge computing controller synchronously reads and records the outside temperature value of the greenhouse measured by the first temperature sensor, the inside temperature value of the greenhouse measured by the second temperature sensor, and the fan speed value measured by the speed sensor at multiple predetermined time points, forming three sets of time-aligned original time-series data sequences. The three sets of original time-series data sequences are cached locally and basic data cleaning, including missing value imputation and obvious outlier removal, is performed to generate a standardized time-series dataset to be processed.

3. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The shared noise-resistant temporal encoder includes a one-dimensional convolutional layer, a positional encoding layer, and a Transformer encoder layer connected in sequence. The one-dimensional convolutional layer is used to extract local features and perform preliminary smoothing filtering on the input temporal data. The positional encoding layer is used to inject sequential information into the temporal data. The Transformer encoder layer learns long-term temporal dependencies and generates the contextual temporal feature representation through a self-attention mechanism.

4. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The temporal cross-attention module is a multi-head cross-attention mechanism. It captures diverse nonlinear correlation patterns between the temperature difference-related information and the historical state information of the fan speed from different subspaces through multiple attention heads that are computed in parallel. The outputs of all attention heads are concatenated and linearly transformed to form a unified context-aware transfer feature vector.

5. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The system includes an interpretable fuzzy inference branch, which receives the context-aware transition feature vector, maps the key dimensions of the context-aware transition feature vector to predefined fuzzy variables, performs inference through a fuzzy rule base, and outputs speed adjustment suggestions.

6. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The regression head is a fully connected neural network layer. The input of the regression head is the context-aware transfer feature vector. After the context-aware transfer feature vector undergoes nonlinear transformation and information compression through one or more fully connected layers, it is output through a linear activation function layer to generate continuous scalar values, which are the feedforward control quantities.

7. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, The proportional feedback fine-tuning unit obtains the real-time feedback adjustment amount by multiplying the difference between the instantaneous temperature difference inside and outside the greenhouse at the current moment by a preset proportional coefficient, and then weights and sums the real-time feedback adjustment amount with the continuous value output by the regression head to form the final fan speed control command.

8. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, During the training phase, the joint optimization objective of the shared noise-resistant temporal encoder and the temporal cross-attention module includes a main regression loss function and an auxiliary temporal smoothness constraint loss function, which is applied to the intermediate layer or output layer of the context temporal feature representation.

9. The smart agricultural greenhouse environment monitoring system based on the Internet of Things according to claim 1, characterized in that, Also includes: Visual monitoring and pest identification module: Captures crop images through a camera and performs pest detection and classification based on the ResNet60 deep learning model; Multimodal data fusion module: fuses pest identification results with environmental sensor data for feature fusion; Intelligent decision support module: Based on fusion features and agricultural knowledge graph, it generates suggestions for pest and disease control and environmental regulation.

Images