Constant-temperature incubator intelligent control method and system based on deep learning

Abstract

The invention discloses an intelligent control method and system for a constant-temperature incubator based on deep learning. The method comprises the following steps: acquiring an equipment state vector in real time; analyzing the equipment state vector by using an improved model, and predicting an environmental parameter demand trend in a culture process to obtain a prediction result; modeling a control strategy into an MDP Markov decision process through a DRL deep reinforcement learning framework according to the prediction result, and dynamically optimizing power distribution of a heater, a refrigerator, a humidifier and a fan by taking minimization of energy consumption and temperature fluctuation as objective functions to obtain an initial control strategy; and adjusting the power of the actuator according to the initial control strategy by using a fuzzy PID algorithm, monitoring sensor data in real time through an isolated forest model, and adjusting the initial control strategy when abnormality is detected to obtain a target control strategy. The environment demand change is pre-judged in advance, and the hysteresis problem of traditional control is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of incubator control, and particularly relates to an intelligent control method and system for an incubator based on deep learning. BACKGROUND

[0002] The control method of the existing incubator mainly relies on the traditional PID algorithm or static rules, and it is difficult to cope with complex and changeable culture requirements. The traditional control method mainly adopts a simple PID control algorithm, and only feedback adjustment can be made according to the current temperature deviation. The response to the nonlinear change of the environmental parameters is lagged, and overshoot or oscillation is prone to occur. The temperature fluctuation is large, and it is difficult to meet the high-precision requirements of cell culture. SUMMARY

[0003] The present application relates to the technical field of incubator control, and particularly relates to an intelligent control method and system for an incubator based on deep learning.

[0004] To achieve the above-mentioned purpose, the technical scheme of the present application is that, further, in the intelligent control method for the incubator based on deep learning, the intelligent control method for the incubator based on deep learning comprises the following steps: Real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door body state data in the incubator, preprocessing of the collected data, and obtaining of the equipment state vector; Using an improved Transformer model to analyze the equipment state vector, predicting the environmental parameter demand trend in the culture process, and obtaining a prediction result; According to the prediction result, modeling the control strategy as an MDP Markov decision process through a DRL deep reinforcement learning framework, taking the minimum energy consumption and temperature fluctuation as the objective function, dynamically optimizing the power distribution of the heater, the cooler, the humidifier and the fan, and obtaining an initial control strategy; Using a fuzzy PID algorithm to adjust the actuator power according to the initial control strategy, and using an isolation forest model to monitor the sensor data in real time. When an anomaly is detected, the initial control strategy is adjusted to obtain a target control strategy.

[0005] Further, in the intelligent control method for the incubator based on deep learning, the real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door body state data in the incubator, the preprocessing of the collected data, and the obtaining of the equipment state vector comprise: Real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door body state data in the incubator; Adopting a sliding window mean filter to eliminate sensor fault values and short-time noise, and mapping multi-source heterogeneous data to a unified range to obtain processed data; The time sequence features of the processed data are extracted, and the preprocessed temperature, humidity, CO2 concentration, air flow speed and door state are integrated into a multi-dimensional vector to obtain an equipment state vector.

[0006] Further, in the above-mentioned intelligent control method for constant-temperature incubator based on deep learning, the improved Transformer model is used to analyze the equipment state vector, predict the demand trend of the environmental parameters in the cultivation process, and obtain a prediction result, which includes: A space-time attention mechanism is added to the Transformer model, wherein the space attention analyzes the spatial dependence of the sensor network, the time attention captures the long-period law, and a sine-cosine position coding is used to enhance the time sequence perception of the model to the time sequence data, thereby obtaining an improved Transformer model. The improved Transformer model is used to analyze the equipment state vector, predict the demand trend of the environmental parameters in the cultivation process, and obtain a prediction result.

[0007] Further, in the above-mentioned intelligent control method for constant-temperature incubator based on deep learning, according to the prediction result, a control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework, a power distribution of a heater, a cooler, a humidifier and a fan is dynamically optimized to minimize energy consumption and temperature fluctuation as an objective function, and an initial control strategy is obtained, which includes: A state space of the MDP Markov decision process is defined, including the equipment state vector and the prediction result, and an action space is defined as the heater power, the cooler power, the humidifier power and the fan speed, and each action is subjected to physical constraints. The minimization of energy consumption and temperature fluctuation is taken as the core, and a reward function is established by comprehensively considering the control effect of CO2 and humidity.

[0008] Further, in the above-mentioned intelligent control method for constant-temperature incubator based on deep learning, according to the prediction result, a control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework, a power distribution of a heater, a cooler, a humidifier and a fan is dynamically optimized to minimize energy consumption and temperature fluctuation as an objective function, and an initial control strategy is obtained, which includes: A DDPG deep deterministic policy gradient algorithm is used to learn an optimal control strategy through an Actor-Critic architecture. The Actor network is used to input the state and output continuous actions of the actuator power, and the Critic network is used to evaluate the value of the current action to guide the optimization of the Actor. Action selection is added with Ornstein-Uhlenbeck noise to balance exploration and utilization, the noise intensity decays with the training process, and the actuator power sequence of the next 5 minutes is output to obtain the initial control strategy.

[0009] Further, in the above-mentioned intelligent control method for constant-temperature incubator based on deep learning, the fuzzy PID algorithm is used to adjust the actuator power according to the initial control strategy, the Isolation Forest model is used to monitor the sensor data in real time, and the initial control strategy is adjusted when an anomaly is detected to obtain a target control strategy, comprising: The fuzzy PID algorithm is used to take the deviation and deviation rate of the initial control strategy and the current measured value as input, and output the actuator power correction amount.

[0010] Further, in the above-mentioned intelligent control method for constant-temperature incubator based on deep learning, the fuzzy PID algorithm is used to adjust the actuator power according to the initial control strategy, the Isolation Forest model is used to monitor the sensor data in real time, and the initial control strategy is adjusted when an anomaly is detected to obtain a target control strategy, comprising: The Isolation Forest model is used to monitor whether the sensor data deviates from the normal distribution and the actuator feedback is abnormal, the model is trained with historical normal operation data, and the abnormal score threshold is set to 0.8; If it is a mild anomaly, the strong correction mode of the fuzzy PID is enabled, and the deviation is recorded for subsequent model calibration; if it is a severe anomaly, a safety strategy is triggered immediately and an audible and visual alarm is issued, and the alarm information is pushed to the management personnel through the Internet of Things module.

[0011] Further, in an intelligent control system for constant-temperature incubator based on deep learning, the intelligent control system for constant-temperature incubator based on deep learning comprises the following modules: A state vector acquisition module is used to acquire temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator in real time, preprocess the acquired data, and obtain a device state vector; A device state prediction module is used to analyze the device state vector using an improved Transformer model to predict the environmental parameter demand trend in the incubation process and obtain a prediction result; A control strategy generation module is used to model the control strategy as a MDP Markov decision process through a DRL deep reinforcement learning framework according to the prediction result, take the minimization of energy consumption and temperature fluctuation as the objective function, dynamically optimize the power distribution of the heater, cooler, humidifier and fan, and obtain an initial control strategy; A control strategy optimization module is used to adjust the actuator power according to the initial control strategy using the fuzzy PID algorithm, monitor the sensor data in real time using the Isolation Forest model, and adjust the initial control strategy when an anomaly is detected to obtain a target control strategy.

[0012] Furthermore, in a deep learning-based intelligent control system for a constant temperature incubator, the control strategy optimization module includes the following sub-modules: The correction submodule is used to take the deviation between the initial control strategy and the current measured value and the rate of change of the deviation as inputs using the fuzzy PID algorithm, and outputs the actuator power correction amount.

[0013] Furthermore, in a deep learning-based intelligent control system for a constant temperature incubator, the control strategy optimization module includes the following sub-modules: The judgment submodule is used to monitor whether sensor data deviates from the normal distribution and whether actuator feedback is abnormal through the isolated forest model. The model is trained with historical normal operation data and the abnormal score threshold is set to 0.8. The control submodule is used to enable the strong correction mode of fuzzy PID in case of a minor anomaly, and record the deviation for subsequent model calibration; in case of a severe anomaly, it immediately triggers the safety policy and issues an audible and visual alarm, and pushes the alarm information to the management personnel through the IoT module.

[0014] Its beneficial effects are as follows: 1. The improved Transformer model is used to predict trends in multi-parameter time-series data, anticipating changes in environmental requirements 60 minutes in advance, thus solving the lag problem of traditional control; the fuzzy PID algorithm combined with the DRL initial strategy achieves dual regulation, keeping temperature fluctuations within ±0.1℃, and improving the control accuracy of humidity and CO2 concentration to meet the requirements of high-precision cultivation. 2. The DRL framework uses minimizing energy consumption as the objective function, optimizing the power allocation of actuators such as heaters and coolers to avoid ineffective energy consumption; combined with parameter prediction, it achieves advance adjustment, reducing frequent start-stop of actuators, and experimental verification shows that it can reduce energy consumption and significantly reduce long-term operating costs. 3. The isolated forest model monitors sensor data and actuator status in real time, identifying anomalies within 1 second and triggering an adaptive adjustment strategy; in the case of mild anomalies, stability is maintained through fuzzy PID parameter correction, and in the case of severe anomalies, a safety mechanism is automatically activated and an alarm is triggered, reducing the risk of cultivation failure. 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 a deep learning-based intelligent control method for a constant temperature incubator according to the present invention. Figure 2 This is a schematic diagram of a second embodiment of a deep learning-based intelligent control method for a constant temperature incubator according to the present invention. Figure 3 This is a schematic diagram of the first embodiment of an intelligent control system for a constant temperature incubator based on deep learning, as described in this invention. 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 embodiments described herein are merely illustrative and 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 with reference to the accompanying drawings. Figure 1 As shown, a deep learning-based intelligent control method for a constant temperature incubator includes the following steps: Step 101: Real-time acquisition of temperature, humidity, CO2 concentration, airflow velocity, and door status data within the incubator. Preprocessing the acquired data yields a device status vector. Specifically, in this embodiment, real-time acquisition of temperature, humidity, CO2 concentration, airflow velocity, and door status data within the incubator is performed. A sliding window mean filter is used to remove sensor fault values ​​and short-term noise, and multi-source heterogeneous data is mapped to a unified range to obtain processed data. The time-series features of the processed data are extracted, and the preprocessed temperature, humidity, CO2 concentration, airflow velocity, and door status data are integrated into a multi-dimensional vector to obtain the device status vector. Real-time acquisition of temperature, humidity, CO2 concentration, airflow velocity, and door status data within the incubator. To achieve comprehensive monitoring of the incubator environment, multiple types of high-precision sensors and a synchronous acquisition mechanism are required: temperature acquisition utilizes 3-5 distributed PT1000 platinum resistance sensors, covering key areas such as the top, middle, and sample zone of the incubator to ensure the capture of spatial temperature gradients; humidity monitoring employs capacitive sensors, installed in areas with stable airflow to avoid direct interference from humidifiers, achieving an accuracy of ±2%RH; CO2 concentration is detected in real-time using a non-dispersive infrared (NDIR) sensor with a sampling frequency of 1Hz to track concentration changes caused by culture metabolism; airflow velocity is collected by a thermal anemometer near the fan outlet to reflect the uniformity of circulating airflow; and door status is sampled at a high frequency of 10Hz using a magnetic proximity switch to accurately capture transient changes during switching. All sensor data are uniformly received through an embedded control system such as the STM32H7 series, and time-stamped synchronously using a hardware clock to ensure that the time deviation of multi-source data is ≤10ms, providing a spatiotemporally consistent raw data foundation for subsequent analysis.

[0020] A sliding window mean filter is used to remove sensor fault values ​​and short-term noise, and to map multi-source heterogeneous data to a unified range to obtain processed data. For random noise that may exist in the raw sensor data, such as transient jumps caused by electronic interference, and fault values ​​such as abnormal readings caused by sensor offline, a sliding window mean filter is used for preprocessing: the window size is set to 5-10 sampling points corresponding to 5-10 seconds, adapted to a standard 1Hz sampling frequency. The mean value within the window is used to replace the current point value, smoothing short-term fluctuations. For obvious outliers exceeding the 3σ range, such as a temperature surge of more than 10℃, they are first marked as fault points, and then linear interpolation of the effective values ​​at both ends of the window is used to fill in the fault data, avoiding contamination by fault data. Simultaneously, due to significant differences in the physical dimensions of various parameters (e.g., temperature is in °C, CO2 is in %), standardization is required to map them to a unified range: the mean μ and standard deviation σ of each parameter are calculated based on historical normal operation data, and the data is converted into standardized values ​​with a mean of 0 and a variance of 1, eliminating the impact of magnitude differences on subsequent model analysis and ensuring that parameters such as temperature and humidity have balanced weights in the feature space.

[0021] Time-series features of the processed data are extracted, and the preprocessed temperature, humidity, CO2 concentration, airflow velocity, and door status are integrated into a multi-dimensional vector to obtain the equipment state vector. Based on the preprocessed data, key time-series features need to be extracted to enrich the state description: for temperature, the mean value of the distributed sensors is calculated to reflect the overall temperature level, the maximum deviation reflects spatial uniformity, and the 5-minute sliding window change rate reflects the heating / cooling trend; for humidity, CO2 concentration, and airflow velocity, the current value, the cumulative change over 10 minutes reflects the long-term trend, and the fluctuation amplitude within 30 seconds reflects stability, respectively. For door status, in addition to retaining the real-time open / close state 0 / 1, the number of opening and closing times in the past 5 minutes and the cumulative opening duration are added to reflect the degree of environmental interference. Subsequently, these features are integrated with the time labels of the cultivation process, such as cultivation duration and stage: preheating / stabilization / end, to form a multi-dimensional vector containing temperature features, humidity features, CO2 features, airflow features, door status features, and time labels. Finally, an equipment state vector with dimensions of 17-20 is constructed to comprehensively characterize the real-time operating status of the incubator and provide structured input for subsequent model prediction and strategy optimization.

[0022] Step 102: Analyze the equipment state vector using the improved Transformer model, predict the trend of environmental parameter demand during the cultivation process, and obtain the prediction results; Specifically, this embodiment adds a spatiotemporal attention mechanism to the Transformer model. Spatial attention analyzes the spatial dependencies of the sensor network, temporal attention captures long-period patterns, and sine-cosine position coding enhances the model's temporal awareness of time-series data, resulting in an improved Transformer model. The improved Transformer model is then used to analyze the device state vector and predict the environmental parameter demand trends during the cultivation process, yielding prediction results. The addition of a spatiotemporal attention mechanism to the Transformer model, where spatial attention analyzes the spatial dependencies of the sensor network, temporal attention captures long-period patterns, and sine-cosine position coding enhances the model's temporal awareness of time-series data, results in an improved Transformer model.

[0023] To enhance the Transformer model's ability to analyze multi-source heterogeneous time-series data from incubators, model improvements are achieved by integrating spatiotemporal attention mechanisms and positional encoding. The spatial attention module constructs a sensor position correlation matrix based on the layout characteristics of the distributed sensor network. Using the physical coordinates of sensors such as temperature probes and humidity sensors in different areas as a foundation, it captures spatial dependencies by calculating attention weights—for example, the numerical correlation between the top temperature sensor and the sample area temperature sensor, and the spatial coupling between the fan outlet speed and the humidity distribution within the incubator. This enables the model to identify the spatial transmission patterns of environmental parameters. The temporal attention module employs a multi-layered, multi-head structure, capturing long-term patterns by setting different attention windows, such as a short-term window of 30 minutes and a long-term window of 24 hours. Examples include the impact of diurnal temperature fluctuations on the incubator interior and the periodic accumulation trend of CO2 concentration during culture metabolism. Simultaneously, sine-cosine positional encoding is introduced to add temporal markers to each time step, enabling the model to distinguish the sequence of different time points and enhancing its perception of dynamic changes in time-series data. Through the above improvements, the model can simultaneously analyze the spatial distribution characteristics and temporal evolution of environmental parameters, solving the problem that the standard Transformer is insufficient in capturing the spatial correlation of multi-source sensor data.

[0024] An improved Transformer model is used to analyze the equipment state vectors and predict the environmental parameter demand trends during the cultivation process, yielding prediction results. Based on the improved Transformer model, deep analysis of equipment state vectors enables accurate prediction of environmental parameter demand trends during the cultivation process. Equipment state vectors with dimensions 17-20, including multi-dimensional features such as temperature, humidity, CO2 concentration, airflow velocity, and gate status, are input into the model. First, an embedding layer transforms discrete features, such as gate status labels, into continuous vectors, which are then concatenated with continuous features, such as the average temperature, to form the model input sequence. During spatiotemporal attention processing, the spatial attention layer focuses on the spatial correlations of various sensor data, such as the spatial response relationship between sample area temperature and heater power, while the temporal attention layer mines long-period patterns in historical data, such as the sudden increase in temperature demand during specific cultivation stages. Combined with the temporal information provided by sine-cosine position encoding, the model can analyze the dynamic coupling patterns between parameters. The decoder uses an autoregressive generation mode, using current and historical state vectors as conditions to progressively predict environmental parameter demand trends for the next 60 minutes. The output includes a time-series sequence of temperature setpoints and spatial uniformity requirements, humidity target ranges, and CO2 concentration metabolic demand thresholds, along with confidence scores for each predicted value reflecting the model's reliability assessment of the prediction results. The prediction results not only provide quantitative targets for future environmental parameters, but also provide a basis for the early adjustment of subsequent control strategies by using the trend change rate, such as the temperature needing to increase by 0.5℃ in the next 10 minutes, thus avoiding the lag of traditional passive control.

[0025] Step 103: Based on the prediction results, the control strategy is modeled as an MDP Markov decision process using the DRL deep reinforcement learning framework. With minimizing energy consumption and temperature fluctuations as the objective function, the power allocation of the heater, cooler, humidifier and fan is dynamically optimized to obtain the initial control strategy. Specifically, this embodiment defines the state space of the MDP Markov Decision Process, including the device state vector and prediction results. The action space is defined as heater power, cooler power, humidifier power, and fan speed, with each action subject to physical constraints. A reward function is established by integrating CO2 and humidity control effects, with energy consumption minimization and temperature fluctuation minimization as the core objectives. The DDPG deep deterministic policy gradient algorithm is employed, and the optimal control policy is learned through an Actor-Critic architecture. The Actor network is used to input the state and output continuous actions of actuator power, while the Critic network is used to evaluate the value of the current action and guide Actor optimization. Ornstein-Uhlenbeck noise balance exploration and utilization are incorporated during action selection, with noise intensity decreasing as training progresses. The actuator power sequence for the next 5 minutes is output to obtain the initial control policy.

[0026] The state space of the MDP Markov decision process is defined, including the equipment state vector and prediction results. The action space is defined as heater power, cooler power, humidifier power, and fan speed. Each action is physically constrained. To build an MDP model that fits the incubator control scenario, the state space needs to fully integrate the current equipment state and future demand trends. Specifically, it includes a 17-20 dimensional preprocessed equipment state vector, covering temperature spatial distribution, humidity fluctuation, CO2 concentration, airflow speed, and door state characteristics, as well as a 180-dimensional prediction result of environmental parameters for the next 60 minutes from the improved Transformer model, including the time series and confidence levels of temperature, humidity, and CO2. The two are concatenated through feature splicing to form a state space of approximately 200 dimensions, ensuring that the model can make decisions based on the current state and future trends. The motion space focuses on the controllable variables of the actuators, setting heater power (0-100%), cooler power (0-100%), humidifier power (0-100%), and fan speed (0-100%) as continuous motion dimensions, and imposing strict physical constraints: the heater and cooler cannot operate simultaneously to avoid energy conflict; the power switching rate is ≤20% / s to prevent actuator overload; the fan speed must be coordinated with temperature and humidity regulation, such as increasing the speed during humidification to evenly distribute water vapor; the upper limit of humidifier power is dynamically adjusted according to the current humidity; when the humidity is ≥90%RH, the power is forcibly reduced to below 30% to avoid condensation. These constraints are embedded in the model through an action masking mechanism to ensure that the output actions conform to the physical characteristics of the equipment.

[0027] With energy consumption minimization and temperature fluctuation minimization as the core objectives, a reward function is established by integrating CO2 and humidity control effects. The reward function design must balance control accuracy and energy efficiency, forming a multi-objective optimization orientation. The core terms include energy cost and temperature fluctuation penalty. The energy consumption term is quantified by converting the power of each actuator, such as heater power × rated power coefficient + cooler power × energy efficiency ratio. The temperature fluctuation term uses the absolute deviation between the current measured temperature and the set value, combined with the maximum temperature difference in the space to reflect uniformity through weighted calculation. Simultaneously, the difference between the measured and target values ​​of CO2 concentration deviation (DeltaCO_2_t) and humidity deviation are included as auxiliary terms, with priority distinguished by weighting coefficients: temperature weight β=0.4, energy consumption α=0.3, CO2 γ=0.2, and humidity δ=0.1, ensuring that core parameters are prioritized. The final reward function is defined as follows: [Further details on the reward function would be needed for a complete translation]. This transforms the minimization objective into maximizing the reward, guiding the agent to learn a low-energy-consumption + high-precision control strategy. In addition, for interference scenarios such as door opening, a reward and compensation mechanism is set up, such as temporarily reducing the energy consumption weight to 0.1 when the door is opened, prioritizing the rapid recovery of temperature, and enhancing the adaptability of the strategy.

[0028] The DDPG (Deep Deterministic Policy Gradient) algorithm is employed to learn the optimal control policy through an Actor-Critic architecture. Given the continuous power and rotational speed of the incubator actuators (0-100% continuous values), the DDPG algorithm is chosen as the reinforcement learning framework. Its core advantage lies in its ability to directly output deterministic continuous actions without discretization. The Actor-Critic architecture utilizes a dual-network collaborative learning approach: the Actor network generates the policy, taking the MDP state space as input, passing through three fully connected layers with hidden layers of dimensions 256 and 128, and using ReLU activation to map to the action space, outputting the specific power / rotational speed of each actuator; the Critic network evaluates the value, taking the state + action pair as input, and calculating the Q-value of the action (the expected long-term cumulative reward) using a similar network structure, serving as feedback for Actor optimization. To avoid overfitting and training instability, an experience replay pool is introduced to store historical states, actions, rewards, and next-state samples, using random sampling to break sample correlation. Simultaneously, a target Actor and a target Critic network are set, with a soft update rate τ=0.001 slowly tracking the main network parameters to ensure training stability. This architecture enables the model to efficiently explore the continuous action space and gradually approach the optimal control strategy.

[0029] The Actor network takes the state as input and outputs continuous actions, including actuator power. The Critic network evaluates the value of the current action and guides Actor optimization. The Actor network, acting as the policy function carrier, takes the complete state of the MDP (Multi-Device Programming Logic) including the device state vector and prediction results as input. After the high-dimensional features are compressed to 128 dimensions by the embedding layer, key features are extracted through a fully connected layer with residual connections containing BatchNorm to prevent gradient vanishing. Finally, the output layer activation function, using the Sigmoid function, constrains the action values ​​to the range of 0-100%, directly outputting continuous values ​​for the power of the heater, cooler, and humidifier, and the fan speed. The Critic network takes the current state and the Actor's output action as joint input. It captures the correlation between state and action features through a cross-attention mechanism, such as the synergistic effect of high humidity and low fan speed. A multilayer perceptron calculates the Q-value to quantify the long-term value of the action in the current state. During training, the Critic updates its parameters by minimizing the Q-value prediction error. The target Q-value = immediate reward + γ × target. The Critic then updates its parameters based on the Q-value of the next state it evaluates. The Actor, in turn, uses the Critic's Q-value gradient to adjust the network weights using the policy gradient theorem, maximizing the Q-value of the output action, thus forming a closed loop of generation-evaluation-optimization. This collaborative mechanism ensures that the actions generated by the Actor not only meet the requirements of the current state but also optimize the cumulative reward in the long run.

[0030] Ornstein-Uhlenbeck noise is incorporated into the action selection process for balance exploration and utilization. The noise intensity decays as training progresses, and the actuator power sequence for the next 5 minutes is output to obtain the initial control strategy. This balance strategy learning involves exploring new actions to discover better strategies and utilizing known effective actions. Ornstein-Uhlenbeck noise is superimposed on the actions output by the Actor. This noise is time-dependent, with the current noise correlated with the noise at the previous moment, better meeting the continuity requirements of incubator environmental control and avoiding abrupt action changes. Its mathematical model is: θ is the regression coefficient, σ is the noise intensity, and N is a normal distribution. Initially, a high noise intensity σ = 0.2 is set to encourage exploration of the effects of different power combinations. With each training iteration, σ decays exponentially at a rate of 0.95 every 1000 steps, eventually decreasing to 0.01, allowing the strategy to gradually converge to a stable solution. The model receives updated state and prediction results every 10 minutes and outputs the actuator power sequence for the next 5 minutes with a time step of 1 minute, for a total of 5 sets of action values, forming the initial control strategy. This sequence includes both optimization directions learned from historical data and retains a certain degree of exploratory nature to adapt to new scenarios, providing a benchmark for subsequent fine-tuning of fuzzy PID.

[0031] Step 104: Use the fuzzy PID algorithm to adjust the actuator power according to the initial control strategy, monitor sensor data in real time through the isolated forest model, and adjust the initial control strategy when an anomaly is detected to obtain the target control strategy.

[0032] Specifically, in this embodiment, a fuzzy PID algorithm is used with the deviation between the initial control strategy and the current measured value, and the rate of change of the deviation, as inputs, and the actuator power correction is output. An isolated forest model is used to monitor whether sensor data deviates from the normal distribution and whether actuator feedback is abnormal. The model is trained with historical normal operation data, and the abnormality score threshold is set to 0.8. If it is a minor abnormality, the strong correction mode of the fuzzy PID is activated, and the deviation is recorded for subsequent model calibration. If it is a severe abnormality, a safety strategy is immediately triggered and an audible and visual alarm is issued. The alarm information is pushed to the management personnel through the IoT module.

[0033] The fuzzy PID algorithm uses the deviation and rate of change of the initial control strategy from the current measured value as input, and outputs the actuator power correction. By combining the nonlinear processing capabilities of fuzzy logic with the precise adjustment characteristics of PID, the fuzzy PID algorithm achieves dynamic correction of the initial control strategy. The deviation between the target parameters output by the initial control strategy (such as target temperature and humidity) and the current measured values ​​(such as the measured value and the rate of change of the deviation), reflecting the trend of deviation change, is used as input to the fuzzy controller. These two values ​​are divided into seven fuzzy subsets (negative large, negative medium, negative small, zero, positive small, positive medium, positive large). Based on expert experience, 56 fuzzy rules are designed. For example, if the value is positive large and the value is positive small, the heater power needs to be significantly increased. The PID parameter correction is output using the centroid method through fuzzy inference and declarative processing. These corrections adjust the proportional, integral, and derivative coefficients of the PID controller in real time—for example, increasing them to speed up the response when the deviation is large and changes slowly, and increasing them to suppress oscillations when the deviation is small and fluctuates frequently. Ultimately, the actuator power correction range of ±10% of the fuzzy PID output is superimposed with the power value of the initial control strategy to form an execution command that takes into account both global optimization and local fine-tuning. This retains the energy consumption optimization advantage of the DRL strategy while eliminating local deviations through PID feedback adjustment, thereby improving control accuracy.

[0034] The isolated forest model monitors whether sensor data deviates from the normal distribution and whether actuator feedback is abnormal. The model is trained using historical normal operation data, with an anomaly score threshold set to 0.8. If the anomaly is slight, a strong correction mode of fuzzy PID is activated, and the deviation is recorded for subsequent model calibration. If the anomaly is severe, a safety policy is immediately triggered and an audible and visual alarm is issued, and the alarm information is pushed to management personnel via the IoT module. The isolated forest model achieves rapid detection of abnormal data by constructing multiple random binary trees: the training set is the normal operation data of the past 3 months, excluding special periods such as frequent door opening and closing and actuator maintenance. It covers the normal distribution characteristics of parameters such as temperature fluctuation of ±0.3℃, humidity fluctuation of ±3%RH, and CO2 concentration fluctuation of ±0.2%. The model constructs a tree structure by randomly selecting features and split points, and calculates the anomaly score for each sample. Normal data is isolated in deeper layers of the tree due to its dense distribution and has a low score, while abnormal data is isolated in shallower layers due to its deviation from the group and has a high score. The anomaly score threshold is set to 0.8. An abnormal response is triggered when the score of sensor data, such as a sudden temperature rise of 0.5℃ / s, or actuator feedback, such as a deviation of >20% between commanded power and actual output, exceeds the threshold. Mild anomaly scores are 0.8~0.9. For example, when there is a slow deviation caused by sensor drift, the fuzzy PID automatically switches to strong correction mode, increases the weight to 1.5 times the original, shortens the integration time, and quickly cancels the deviation. At the same time, the sensor data, control commands, and deviation values ​​during the abnormal period are stored in the log for use in updating the training data of Transformer and DRL for model calibration the next day. Severe anomaly scores are >0.9.

[0035] Its beneficial effects are as follows: 1. The improved Transformer model is used to predict trends in multi-parameter time-series data, anticipating changes in environmental requirements 60 minutes in advance, thus solving the lag problem of traditional control; the fuzzy PID algorithm combined with the DRL initial strategy achieves dual regulation, keeping temperature fluctuations within ±0.1℃, and improving the control accuracy of humidity and CO2 concentration to meet the requirements of high-precision cultivation. 2. The DRL framework uses minimizing energy consumption as the objective function, optimizing the power allocation of actuators such as heaters and coolers to avoid ineffective energy consumption; combined with parameter prediction, it achieves advance adjustment, reducing frequent start-stop of actuators, and experimental verification shows that it can reduce energy consumption and significantly reduce long-term operating costs. 3. The isolated forest model monitors sensor data and actuator status in real time, identifying anomalies within 1 second and triggering an adaptive adjustment strategy; in the case of mild anomalies, stability is maintained through fuzzy PID parameter correction, and in the case of severe anomalies, a safety mechanism is automatically activated and an alarm is triggered, reducing the risk of cultivation failure.

[0036] Please see Figure 2In a deep learning-based intelligent control method for a constant temperature incubator, based on the prediction results, the control strategy is modeled as an MDP Markov decision process using the DRL deep reinforcement learning framework. With minimizing energy consumption and temperature fluctuations as the objective function, the power allocation of the heater, cooler, humidifier, and fan is dynamically optimized to obtain the initial control strategy, which includes the following steps: Step 201: Use the DDPG deep deterministic policy gradient algorithm to learn the optimal control policy through the Actor-Critic architecture; Step 202: The Actor network is used to input the state and output the actuator power in a continuous action, while the Critic network is used to evaluate the value of the current action and guide the Actor optimization. Step 203: When selecting actions, Ornstein-Uhlenbeck noise balance exploration and utilization are added. The noise intensity decreases as the training process progresses, and the actuator power sequence for the next 5 minutes is output to obtain the initial control strategy.

[0037] The above describes an embodiment of the intelligent control method for a constant temperature incubator based on deep learning according to the present invention. Please refer to [link / reference]. Figure 3 In a deep learning-based intelligent control system for a constant temperature incubator, the system includes the following modules: The state vector acquisition module is used to collect real-time data on temperature, humidity, CO2 concentration, airflow speed and door status inside the incubator, and preprocess the collected data to obtain the device state vector. The equipment status prediction module is used to analyze the equipment status vector using an improved Transformer model, predict the environmental parameter demand trend during the cultivation process, and obtain the prediction results. The control strategy generation module is used to model the control strategy as an MDP Markov decision process based on the prediction results through the DRL deep reinforcement learning framework. With minimizing energy consumption and temperature fluctuation as the objective function, it dynamically optimizes the power allocation of heater, cooler, humidifier and fan to obtain the initial control strategy. The control strategy optimization module is used to adjust the actuator power according to the initial control strategy using the fuzzy PID algorithm, monitor sensor data in real time through the isolated forest model, and adjust the initial control strategy when an anomaly is detected to obtain the target control strategy.

[0038] 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 claims and their equivalents.

Claims

1. A deep learning-based intelligent control method for constant-temperature incubators, characterized in that, The intelligent control method of the constant-temperature incubator based on deep learning comprises the following steps: Real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator, preprocessing of the collected data, and obtaining of an equipment state vector; Analysis of the equipment state vector by using an improved Transformer model, prediction of the environmental parameter demand trend in the cultivation process, and obtaining of a prediction result; According to the prediction result, modeling the control strategy as an MDP Markov decision process through a DRL deep reinforcement learning framework, taking the minimization of energy consumption and temperature fluctuation as the objective function, dynamically optimizing the power distribution of the heater, cooler, humidifier and fan, and obtaining an initial control strategy; Adjusting the actuator power according to the initial control strategy by using a fuzzy PID algorithm, real-time monitoring of sensor data by using an isolation forest model, adjusting the initial control strategy when an anomaly is detected, and obtaining a target control strategy.

2. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, characterized in that, The real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator, preprocessing of the collected data, and obtaining of an equipment state vector comprises: Real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator; Using a sliding window mean filter to eliminate sensor fault values and short-time noise and mapping multi-source heterogeneous data to a unified range to obtain processed data; Extracting the time series features of the processed data, integrating the preprocessed temperature, humidity, CO2 concentration, air flow speed and door state into a multi-dimensional vector, and obtaining an equipment state vector.

3. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, characterized in that, The analysis of the equipment state vector by using an improved Transformer model, the prediction of the environmental parameter demand trend in the cultivation process, and the obtaining of a prediction result comprise: Adding a space-time attention mechanism in the Transformer model, wherein the space attention analyzes the spatial dependence of the sensor network, the time attention captures long-period rules, and a sine-cosine position encoding is used to enhance the model to improve the time series data time sequence perception, thereby obtaining an improved Transformer model; Analysis of the equipment state vector by using an improved Transformer model, prediction of the environmental parameter demand trend in the cultivation process, and obtaining of a prediction result.

4. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, characterized in that, According to the prediction result, modeling the control strategy as an MDP Markov decision process through a DRL deep reinforcement learning framework, taking the minimization of energy consumption and temperature fluctuation as the objective function, dynamically optimizing the power distribution of the heater, cooler, humidifier and fan, and obtaining an initial control strategy, which comprises: Defining the state space of the MDP Markov decision process, including the equipment state vector and the prediction result, and defining the action space as the heater power, cooler power, humidifier power and fan speed, each action being subjected to physical constraints; Taking the minimization of energy consumption and temperature fluctuation as the core, and establishing a reward function by comprehensively considering the CO2 and humidity control effect.

5. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, characterized in that, The control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework according to the prediction result, a power distribution of a heater, a cooler, a humidifier and a fan is dynamically optimized to minimize energy consumption and temperature fluctuation as an objective function, and an initial control strategy is obtained, including: An optimal control strategy is learned through an Actor-Critic architecture by using a DDPG deep deterministic policy gradient algorithm; The Actor network is used for inputting a state and outputting a continuous action of an actuator power, and the Critic network is used for evaluating a value of a current action and guiding the Actor optimization; When an action is selected, Ornstein-Uhlenbeck noise is added to balance exploration and utilization, the noise intensity is attenuated with the training process, a future 5-minute actuator power sequence is output, and the initial control strategy is obtained.

6. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, wherein, The initial control strategy is adjusted according to the actuator power by using a fuzzy PID algorithm, sensor data is monitored in real time through an Isolation Forest model, the initial control strategy is adjusted when an anomaly is detected, and a target control strategy is obtained, including: The deviation and deviation change rate of the initial control strategy and the current measured value are input into the fuzzy PID algorithm, and an actuator power correction amount is output.

7. The intelligent control method of a constant-temperature incubator based on deep learning according to claim 1, characterized in that, The initial control strategy is adjusted according to the actuator power by using a fuzzy PID algorithm, sensor data is monitored in real time through an Isolation Forest model, the initial control strategy is adjusted when an anomaly is detected, and a target control strategy is obtained, including: Whether the sensor data deviates from a normal distribution and an actuator feedback is abnormal is monitored through the Isolation Forest model, the model is trained with historical normal operation data, and an abnormal score threshold is set to 0.8; If it is a mild anomaly, a strong correction mode of the fuzzy PID is enabled, and the deviation is recorded for subsequent model calibration, if it is a severe anomaly, a safety strategy is triggered immediately and an audible and visual alarm is sent, and alarm information is pushed to a manager through an Internet of Things module.

8. A deep learning-based intelligent control system for constant-temperature incubators, characterized in that, The intelligent control system of the constant-temperature incubator based on deep learning includes the following modules: A state vector acquisition module is configured to collect temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator in real time, pre-process the collected data, and obtain a device state vector; A device state prediction module is configured to analyze the device state vector by using an improved Transformer model, predict a demand trend of environmental parameters in the incubation process, and obtain a prediction result; A control strategy generation module is configured to model a control strategy as an MDP Markov decision process through a DRL deep reinforcement learning framework according to the prediction result, dynamically optimize a power distribution of a heater, a cooler, a humidifier and a fan to minimize energy consumption and temperature fluctuation as an objective function, and obtain an initial control strategy; A control strategy optimization module is configured to adjust an actuator power according to the initial control strategy by using a fuzzy PID algorithm, monitor sensor data in real time through an Isolation Forest model, adjust the initial control strategy when an anomaly is detected, and obtain a target control strategy.

9. The intelligent control system for constant-temperature incubator based on deep learning according to claim 8, characterized in that, The control strategy optimization module includes the following sub-modules: The correction submodule is configured to utilize a fuzzy PID algorithm to take the deviation and the rate of change of the deviation between the initial control strategy and the current measured value as input, and output a power correction amount of the actuator.

10. The intelligent control system of constant-temperature incubator based on deep learning according to claim 8, characterized in that, The control strategy optimization module comprises the following submodules: The judgment submodule is configured to monitor whether the sensor data deviates from the normal distribution and the actuator feedback is abnormal through an isolation forest model, train the model with historical normal operation data, and set the abnormal score threshold to 0.

8. The control submodule is configured to enable a strong correction mode of the fuzzy PID if the abnormality is mild, and record the deviation for subsequent model calibration; if the abnormality is severe, trigger a safety strategy immediately and issue an audible and visual alarm, and push the alarm information to the manager through the Internet of Things module.

Images