Intelligent management method and system for edible mushroom planting environment based on big data

Abstract

This invention discloses an intelligent management method and system for edible mushroom cultivation environment based on big data. It acquires sensor data, 3D image data, and growth data, preprocesses the collected data to obtain multimodal data, establishes a multimodal feature extraction network model, and uses an LSTM (Long Short-Term Memory) network to extract the time-series features of the sensors in the multimodal data. The model performs semantic segmentation and volume calculation on point cloud data and encodes the growth data. Multi-source features are weighted and fused using a cross-modal attention mechanism to generate associated feature data. Based on this associated feature data, an optimal environmental parameter adjustment strategy is output using a reinforcement learning (RL) framework, with maximizing yield and quality as the reward function. According to the optimal environmental parameter adjustment strategy, the system controls the air conditioning, humidifier, supplemental lighting, and ventilation equipment. This significantly improves management accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method and system for intelligent management of edible mushroom cultivation environments based on big data. Background Technology

[0002] In the field of large-scale edible fungi cultivation, traditional management models have long relied on manual experience, resulting in a single dimension of environmental monitoring, mainly collecting basic data such as temperature and humidity, and lacking multimodal data linkage such as three-dimensional spatial morphology and full-cycle growth status. Environmental control is lagging and extensive, and is prone to fluctuations in mushroom yield and inconsistent quality due to parameter imbalances. At the same time, there are also problems such as high energy consumption and untimely disease control, making the environmental management of edible fungi cultivation inefficient and difficult to adapt to the needs of modern precision cultivation. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an intelligent management method and system for edible fungi cultivation environment based on big data.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned intelligent management method for edible fungi cultivation environment based on big data: Acquire sensor data, 3D image data, and growth data; preprocess the acquired data to obtain multimodal data; A multimodal feature extraction network model is established, and the time-series features of the sensors in the multimodal data are extracted using an LSTM (Long Short-Term Memory) network. The point cloud data is semantically segmented and its volume is calculated using a PointNet++ model, and the grown data is One-Hot encoded. Multi-source features are weighted and fused through a cross-modal attention mechanism to generate associated feature data. Based on the aforementioned associated feature data, an optimal environmental parameter adjustment strategy is output using a reinforcement learning (RL) framework, with the reward function being the maximization of output and quality. The system controls the air conditioner, humidifier, supplementary lighting, and ventilation equipment based on the optimal environmental parameter adjustment strategy.

[0005] Furthermore, in the aforementioned intelligent management method for edible fungi cultivation environment based on big data, the acquisition of sensor data, three-dimensional image data, and growth data, and the preprocessing of the collected data to obtain multimodal data, include: Acquire sensor data, 3D image data, and growth data. The sensor data should include at least temperature and humidity data, carbon dioxide concentration data, light intensity data, and oxygen concentration data. The three-dimensional image data includes the spatial distribution of the mushrooms, the outline morphology of the cap, and the growth posture of the stipe; the growth data includes mycelial development data, primordium formation data, fruiting body growth data, and maturity data. The 3σ principle is used to identify and remove outliers in sensor data, and the missing data after removal is filled by linear interpolation of data from adjacent time periods. The moving average method is used to smooth the time series data of the sensor, eliminate high-frequency noise, and map sensor data of different dimensions to the [0,1] interval to obtain preprocessed time series data.

[0006] Furthermore, in the aforementioned intelligent management method for edible fungi cultivation environment based on big data, the acquisition of sensor data, three-dimensional image data, and growth data, and the preprocessing of the collected data to obtain multimodal data, include: Isolated noise points in the 3D point cloud data were removed using a statistical filtering algorithm, while retaining the effective point cloud data of the mushrooms and the matrix. Point cloud data from different perspectives collected by multiple scanners were registered and stitched together to obtain a complete three-dimensional point cloud model of the planting area. The image data was then subjected to contrast enhancement and deblurring to obtain preprocessed image data. The growth data of cap diameter and stipe height were normalized to unify the data scale, resulting in preprocessed growth data. Based on the acquisition timestamp, preprocessed time series data, preprocessed image data, and preprocessed growth data are correlated to obtain multimodal data.

[0007] Furthermore, in the aforementioned intelligent management method for edible fungus cultivation environment based on big data, the step of establishing a multimodal feature extraction network model and using an LSTM (Long Short-Term Memory) network to extract the time-series features of sensors in the multimodal data includes: The time series samples are input into a pre-trained LSTM (Long Short-Term Memory) network. The network captures the long-term dependencies and short-term fluctuation characteristics of sensor data through layer-by-layer gating units, and outputs a time series feature vector that can characterize the time series change pattern of the environment.

[0008] Furthermore, in the aforementioned intelligent management method for edible fungi cultivation environment based on big data, the step of performing semantic segmentation and volume calculation on point cloud data using the PointNet++ model includes: The preprocessed 3D point cloud data is input into the PointNet++ model. The model divides the point cloud into multiple local regions through hierarchical sampling and local feature aggregation, then extracts the local features of each region and merges them into global features step by step. The semantic segmentation branch is used to classify and identify the mushroom fruiting bodies, mycelium, substrate, and planting rack in the point cloud data, and to separate the mushroom region from the background region to obtain the independent point cloud region of the mushroom. Based on the independent point cloud region of mushrooms, the point cloud is transformed into a three-dimensional voxel mesh through spatial voxelization. The number of voxels covered by the mushroom point cloud is counted, and the actual volume of a single mushroom and a group of mushrooms is calculated by combining the voxel resolution, thus obtaining the spatial morphological feature vector of mushrooms.

[0009] Furthermore, in the aforementioned intelligent management method for edible fungi cultivation environment based on big data, the step of performing One-Hot encoding on growth data and generating associated feature data through weighted fusion of multi-source features via a cross-modal attention mechanism includes: The correlation weights between each modality feature and the yield and quality of edible fungi are calculated using an attention mechanism. The feature vectors of the three modalities are weighted and summed based on the weight coefficients output by the attention mechanism. The fused features are then dimension-mapped and integrated through a fully connected layer to generate associated feature data.

[0010] Furthermore, in the aforementioned intelligent management method for edible fungi cultivation environment based on big data, the step of using the associated feature data and a reinforcement learning framework (RL) to output an optimal environmental parameter adjustment strategy with maximizing yield and quality as the reward function includes: The temporal feature vector, the spatial morphological feature vector of mushrooms, and the growth state feature vector are defined as the state space; the set of executable actions of the intelligent agent is defined as the action space, and the parameter adjustment range of each environmental control device is defined as the action space. A deep Q-network (DQN) is used as the core algorithm for reinforcement learning. Associated feature data is input into the network, and the agent explores actions through an ε-greedy policy. Real-time collection of greenhouse environment data and mushroom growth data, updating of environmental status and calculation of reward value, outputting the Q value of each action, and selecting the action combination with the highest Q value as the optimal environmental parameter adjustment strategy.

[0011] Furthermore, in the intelligent management system for edible mushroom cultivation environment based on big data, the intelligent management system for edible mushroom cultivation environment based on big data includes the following modules: The multimodal data acquisition module is used to acquire sensor data, 3D image data, and growth data, and preprocess the acquired data to obtain multimodal data. The data feature association module is used to establish a multimodal feature extraction network model. It uses an LSTM (Long Short-Term Memory) network to extract the time-series features of the sensors in the multimodal data, performs semantic segmentation and volume calculation on the point cloud data using a PointNet++ model, and performs One-Hot encoding on the grown data. It also generates associated feature data by weighted fusion of multi-source features through a cross-modal attention mechanism. The environmental parameter adjustment module is used to output the optimal environmental parameter adjustment strategy based on the associated feature data and through a reinforcement learning framework using RL, with the reward function being the maximization of output and quality. The environmental intelligent management module is used to adjust strategies based on the optimal environmental parameters and control the air conditioning, humidifier, supplementary lighting, and ventilation equipment in the system.

[0012] Furthermore, in the intelligent management system for edible fungi cultivation environment based on big data, the data feature association module includes the following sub-modules: The mapping submodule is used to calculate the correlation weights between each modal feature and the yield and quality of edible fungi through an attention mechanism. Based on the weight coefficients output by the attention mechanism, the feature vectors of the three modalities are weighted and summed. The fused features are then dimensionally mapped and integrated through a fully connected layer to generate associated feature data.

[0013] Furthermore, in the intelligent management system for edible fungi cultivation environment based on big data, the environmental parameter adjustment module includes the following sub-modules: Define a submodule to define the temporal feature vector, the spatial morphological feature vector of mushrooms, and the growth state feature vector as the state space; set the set of executable actions for the intelligent agent, and define the parameter adjustment range of each environmental control device as the action space. The exploration submodule employs a Deep Q-Network (DQN) as the core algorithm for reinforcement learning, inputting associated feature data into the network, and the agent explores actions using an ε-greedy strategy; The output submodule is used to collect greenhouse environmental data and mushroom growth data in real time, update the environmental status and calculate the reward value, output the Q value of each action, and select the action combination with the highest Q value as the optimal environmental parameter adjustment strategy.

[0014] Its beneficial effects lie in the construction of a comprehensive monitoring network through multiple types of sensors and a 3D vision system, forming a multimodal dataset through standardized preprocessing, providing reliable data support for decision-making. Relying on LSTM, PointNet++, and cross-modal attention mechanisms, it achieves deep fusion of multi-dimensional features, accurately establishing a correlation model between environment and growth. Based on reinforcement learning, it generates optimal control strategies, balancing yield, quality, and energy consumption, avoiding resource waste driven by a single objective. Fourthly, it constructs a closed-loop control system, realizing coordinated equipment execution and dynamic strategy optimization, significantly improving management accuracy and efficiency, reducing reliance on manual labor, and steadily increasing the yield of high-quality mushrooms. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the intelligent management method for edible fungi cultivation environment based on big data in this invention. Figure 2 This is a schematic diagram of the second embodiment of the intelligent management method for edible fungi cultivation environment based on big data in this invention. Figure 3 This is a schematic diagram of the first embodiment of the intelligent management system for edible fungi cultivation environment based on big data in this invention. Detailed Implementation

[0017] 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 of the invention and are not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a smart management method for edible mushroom cultivation environment based on big data is proposed. This method includes the following steps: Step 101: Acquire sensor data, 3D image data, and growth data; preprocess the acquired data to obtain multimodal data; Specifically, in this embodiment, sensor data, three-dimensional image data, and growth data are acquired. The sensor data includes at least temperature and humidity data, carbon dioxide concentration data, light intensity data, and oxygen concentration data. The three-dimensional image data includes the spatial distribution of the mushroom, the outline of the cap, and the growth posture of the stipe. The growth data includes mycelial development data, primordium formation data, fruiting body growth data, and maturity data. The 3σ principle is used to identify and remove outliers in the sensor data, and the missing data after removal is filled by linear interpolation of adjacent time periods. The moving average method is used to smooth the time series data of the sensor, eliminate high-frequency noise, and map sensor data of different dimensions to the [0,1] interval to obtain preprocessed time series data. Isolated noise points in the 3D point cloud data were removed using a statistical filtering algorithm, retaining the effective point cloud data of the mushrooms and the substrate. Point cloud data from different perspectives collected by multiple scanners were registered and stitched together to obtain a complete 3D point cloud model of the planting area. The image data was then subjected to contrast enhancement and deblurring to obtain preprocessed image data. The growth data of cap diameter and stipe height were normalized to unify the data scale, resulting in preprocessed growth data. Based on the collection timestamp, the preprocessed time series data, preprocessed image data, and preprocessed growth data were correlated to obtain multimodal data.

[0020] The core objective of the multi-source data acquisition and preprocessing stage is to obtain comprehensive and accurate basic data on the planting environment and edible fungi growth. Standardized preprocessing is used to eliminate data noise and unify data formats, laying a reliable data foundation for subsequent feature extraction and intelligent decision-making. This stage is mainly divided into two sub-modules: data acquisition and data preprocessing. During the data acquisition phase, sensor data collection is the core of environmental monitoring. To achieve full-dimensional environmental coverage of the cultivation shed, a multi-type, high-density sensor network needs to be constructed: temperature and humidity sensors are arranged at a density of 5 meters per sensor at different heights of the cultivation racks and at the four corners of the shed, collecting air and substrate temperature and humidity data every 10 minutes to ensure the capture of temperature and humidity differences in different areas; carbon dioxide concentration sensors are strategically installed in the middle of the shed and near the ventilation openings, collecting data every 15 minutes to accurately monitor concentration changes in key areas of gas circulation; light intensity sensors are suspended below the supplemental lighting and above the cultivation layer, with a collection frequency of once every 30 minutes to match the adjustment cycle of the supplemental lighting; substrate pH and moisture content sensors are buried 5 cm deep in the substrate, collecting data every hour to ensure real-time monitoring of key indicators of the mushroom growth substrate; oxygen concentration sensors are mainly deployed at the bottom of the shed and in areas with dense mushroom bags, collecting data every 20 minutes to promptly detect the risk of oxygen deficiency. All sensor data is synchronously transmitted to the local terminal and cloud database through an industrial IoT gateway, achieving dual data backup and avoiding data loss due to equipment failure. The 3D vision system focuses on the spatial morphology of the mushrooms. Multiple high-precision 3D laser scanners and RGB-D depth cameras are installed on the top of the cultivation shed. Scannings are conducted three times a day (morning, noon, and evening) to cover the different growth stages of the mushrooms. The laser scanners are responsible for constructing a 3D point cloud model of the mushrooms and substrate, while the depth cameras capture images that fuse color and depth, fully capturing key visual features such as cap outlines, stipe postures, and population distribution. Growth data collection adopts a combination of manual sampling and intelligent recognition. Technicians randomly select 5% of the mushroom bags daily to measure core indicators such as cap diameter and stipe height. At the same time, image recognition algorithms are used to analyze the 3D vision images in batches to supplement population growth data. Data on growth stages, disease conditions, and other categories are also recorded simultaneously to form a complete growth log. In the preprocessing stage, sensor data first needs to be filtered for outliers using the 3σ principle, then filled with gaps using linear interpolation, smoothed for noise using the moving average method, and finally normalized to the [0,1] interval to unify the dimensions. 3D vision data needs to undergo statistical filtering for noise reduction, multi-view registration and stitching, and image enhancement processing to eliminate scanning errors and blind spots. Growth data needs to be standardized for numerical data and cross-verified and corrected by manual data. Finally, based on the timestamp, the three types of preprocessed data are associated and bound to generate a standardized multimodal dataset, which is stored in the cloud to provide high-quality data support for subsequent stages.

[0021] Step 102: Establish a multimodal feature extraction network model, use an LSTM (Long Short-Term Memory) network to extract time-series features from sensors in multimodal data, perform semantic segmentation and volume calculation on point cloud data using the PointNet++ model, and perform One-Hot encoding on the grown data; use a cross-modal attention mechanism to weightedly fuse multi-source features to generate associated feature data; Specifically, in this embodiment, time-series samples are input into a pre-trained LSTM (Long Short-Term Memory) network. The network captures the long-term dependencies and short-term fluctuations of sensor data through layer-by-layer gating units, outputting a time-series feature vector that characterizes the temporal changes in the environment. The pre-processed 3D point cloud data is input into the PointNet++ model. The model divides the point cloud into multiple local regions through hierarchical sampling and local feature aggregation, then extracts the local features of each region and merges them into global features. The semantic segmentation branch classifies and identifies the mushroom fruiting bodies, mycelium, substrate, and planting racks in the point cloud data, separating the mushroom region from the background region to obtain independent mushroom point cloud regions. Based on the independent mushroom point cloud regions, the point cloud is transformed into a 3D voxel mesh using a spatial voxelization method. The number of voxels covered by the mushroom point cloud is counted, and the actual volume of a single mushroom and a group of mushrooms is calculated based on the voxel resolution, resulting in a mushroom spatial morphology feature vector. The correlation weights between each modality feature and the yield and quality of edible fungi are calculated using an attention mechanism. The feature vectors of the three modalities are weighted and summed based on the weight coefficients output by the attention mechanism. The fused features are then dimension-mapped and integrated through a fully connected layer to generate associated feature data.

[0022] The core of the multimodal feature extraction and fusion process is to extract core features from different types of raw data and establish a deep correlation between the environment and mushroom growth through cross-modal technology, forming associated feature data that can support intelligent decision-making. This process is divided into two key sub-processes: single-modal feature extraction and cross-modal feature fusion. In the single-modal feature extraction stage, feature mining of sensor time-series data is achieved using an LSTM (Long Short-Term Memory) network. First, the preprocessed sensor data is reconstructed into time-series windows on a one-day basis. Each window contains 144 temperature and humidity data points, 96 carbon dioxide data points, and other multi-dimensional time-series samples. Then, the samples are input into a pre-trained LSTM network. The gating unit composed of the input gate, forget gate, and output gate can accurately capture the long-term dependence and short-term fluctuations of the data. For example, it can identify the inhibitory effect of three consecutive days of low temperature and humidity on mycelial development and capture the instantaneous change of carbon dioxide during ventilation periods. Finally, it outputs a high-dimensional feature vector representing the temporal patterns of the environment. Feature extraction of the 3D visual point cloud data is accomplished by the PointNet++ model. This model first divides the point cloud into multiple local regions through hierarchical sampling, then aggregates local features layer by layer to form global features. Simultaneously, a semantic segmentation branch is activated to accurately classify objects such as mushroom fruiting bodies, mycelium, substrate, and planting racks, achieving separation of mushroom regions from the background. Based on this, the mushroom point cloud is transformed into a 3D mesh through spatial voxelization, the number of voxels is counted, and the volume of a single mushroom and the group is calculated in combination with the resolution to generate a spatial morphological feature vector. Feature extraction of growth data requires differentiated processing based on data type. One-Hot encoding is used for categorical data such as growth stage and disease level. For example, the four growth stages are converted into one-hot vectors, and then the encoded categorical features are concatenated with normalized numerical growth features to form a complete growth feature vector. After completing the extraction of single-modal features, multi-source feature fusion needs to be achieved through a cross-modal attention mechanism. First, a cross-modal attention network is constructed, and the three types of single-modal features are input into the network. The network dynamically allocates weights according to the growth stage of the mushroom. For example, during the fruiting body growth stage, the weights of light time sequence and cap volume features are increased, while during the mycelial development stage, the proportions of substrate temperature and humidity and pH value features are strengthened. Then, the three types of feature vectors are weighted and summed according to the weight coefficients, and the dimensions are integrated through a fully connected layer. Finally, associated feature data that can simultaneously represent environmental changes, spatial morphology, and growth status are generated, building a core feature bridge for reinforcement learning decision-making.

[0023] Data temporal reconstruction: Preprocessed sensor data is divided into fixed-length temporal windows according to the time dimension to construct sensor temporal data samples. LSTM network training and feature extraction: The temporal samples are input into a pre-trained LSTM (Long Short-Term Memory) network. The network captures long-term dependencies (such as the impact of three consecutive days of temperature and humidity changes on mycelial growth) and short-term fluctuation characteristics (such as a sudden drop in carbon dioxide concentration during ventilation) of sensor data through layer-by-layer gating units (input gate, forget gate, output gate). Finally, it outputs a high-dimensional sensor temporal feature vector that characterizes the temporal changes in the environment.

[0024] Point cloud semantic segmentation: The preprocessed 3D point cloud data is input into the PointNet++ model. The model first divides the point cloud into multiple local regions through hierarchical sampling and local feature aggregation, then extracts the local features of each region and merges them into global features step by step. At the same time, through the semantic segmentation branch, different objects such as mushroom fruiting bodies, mycelium, substrate, and planting racks in the point cloud data are classified and identified to achieve accurate separation of mushroom areas from background areas.

[0025] Mushroom volume calculation: Based on the independent point cloud regions of mushrooms obtained by semantic segmentation, the point cloud is transformed into a three-dimensional voxel mesh through spatial voxelization. The number of voxels covered by the mushroom point cloud is counted, and the actual volume of a single mushroom and a group of mushrooms is calculated by combining the voxel resolution, forming a spatial morphological feature vector of mushrooms.

[0026] For categorical data (such as growth stage and disease severity) in growth data, One-Hot encoding is performed: for example, the growth stage is divided into four categories: "mycelial development stage", "primordia formation stage", "fruiting body growth stage" and "maturity stage", and encoded as feature vectors of [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1], respectively; the disease severity is divided into four categories: "no disease", "mild disease", "moderate disease" and "severe disease", and corresponding one-hot encoded vectors are generated.

[0027] The encoded categorical features are concatenated with normalized numerical growth features (such as cap diameter and single mushroom weight) to form a complete growth feature vector.

[0028] A cross-modal attention network was constructed, using temporal features of sensors, three-dimensional visual spatial features, and growth features as network inputs. The correlation weights between each modal feature and the yield and quality of edible fungi were calculated through an attention mechanism. For example, when the fruiting body is in the growth stage, the temporal features of light intensity and the spatial features of cap volume have a greater impact on the yield, and the network will automatically assign higher weights to these two types of features. When the mycelium is in the development stage, the temporal features of substrate temperature and humidity and pH value will have significantly increased weights.

[0029] Based on the weight coefficients output by the attention mechanism, the feature vectors of the three modalities are weighted and summed. Then, the fused features are dimension-mapped and integrated through a fully connected layer, ultimately generating associated feature data that can simultaneously represent environmental changes, spatial morphology, and growth status, providing core feature support for subsequent reinforcement learning decisions.

[0030] Step 103: Based on the associated feature data, using a reinforcement learning framework (RL) with the reward function of maximizing output and quality, output the optimal environmental parameter adjustment strategy; Specifically, in this embodiment, the temporal feature vector, the mushroom spatial morphology feature vector, and the growth state feature vector are defined as the state space; the set of executable actions of the agent is set, and the parameter adjustment range of each environmental control device is defined as the action space; the DQN deep Q network is used as the core algorithm of reinforcement learning, the associated feature data is input into the network, and the agent explores actions through an ε-greedy strategy; real-time collection of greenhouse environmental data and mushroom growth data, updating of environmental state and calculation of reward value, output of Q value of each action, and selection of the action combination with the highest Q value as the optimal environmental parameter adjustment strategy.

[0031] The core of the optimal environmental parameter adjustment strategy generation stage based on reinforcement learning is to leverage a reinforcement learning framework and, based on associated feature data, generate an optimal control scheme that balances yield, quality, and energy consumption through continuous policy iteration. This stage encompasses two main sub-modules: reinforcement learning framework construction and strategy training and optimization. In the framework construction phase, the definitions of the agent and the environment must first be clarified. The environmental parameter decision module is defined as the agent, whose core task is to output precise control commands based on associated feature data. The entire cultivation greenhouse and all mushroom cultivation objects are defined as the reinforcement learning environment, and the real-time state of the environment is fully represented by the associated feature data. Next, the state space and action space need to be scientifically defined. The state space directly uses high-dimensional vectors of the associated feature data, encompassing all key information about the environment, morphology, and growth. The action space focuses on the adjustment range of various control devices, specifically divided into three levels: air conditioning temperature ±0.5℃ to ±1.5℃, humidifier humidity ±3%RH to ±8%RH, supplementary lighting ±100lux to ±300lux, and ventilation equipment low, medium, and high. It also includes combined actions for starting and stopping equipment, ensuring the agent has sufficient control dimensions. The design of the reward function is the core of the framework. A multi-objective reward system needs to be constructed. The output reward is set based on the ratio of the total output per unit period to the historical best output, with the excess output being weighted proportionally. The quality reward is based on the proportion of high-quality mushrooms. The higher the quality rate, the higher the reward value. If disease occurs, the corresponding reward will be deducted. The energy consumption penalty is set with a penalty coefficient for energy consumption exceeding the threshold. The comprehensive reward value is the weighted sum of the output and quality rewards minus the energy consumption penalty, so as to achieve a balanced control of the three. In the strategy training and optimization phase, a deep Q-network (DQN) is used as the core algorithm. Initially, the agent executes an ε-greedy policy with an ε value of 0.9, prioritizing random exploration of different control actions to cover more environmental scenarios. As the training rounds increase, the ε value is reduced to 0.1, and the agent prioritizes historically optimal actions. After each action is executed, environmental and growth data are collected in real time to update the state and calculate the reward. The experience of state-action-reward-new state is stored in the replay pool, and batches of experience are periodically extracted to train the network to update parameters and optimize the Q value. When the fluctuation of the comprehensive reward value is less than 5% for 100 consecutive rounds, the network is considered to have converged and the parameters are fixed. At this time, real-time associated features are input into the network, and the action combination with the highest Q value is selected as the optimal policy. This clarifies the adjustment range and start / stop status of each device, providing a decision-making basis for precise control.

[0032] Step 104: Adjust the strategy according to the optimal environmental parameters and control the air conditioner, humidifier, supplementary lighting and ventilation equipment in the system.

[0033] Specifically, in the link of the linkage control of the planting environment equipment based on the optimal strategy in this embodiment, the core is to transform the abstract strategy generated by reinforcement learning into control instructions executable by the equipment, achieve precise environmental regulation, and build a closed-loop optimization system. This link is divided into two sub-links: the parsing and execution of the regulation instructions and the feedback optimization. In the stage of parsing and executing the regulation instructions, first, it is necessary to build a device control instruction parsing module. The core of this module is to complete the conversion from the policy parameters to the device signals. For example, the instruction to reduce the temperature by 1°C is converted into the frequency adjustment signal of the air conditioner compressor, the instruction to increase the humidity by 5%RH is converted into the atomization amount control signal of the humidifier, the instruction to adjust the light is converted into the power signal of the supplementary light, and the ventilation requirement is converted into the rotation speed signal of the exhaust fan, ensuring that the abstract strategy is implemented as specific device operation parameters. The linkage execution of the devices needs to follow the principle of priority sorting + synchronous linkage. For example, when there are both high temperature and excessive carbon dioxide in the greenhouse, the air conditioner will be started first to cool down, and at the same time, the ventilation equipment will be linked to discharge the high-concentration carbon dioxide, avoiding the regulation imbalance caused by the operation of a single device. If the light needs to be increased during the growth period of the sub-entity, the power and duration of the supplementary light will be adjusted synchronously, and the temperature will be monitored in real time. Once the temperature rises due to the heat generated by the supplementary light, local ventilation will be immediately started to ensure the environmental stability. All control signals are finally sent to the device execution terminal through the industrial PLC programmable logic controller. The high reliability of the PLC can ensure that the device operates precisely according to the instructions. At the same time, the system will automatically record the start and stop time, operation parameters and other logs of the device, retaining data for subsequent energy consumption analysis and strategy optimization. In the stage of execution feedback and closed-loop optimization, after the device completes the regulation action, the sensor network and the three-dimensional vision system will continuously monitor at the original frequency, and transmit the data such as the change of environmental parameters and the growth state of the mushrooms to the cloud data warehouse in real time, forming a complete feedback chain of the regulation effect. After the feedback data is preprocessed and feature extracted, it will be re-input into the reinforcement learning model. The model will fine-tune the decision-making strategy by combining the new state data and the actual reward value. For example, if the high-quality rate of the mushrooms does not meet the expectation after a certain regulation, the model will re-optimize the feature weights and action selection logic, continuously improving the regulation accuracy. Finally, a complete closed-loop of data collection - feature fusion - strategy decision - device execution - effect feedback - strategy optimization is formed, realizing the full-process intelligent management of the edible mushroom planting environment.

[0034] The beneficial effects are as follows: By constructing a comprehensive monitoring network with multiple types of sensors and a three-dimensional vision system, a multi-modal data set is formed through standardized preprocessing, providing reliable data support for decision-making. Relying on LSTM, PointNet++, and cross-modal attention mechanisms, deep fusion of multi-dimensional features is achieved, and an accurate correlation model between the environment and growth is established. Based on reinforcement learning, an optimal regulation strategy is generated, taking into account the balance of yield, quality, and energy consumption, and avoiding resource waste caused by a single target orientation. Fourth, a closed-loop regulation system is constructed to realize the linkage execution of devices and the dynamic optimization of strategies, greatly improving the management accuracy and efficiency, reducing the dependence on manual labor, and steadily increasing the output rate of high-quality mushrooms.

[0035] Please see Figure 2 In the intelligent management method for edible fungi cultivation environment based on big data, sensor data, 3D image data, and growth data are acquired. The collected data is then preprocessed to obtain multimodal data, including the following steps: Step 201: Remove isolated noise points from the 3D point cloud data using a statistical filtering algorithm, retaining the effective point clouds of the mushrooms and the matrix; Step 202: Register and stitch together point cloud data from different perspectives collected by multiple scanners to obtain a complete 3D point cloud model of the planting area. Perform contrast enhancement and deblurring on the image data to obtain preprocessed image data. Step 203: Normalize the growth data of cap diameter and stipe height to unify the data scale and obtain preprocessed growth data; Step 204: Based on the acquisition timestamp, correlate the preprocessed time series data, preprocessed image data, and preprocessed growth data to obtain multimodal data.

[0036] The above describes embodiments of the intelligent management method for edible fungi cultivation environment based on big data according to the present invention. Please refer to [link / reference]. Figure 3 The intelligent management system for edible mushroom cultivation environment based on big data includes the following modules: The multimodal data acquisition module is used to acquire sensor data, 3D image data, and growth data, and preprocess the acquired data to obtain multimodal data. The data feature association module is used to establish a multimodal feature extraction network model. It uses an LSTM (Long Short-Term Memory) network to extract time-series features from sensors in multimodal data, performs semantic segmentation and volume calculation on point cloud data using a PointNet++ model, and performs One-Hot encoding on the grown data. It also generates associated feature data by weighted fusion of multi-source features through a cross-modal attention mechanism. The environmental parameter adjustment module is used to output the optimal environmental parameter adjustment strategy based on associated feature data and through a reinforcement learning framework (RL) with the reward function being the maximization of output and quality. The intelligent environmental management module is used to adjust strategies based on optimal environmental parameters and control the air conditioning, humidifier, supplemental lighting, and ventilation equipment in the system.

[0037] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A method for intelligent management of edible fungi cultivation environment based on big data, characterized in that... The intelligent management method for edible fungi cultivation environment based on big data includes the following steps: Acquire sensor data, 3D image data, and growth data; preprocess the acquired data to obtain multimodal data; A multimodal feature extraction network model is established, and the time-series features of the sensors in the multimodal data are extracted using an LSTM (Long Short-Term Memory) network. The point cloud data is semantically segmented and its volume is calculated using a PointNet++ model, and the grown data is One-Hot encoded. Multi-source features are weighted and fused through a cross-modal attention mechanism to generate associated feature data. Based on the aforementioned associated feature data, an optimal environmental parameter adjustment strategy is output using a reinforcement learning (RL) framework, with the reward function being the maximization of output and quality. The system controls the air conditioner, humidifier, supplementary lighting, and ventilation equipment based on the optimal environmental parameter adjustment strategy.

2. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... The acquisition of sensor data, 3D image data, and growth data, followed by preprocessing of the acquired data to obtain multimodal data, includes: Acquire sensor data, 3D image data, and growth data. The sensor data should include at least temperature and humidity data, carbon dioxide concentration data, light intensity data, and oxygen concentration data. The three-dimensional image data includes the spatial distribution of the mushrooms, the outline morphology of the cap, and the growth posture of the stipe; the growth data includes mycelial development data, primordium formation data, fruiting body growth data, and maturity data. The 3σ principle is used to identify and remove outliers in sensor data, and the missing data after removal is filled by linear interpolation of data from adjacent time periods. The moving average method is used to smooth the time series data of the sensor, eliminate high-frequency noise, and map sensor data of different dimensions to the [0,1] interval to obtain preprocessed time series data.

3. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... The acquisition of sensor data, 3D image data, and growth data, followed by preprocessing of the acquired data to obtain multimodal data, includes: Isolated noise points in the 3D point cloud data were removed using a statistical filtering algorithm, while retaining the effective point cloud data of the mushrooms and the matrix. Point cloud data from different perspectives collected by multiple scanners were registered and stitched together to obtain a complete three-dimensional point cloud model of the planting area. The image data was then subjected to contrast enhancement and deblurring to obtain preprocessed image data. The growth data of cap diameter and stipe height were normalized to unify the data scale, resulting in preprocessed growth data. Based on the acquisition timestamp, preprocessed time series data, preprocessed image data, and preprocessed growth data are correlated to obtain multimodal data.

4. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... The establishment of a multimodal feature extraction network model, utilizing an LSTM (Long Short-Term Memory) network to extract time-series features of the sensors from the multimodal data, includes: The time series samples are input into a pre-trained LSTM (Long Short-Term Memory) network. The network captures the long-term dependencies and short-term fluctuation characteristics of sensor data through layer-by-layer gating units, and outputs a time series feature vector that can characterize the time series change pattern of the environment.

5. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... The step of performing semantic segmentation and volume calculation on point cloud data using the PointNet++ model includes: The preprocessed 3D point cloud data is input into the PointNet++ model. The model divides the point cloud into multiple local regions through hierarchical sampling and local feature aggregation, then extracts the local features of each region and merges them into global features step by step. The semantic segmentation branch is used to classify and identify the mushroom fruiting bodies, mycelium, substrate, and planting rack in the point cloud data, and to separate the mushroom region from the background region to obtain the independent point cloud region of the mushroom. Based on the independent point cloud region of mushrooms, the point cloud is transformed into a three-dimensional voxel mesh through spatial voxelization. The number of voxels covered by the mushroom point cloud is counted, and the actual volume of a single mushroom and a group of mushrooms is calculated by combining the voxel resolution, thus obtaining the spatial morphological feature vector of mushrooms.

6. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... The growth data is One-Hot encoded; multi-source features are weighted and fused through a cross-modal attention mechanism to generate associated feature data, including: The correlation weights between each modality feature and the yield and quality of edible fungi are calculated using an attention mechanism. The feature vectors of the three modalities are weighted and summed based on the weight coefficients output by the attention mechanism. The fused features are then dimension-mapped and integrated through a fully connected layer to generate associated feature data.

7. The intelligent management method for edible fungi cultivation environment based on big data as described in claim 1, characterized in that... Based on the associated feature data, and using a reinforcement learning (RL) framework, the optimal environmental parameter adjustment strategy is output with maximizing output and quality as the reward function. This includes: The temporal feature vector, the spatial morphological feature vector of mushrooms, and the growth state feature vector are defined as the state space; the set of executable actions of the intelligent agent is defined as the action space, and the parameter adjustment range of each environmental control device is defined as the action space. A deep Q-network (DQN) is used as the core algorithm for reinforcement learning. Associated feature data is input into the network, and the agent explores actions through an ε-greedy policy. Real-time collection of greenhouse environment data and mushroom growth data, updating of environmental status and calculation of reward value, outputting the Q value of each action, and selecting the action combination with the highest Q value as the optimal environmental parameter adjustment strategy.

8. A smart management system for edible fungi cultivation environment based on big data, characterized in that... The intelligent management system for edible fungi cultivation environment based on big data includes the following modules: The multimodal data acquisition module is used to acquire sensor data, 3D image data, and growth data, and preprocess the acquired data to obtain multimodal data. The data feature association module is used to establish a multimodal feature extraction network model. It uses an LSTM (Long Short-Term Memory) network to extract the time-series features of the sensors in the multimodal data, performs semantic segmentation and volume calculation on the point cloud data using a PointNet++ model, and performs One-Hot encoding on the grown data. It also generates associated feature data by weighted fusion of multi-source features through a cross-modal attention mechanism. The environmental parameter adjustment module is used to output the optimal environmental parameter adjustment strategy based on the associated feature data and through a reinforcement learning framework using RL, with the reward function being the maximization of output and quality. The environmental intelligent management module is used to adjust strategies based on the optimal environmental parameters and control the air conditioning, humidifier, supplementary lighting, and ventilation equipment in the system.

9. The intelligent management system for edible fungi cultivation environment based on big data as described in claim 8, characterized in that... The data feature association module includes the following sub-modules: The mapping submodule is used to calculate the correlation weights between each modal feature and the yield and quality of edible fungi through an attention mechanism. Based on the weight coefficients output by the attention mechanism, the feature vectors of the three modalities are weighted and summed. The fused features are then dimensionally mapped and integrated through a fully connected layer to generate associated feature data.

10. The intelligent management system for edible fungi cultivation environment based on big data as described in claim 8, characterized in that... The environmental parameter adjustment module includes the following sub-modules: Define a submodule to define the temporal feature vector, the spatial morphological feature vector of mushrooms, and the growth state feature vector as the state space; set the set of executable actions for the intelligent agent, and define the parameter adjustment range of each environmental control device as the action space. The exploration submodule employs a Deep Q-Network (DQN) as the core algorithm for reinforcement learning, inputting associated feature data into the network, and the agent explores actions using an ε-greedy strategy; The output submodule is used to collect greenhouse environmental data and mushroom growth data in real time, update the environmental status and calculate the reward value, output the Q value of each action, and select the action combination with the highest Q value as the optimal environmental parameter adjustment strategy.

Images