Intelligent control method and system for morel cultivation

Abstract

The application discloses an intelligent control method and system for Morel planting, which divides the growth of Morel into three stages of mycelium, primordium and fruiting body, collects environment and growth data by using multi-source sensors and visual equipment, and constructs a three-stage differentiated deep reinforcement learning model based on deep deterministic policy gradient. The model optimizes the Actor-Critic architecture, adopts phased priority experience replay and policy smoothing regularization, and sets a hierarchical reward function to match the needs of different growth stages. The system realizes closed-loop adaptive control through perception, decision-making, execution and digital twinning, and can accurately and smoothly adjust environmental parameters such as temperature and humidity, light, CO2 / O2 concentration, etc. The application improves the growth stability, yield and quality of Morel, overcomes the defects of traditional manual control and fixed threshold control, and is suitable for intelligent factory planting of Morel.

Description

Technical Field

[0001] This invention relates to the field of morel mushroom cultivation technology, and more specifically, to an intelligent control method and system for morel mushroom cultivation. Background Technology

[0002] Morel mushroom cultivation requires extremely high precision in controlling environmental factors such as temperature, humidity, light, and gas concentrations (CO2, O2). Traditional container cultivation relies heavily on manual experience for regulation, resulting in low matching of environmental parameters, long growth cycles, and unstable yield and quality. Furthermore, existing automated control systems mostly use fixed threshold controls, which cannot adapt to the dynamic needs of morel mushrooms at different growth stages and make it difficult to achieve synergistic optimization of environmental factors within the container. Summary of the Invention

[0003] In view of the problems existing in the prior art, the present invention provides an intelligent control method for morel mushroom cultivation, the method comprising the following steps: Step S1: The growth of morel mushrooms is divided into three core stages: mycelial stage, primordium stage, and fruiting body stage; Step S2: Collect real-time environmental data through sensors inside the cabin, and use high-definition visual acquisition equipment to obtain data on morel mycelial density, primordium quantity, and fruiting body growth morphology, and construct a dual-dimensional multi-source feature dataset of environment and growth. Step S3: Set hierarchical reward weights for the mycelial stage, primordium stage, and fruiting body stage respectively, and construct a deep reinforcement learning model based on deep deterministic policy gradient; Step S4: In the deep reinforcement learning model, optimize the Actor-Critic dual network architecture, design a phased priority experience replay mechanism, and add a policy smoothing regularization term; Step S5: Rapidly iteratively correct the output parameters based on real-time feedback data to form a complete closed-loop control; Step S6: Map the physical environment inside the container with the growth status of the bacterial strain, and finally the deep reinforcement learning model outputs precise and smooth control instructions. Preferably, in step S4, a strategy smoothing regularization term is added to suppress policy update fluctuations, prevent sudden changes in control parameters from damaging morel growth, and adapt to the hierarchical reward function to train a dedicated optimal control strategy for the core needs of each stage. Preferably, in step S2, the sensor includes a high-precision temperature and humidity sensor, a light sensor, a CO2 / O2 concentration sensor, and a soil physicochemical sensor. Preferably, in step S3, during the mycelial stage, the reward weights are set based on mycelial survival rate and pollution resistance; during the primordia stage, the reward weights are set based on primordia differentiation rate and uniformity; and during the fruiting body stage, the reward weights are set based on yield and mushroom shape quality. Preferably, in step S4, the Actor network adopts a fully connected convolutional fusion structure and is responsible for outputting continuous control action values ​​of temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, which is in line with the characteristics of morel mushrooms having a long growth cycle and delayed environmental impact. The present invention also provides an intelligent control system for morel mushroom cultivation, the system comprising an environmental multi-source sensing module, an identification module, a differentiated deep reinforcement learning decision-making module, an execution and collaborative control module, and a digital twin real-time mapping module; The identification module divides the growth of morel mushrooms into three core stages: mycelial stage, primordium stage, and fruiting body stage. The environmental multi-source perception module collects real-time environmental data through sensors inside the cabin, and combines it with the morel mycelial density, primordium quantity, and fruiting body growth morphology data obtained by the recognition module to construct a two-dimensional multi-source feature dataset of environment and growth. The differentiated deep reinforcement learning decision module sets hierarchical reward weights for the mycelial stage, primordium stage, and fruiting body stage respectively, and constructs a deep reinforcement learning model based on deep deterministic policy gradient; optimizes the Actor-Critic dual network architecture, designs a phased priority experience replay mechanism, and adds a policy smoothing regularization term; The execution and collaborative control module rapidly iterates and corrects the output parameters based on real-time feedback data, forming a complete closed-loop control. The digital twin real-time mapping module maps the physical environment inside the container and the growth status of the bacterial strains, and finally the deep reinforcement learning model outputs precise and smooth control commands. Preferably, a policy smoothing regularization term is added to the differentiated deep reinforcement learning decision module to suppress policy update fluctuations, avoid sudden changes in control parameters from damaging morel growth, and adapt to the hierarchical reward function to train a dedicated optimal control policy for the core needs of each stage. Preferably, the sensor includes a high-precision temperature and humidity sensor, a light sensor, a CO2 / O2 concentration sensor, and a soil physicochemical sensor. Preferably, in the identification module, during the mycelial stage, the reward weight is set based on mycelial survival rate and pollution resistance; during the primordium stage, the reward weight is set based on primordium differentiation rate and uniformity; and during the fruiting body stage, the reward weight is set based on yield and mushroom shape quality. Preferably, in the differentiated deep reinforcement learning decision module, the Actor network adopts a fully connected convolutional fusion structure and is responsible for outputting continuous control action values ​​for temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, which is in line with the characteristics of morel mushrooms having a long growth cycle and delayed environmental impact. The innovative aspects of the embodiments in the specification include: 1. Addressing the closed and controllable characteristics of modular morel mushroom plant factories, this solution overcomes the limitations of traditional fixed threshold control and general reinforcement learning control in meeting the heterogeneous growth requirements of morel mushrooms across three stages and their poor adaptability to environmental disturbances. It provides a refined and adaptive intelligent environmental control solution to achieve dynamic and precise control of temperature, humidity, light, and air within the modular facility, thereby improving morel mushroom yield and quality while reducing the cost of manual intervention. 2. This invention constructs a three-stage differentiated deep reinforcement learning model of hyphae-primordia-fruiting bodies and designs a "growth-adaptive hierarchical reward function + environmental perturbation adaptive fine-tuning mechanism", which breaks through the limitations of the general control of traditional single reinforcement learning models. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments or related technologies of this specification, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 A schematic diagram of the intelligent control system for morel cultivation provided by the present invention; Figure 2 A flowchart for intelligent regulation of morel mushroom cultivation provided by the present invention; Attached reference numerals: 10, Multi-source perception module for the mobile cabin environment; 20, Precise identification module for three-stage growth status; 30, Differentiated deep reinforcement learning decision-making module; 40, Execution and collaborative control module; 50, Real-time mapping module for digital twin; 100, Intelligent control system. Detailed Implementation The technical solutions in the embodiments of this specification 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. It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this specification are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. This specification discloses an intelligent control method and system for morel mushroom cultivation, which will be described in detail below. Example 1 like Figure 1 As shown, this invention provides an intelligent control system 100 for morel mushroom cultivation. The system includes a multi-source sensing module 10 for the container environment, a three-stage growth state precise identification module 20, a differentiated deep reinforcement learning decision-making module 30, an execution and collaborative control module 40, and a digital twin real-time mapping module 50. The modular environment multi-source sensing module 10 includes a temperature sensor, a humidity sensor, and a CO2 / O2 gas concentration sensor; the temperature sensor has an accuracy of ±0.1℃ (temperature), the humidity sensor has an accuracy of ±1%RH (humidity), and the CO2 / O2 gas concentration sensor has an accuracy of ±50ppm (CO2); the data acquisition frequency is once every 5 minutes, the feature dataset redundancy removal rate is 42%, and the core feature retention rate is ≥98%, ensuring accurate and efficient data. The three-stage growth state precise identification module 20 refines the growth of morel mushrooms into five sub-stages: mycelial colonization stage, mycelial extension stage, primordium induction stage, primordium differentiation stage, and fruiting body enlargement and fruiting body development stage. These five sub-stages are then classified into three core stages: mycelium-primordium-fruiting body. The differentiated deep reinforcement learning decision module 30 sets hierarchical reward weights for the mycelial stage (focusing on mycelial survival rate and pollution resistance), the primordium stage (focusing on primordium differentiation rate and uniformity), and the fruiting body stage (focusing on yield and mushroom shape quality), and constructs a deep reinforcement learning (DRL) model based on DDPG (deep deterministic policy gradient). This invention makes targeted improvements to the algorithm: First, it optimizes the Actor-Critic dual-network architecture. The Actor network adopts a fully connected convolutional fusion structure, responsible for outputting continuous control action values ​​for temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, which is in line with the characteristics of morel mushrooms' long growth cycle and the lag of environmental influences. Second, it designs a phased priority experience replay mechanism. Based on the reward weights of the three stages of mycelium, primordium, and fruiting body, it divides the priority of experience samples and prioritizes the extraction of high-value samples for training, solving the problem of slow model convergence under small sample planting data. Third, it adds a policy smoothing regularization term to suppress policy update fluctuations and avoid damage to morel mushroom growth caused by sudden changes in control parameters. At the same time, it adapts to a hierarchical reward function to train a dedicated optimal control strategy for the core needs of each stage, avoiding the problem of growth damage caused by "one-size-fits-all" control. Preferably, the hierarchical reward function is as follows: Where the weight coefficients satisfy ; Mycelial stage: ω1=0.7, ω2=0.2, ω3=0.1, ; Primordial periods: ω1=0.2, ω2=0.7, ω3=0.1, ; Fruiting body stage: ω1=0.1, ω2=0.2, ω3=0.7, ; α and β are normalization coefficients ranging from 0 to 1, which can be adaptively adjusted according to the characteristics of the strain. The expression for the smoothing regularization term of the strategy is: In the formula Let π be the probability of choosing action a given state s under a policy with parameter θ after the update. Given a state s, what is the probability of choosing action a before updating π under a policy with parameter θ? λ is set to 0.001~0.01, and the single-time variation of the control parameters is constrained as follows: temperature ≤ ±0.3℃, humidity ≤ ±2%RH, CO2 ≤ ±100ppm. The Actor network structure is: input layer → 3×3 convolutional layer (ReLU) → batch normalization → fully connected layer (256→128→64) → tanh output layer; the Critic network adopts an 8-head attention mechanism, the optimizer is Adam, and the learning rate is 1×10⁻ 4 , batch 32, γ=0.99, τ=0.005; Model convergence criterion: The rate of change of loss over 100 consecutive rounds < 1 × 10⁻ 6 The formula for calculating the priority of phased priority experience replay is as follows: , For TD error, The constant is α=0.6; the experience pool capacity is 100,000, and the three-stage samples are stored in a 3:3:4 ratio, with priority given to sampling high-value samples. Preferably, the specific implementation of the phased priority experience replay can be as follows: the experience pool is divided into three stages of growth data—mycelial stage, primordium stage, and fruiting body stage—and the samples are stored in a 3:3:4 ratio. During model training, based on the preset stratified reward weights for each stage, high-value samples with high reward values ​​and high TD errors in the corresponding stage are preferentially extracted for network updates. Among them, mycelial stage samples prioritize mycelial survival rate and pollution resistance-related experience, primordium stage samples prioritize primordium differentiation rate and uniformity-related experience, and fruiting body stage samples prioritize yield and mushroom shape quality-related experience, thereby improving the model convergence speed and control accuracy in small sample scenarios. The improved DDPG model has a 58% faster convergence speed than the traditional DDPG and a 65% smaller policy update oscillation. With a small sample size (50 sets of planting data) for training, the model's tuning accuracy reaches 92.3%, which is 27.6% higher than the traditional reinforcement learning model. When the execution and collaborative control module 40 detects that the environmental disturbance inside the cabin exceeds the threshold, the model automatically activates the disturbance adaptive fine-tuning mechanism, and quickly iterates and corrects the output parameters based on real-time feedback data, forming a complete closed-loop control of "perception-preprocessing-model inference-instruction output-execution-feedback-adaptation". In this embodiment, the threshold for environmental disturbance triggering adaptive fine-tuning is: temperature fluctuation > ±0.5℃, humidity fluctuation > ±3%RH, CO2 fluctuation > ±150ppm; the control command update cycle is 5min / time, and the disturbance response time is ≤20s; the actuator adopts the RS485 bus control protocol. The control logic of the execution and collaborative control module is as follows: The Actor network outputs continuous control action values ​​for temperature, humidity, light intensity, and CO2 concentration. The system linearly maps these continuous action values ​​to the corresponding adjustment amounts of the actuators. Specifically, the temperature action value is mapped to the power or opening degree of the heating / cooling equipment, the humidity action value is mapped to the humidifier spray volume or dehumidifier operating power, the light intensity value is mapped to the PWM duty cycle of the supplementary lighting lamp, and the CO2 concentration is mapped to the ventilation volume or the opening degree of the supplementary air valve. The multi-device collaboration adopts a priority scheduling + mutual exclusion protection mechanism to avoid conflicts caused by simultaneous heating and cooling, or humidification and dehumidification, ensuring the smooth execution of control commands and keeping environmental parameter fluctuations within the set range. Preferably, the present invention further includes a three-stage adaptive threshold perturbation triggering mechanism, which dynamically and adaptively adjusts the environmental perturbation triggering threshold based on the physiological sensitivity differences of morel mycelial stage, primordium stage, and fruiting body stage. 1) Mycelial stage: It has a relatively high tolerance to environmental changes. The disturbance trigger thresholds are set as follows: temperature fluctuation > ±0.5℃, humidity fluctuation > ±3%RH, and CO2 fluctuation > ±150ppm. 2) Primordial stage: Primordial differentiation is fragile and sensitive to sudden environmental changes. The disturbance trigger threshold is automatically tightened to temperature fluctuation > ±0.3℃, humidity fluctuation > ±2%RH, and CO2 fluctuation > ±100ppm; 3) Fruiting body stage: During the fruiting body expansion stage, stability and growth efficiency need to be considered. The disturbance trigger thresholds are set as follows: temperature fluctuation > ±0.4℃, humidity fluctuation > ±2.5%RH, and CO2 fluctuation > ±120ppm. By using phased adaptive thresholds, the system avoids both significant damage during the primordia stage due to minor disturbances and excessive energy consumption during the mycelial and fruiting body stages. This achieves differentiated intelligent regulation that provides "strict protection during the vulnerable stage and high efficiency during the robust stage," further improving the yield and growth stability of morel mushrooms. Target environmental parameter ranges for each stage: Mycelial stage: Temperature 18~22℃, humidity 65%~75%RH, light 0~50lx, CO2 800~1200ppm; Primordial period: Temperature 13~16℃, humidity 80%~90%RH, light 200~500lx, CO2 500~800ppm; Fruiting body stage: temperature 12~15℃, humidity 75%~85%RH, light 500~1000lx, CO2 300~600ppm. Experimental data: Under closed-loop control, the fluctuation range of environmental parameters in the cabin is ≤ ±0.3℃, ±2%RH, and ±100ppm, and the disturbance response time is ≤20s. Compared with traditional fixed threshold control, the survival rate of morel mycelium is increased by 31.2%, the primordium differentiation rate is increased by 38.7%, the fruiting body deformity rate is reduced by 62.5%, and the overall yield is increased by 35.4%. The digital twin real-time mapping module 50 is built based on Unity3D, with a data synchronization frequency of 1 time / 5s. It achieves a 1:1 mapping through coordinate transformation and spatial interpolation. The position error is ≤±1cm, and the environmental parameter error is ≤±0.1℃, ±1%RH, and ±50ppm. It supports visualization, early warning, and historical backtracking. Preferably, the specific implementation of the real-time digital twin mapping is as follows: a 3D scene of the modular cabin is built using Unity3D, and the spatial position and equipment layout of the physical modular cabin are mapped to the virtual scene at a 1:1 ratio through coordinate transformation; data synchronization adopts a combination of timed synchronization and event-triggered synchronization, and environmental parameters and growth status data are updated at a frequency of 5 seconds / time; spatial interpolation algorithm is used to fuse multi-point sensor data to achieve a smooth visualization of the environmental field and growth status, and an error calibration mechanism is used to ensure that the position error is ≤±1cm, and the environmental parameter error is ≤±0.1℃, ±1% RH, and ±50ppm. The system provides a 1:1 real-time mapping of the physical environment and bacterial growth status within the container. The "environment + growth" dual-dimensional data collected by the sensing module undergoes standardization and normalization preprocessing to remove abnormal noise and unify data volume. This data is then input into a pre-trained improved DDPG differentiated DRL model for refined inference operations. The model relies on an Actor-Critic dual-network architecture for core processing: the Actor network extracts key features based on the input multi-source features through a fully connected convolutional fusion layer, combining the weights of the three-stage differentiated strategy (mycelium-primordium-fruiting body) to generate continuous initial control action values ​​for temperature, humidity, light intensity, and CO2 / O2 concentration. The Critic network introduces an attention mechanism to focus on core growth indicators and environmental disturbance factors of morel mushrooms, evaluating the benefits and correcting deviations of the initial action values, eliminating control parameters that do not conform to growth patterns or have excessive energy consumption. Finally, the model outputs precise and smooth control commands to drive the humidifier, cooling / heating equipment, supplemental lighting, gas exchange devices, and other components to perform coordinated adjustments. Example 2 like Figure 2 The figure shows an intelligent control method for morel mushroom cultivation provided by the present invention, the method comprising the following steps: Step S1: The growth of morel mushrooms is further divided into five sub-stages: mycelial colonization stage, mycelial extension stage, primordium induction stage, primordium differentiation stage, and fruiting body enlargement and fruiting body formation stage. These five sub-stages are then classified into three core stages: mycelium-primordium-fruiting body. Step S2: Real-time environmental data is collected using high-precision temperature and humidity sensors, light sensors, CO2 / O2 concentration sensors, and soil physicochemical sensors inside the container. Combined with high-definition visual acquisition equipment, morphological data such as morel mycelial density, primordium quantity, and fruiting body growth are obtained to construct a dual-dimensional multi-source feature dataset of "environment + growth". Redundant data is removed and core features are retained. Step S3: Set hierarchical reward weights for the mycelial stage (focusing on mycelial survival rate and pollution resistance), primordium stage (focusing on primordium differentiation rate and uniformity), and fruiting body stage (focusing on yield and mushroom shape quality), and construct a deep reinforcement learning (DRL) model based on DDPG (Deep Deterministic Policy Gradient). Step S4: Optimize the Actor-Critic dual-network architecture. The Actor network adopts a fully connected convolutional fusion structure, responsible for outputting continuous control action values ​​for temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, aligning with the characteristics of morel mushrooms' long growth cycle and the lag in environmental influences. A phased priority experience replay mechanism is designed, prioritizing experience samples based on the reward weights of the three stages: mycelium, primordium, and fruiting body. High-value samples are selected for training first, addressing the slow convergence issue of the model under small sample planting data. A policy smoothing regularization term is added to suppress policy update fluctuations and prevent sudden changes in control parameters from damaging morel mushroom growth. Simultaneously, a hierarchical reward function is adapted to train a dedicated optimal control strategy for the core needs of each stage, avoiding growth damage caused by a "one-size-fits-all" approach to control. In this embodiment, the hierarchical reward function, policy smoothing regularization term, Actor-Critic network structure, hyperparameters, and experience replay rules are the same as in Embodiment 1, ensuring the accuracy and reproducibility of the three-stage control strategy. Step S5: When the environmental disturbance inside the cabin exceeds the threshold, the model automatically activates the disturbance adaptive fine-tuning mechanism, and quickly iterates and corrects the output parameters based on real-time feedback data, forming a complete closed-loop control of "perception-preprocessing-model inference-instruction output-execution-feedback-adaptation". Step S6: Real-time 1:1 mapping of the physical environment and bacterial growth status within the shelter. The dual-dimensional data of "environment + growth" collected by the sensing module is first preprocessed by standardization and normalization to remove abnormal noise and unify the data volume. Then, it is input into the trained improved DDPG differential DRL model for refined inference calculation. Finally, the model outputs precise and smooth control commands to drive the humidifier, cooling / heating equipment, supplementary lighting, gas exchange device, etc., to perform and coordinate adjustments. In this embodiment, the dataset consists of 50 groups, divided into training and validation sets in a 7:3 ratio. Z-score standardization is used, and outliers are removed using the 3σ criterion. The core feature retention rate is ≥98%. The improved DDPG model achieves a 58% increase in convergence speed, a 65% reduction in policy oscillation, and a 92.3% accuracy in small-sample training, enabling stable intelligent regulation of morel mushrooms. This invention provides targeted adaptation and improvement to the DDPG algorithm, forming a dedicated optimized algorithm adapted to the morel mushroom facility cultivation scenario. Specific improvements are as follows: First, the Actor-Critic dual-network architecture is optimized to improve the accuracy and adaptability of control decisions. The Actor network adopts a hybrid structure combining fully connected and convolutional layers. Convolutional layers extract temporal and spatial correlation features of environmental factors, while fully connected layers complete feature mapping and action output, ultimately generating continuous control action values ​​for key environmental factors such as temperature, humidity, light intensity, and gas concentration (mainly CO2 concentration), ensuring the continuity and smoothness of control parameters. The Critic network introduces an attention mechanism, dynamically allocating attention weights to focus on the long-term impact of each environmental control action on the morel mushroom growth state, accurately evaluating the long-term benefits of different control strategies. This effectively adapts to the core characteristics of morel mushrooms—long growth cycles and the lag in environmental impact—avoiding control biases caused by short-term gain orientation. Secondly, a phased priority-based experience replay (PER) mechanism is designed. Combined with the hierarchical reward weight system constructed above, the experience samples in the model training process are prioritized. Among them, experience samples that are highly correlated with the core objectives of each growth stage (mycelial survival rate and pollution resistance during the mycelial stage; primordia differentiation rate and growth uniformity during the primordia stage; and yield and mushroom shape quality during the fruiting body stage) and have high reward values ​​are given high priority. During the training process, such high-value samples are extracted first for model updates. This effectively solves the problems of slow model convergence and low training accuracy caused by the difficulty of data collection and insufficient sample size in morel cultivation, and improves the adaptability and generalization ability of the model in small sample scenarios. Finally, a policy smoothing regularization term is added to the algorithm's loss function. By constraining the regulatory actions of the Actor network output, drastic fluctuations during policy updates are suppressed, preventing mechanical damage to morel growth caused by sudden changes in environmental regulation parameters and ensuring the stability of the growth environment. At the same time, this regularization term is deeply adapted to the hierarchical reward function, which can guide the model to train a specific optimal control strategy for the core cultivation needs of morel at different growth stages. This completely eliminates the drawbacks of the "one-size-fits-all" approach in traditional regulation modes, achieving precise matching between environmental regulation strategies at each growth stage and the physiological needs of morel growth, providing core algorithmic support for high-quality, high-yield, and stable morel cultivation. Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention. Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart control method for morel mushroom cultivation, characterized in that: The method includes the following steps: Step S1: The growth of morel mushrooms is divided into three core stages: mycelial stage, primordium stage, and fruiting body stage; Step S2: Collect real-time environmental data through sensors inside the cabin, and use high-definition visual acquisition equipment to obtain data on morel mycelial density, primordium quantity, and fruiting body growth morphology, and construct a dual-dimensional multi-source feature dataset of environment and growth. Step S3: Set hierarchical reward weights for the mycelial stage, primordium stage, and fruiting body stage respectively, and construct a deep reinforcement learning model based on deep deterministic policy gradient; Step S4: In the deep reinforcement learning model, optimize the Actor-Critic dual network architecture, design a phased priority experience replay mechanism, and add a policy smoothing regularization term; Step S5: Rapidly iteratively correct the output parameters based on real-time feedback data to form a complete closed-loop control; Step S6: Map the physical environment inside the container with the growth status of the bacterial strain, and finally the deep reinforcement learning model outputs precise and smooth control instructions.

2. The method according to claim 1, characterized in that: In step S4, a policy smoothing regularization term is added to suppress policy update fluctuations and prevent sudden changes in control parameters from damaging morel growth. At the same time, a hierarchical reward function is adapted to train an exclusive optimal control policy for the core needs of each stage.

3. The method according to claim 1, characterized in that: In step S2, the sensors include a high-precision temperature and humidity sensor, a light sensor, a CO2 / O2 concentration sensor, and a soil physicochemical sensor.

4. The method according to claim 1, characterized in that: In step S3, during the mycelial stage, the reward weights are set based on mycelial survival rate and pollution resistance; during the primordia stage, the reward weights are set based on primordia differentiation rate and uniformity; and during the fruiting body stage, the reward weights are set based on yield and mushroom shape quality.

5. The method according to claim 1, characterized in that: In step S4, the Actor network adopts a fully connected convolutional fusion structure and is responsible for outputting continuous control action values ​​for temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, which is in line with the characteristics of morel mushrooms having a long growth cycle and the lag of environmental influences.

6. An intelligent control system for morel mushroom cultivation, characterized in that: The system includes an environmental multi-source perception module, an identification module, a differentiated deep reinforcement learning decision-making module, an execution and collaborative control module, and a digital twin real-time mapping module. The identification module divides the growth of morel mushrooms into three core stages: mycelial stage, primordium stage, and fruiting body stage. The environmental multi-source perception module collects real-time environmental data through sensors inside the cabin, and combines it with the morel mycelial density, primordium quantity, and fruiting body growth morphology data obtained by the recognition module to construct a two-dimensional multi-source feature dataset of environment and growth. The differentiated deep reinforcement learning decision module sets hierarchical reward weights for the mycelial stage, primordium stage, and fruiting body stage respectively, and constructs a deep reinforcement learning model based on deep deterministic policy gradient; optimizes the Actor-Critic dual network architecture, designs a phased priority experience replay mechanism, and adds a policy smoothing regularization term; The execution and collaborative control module rapidly iterates and corrects the output parameters based on real-time feedback data, forming a complete closed-loop control. The digital twin real-time mapping module maps the physical environment inside the container and the growth status of the bacterial strains, and finally the deep reinforcement learning model outputs precise and smooth control commands.

7. The system according to claim 1, characterized in that: A policy smoothing regularization term is added to the differentiated deep reinforcement learning decision module to suppress policy update fluctuations and avoid sudden changes in control parameters that could damage morel growth. At the same time, it adapts to the hierarchical reward function and trains a dedicated optimal control policy for the core needs of each stage.

8. The system according to claim 1, characterized in that: The sensors include a high-precision temperature and humidity sensor, a light sensor, a CO2 / O2 concentration sensor, and a soil physicochemical sensor.

9. The system according to claim 1, characterized in that: In the identification module, during the mycelial stage, the reward weight is set based on mycelial survival rate and pollution resistance; during the primordium stage, the reward weight is set based on primordium differentiation rate and uniformity; and during the fruiting body stage, the reward weight is set based on yield and mushroom shape quality.

10. The system according to claim 1, characterized in that: In the differentiated deep reinforcement learning decision module, the Actor network adopts a fully connected convolutional fusion structure and is responsible for outputting continuous control action values ​​for temperature, humidity, light, and gas concentration. The Critic network introduces an attention mechanism to accurately evaluate the long-term benefits of each action, which is in line with the characteristics of morel mushrooms having a long growth cycle and delayed environmental impact.

Images