A deep learning-based intelligent control method and system for constant-temperature incubators

Abstract

The application discloses a kind of intelligent control method and system of constant-temperature incubator based on deep learning, by real-time acquisition equipment state vector;Utilize the improved model analysis the equipment state vector, predict the environmental parameter demand trend in culture process, obtain prediction result;According to the prediction result, the control strategy is modeled as MDP Markov decision process by DRL deep reinforcement learning framework, to minimize energy consumption and temperature fluctuation as objective function, dynamically optimize the power distribution of heater, refrigerator, humidifier and fan, obtain initial control strategy;Using fuzzy PID algorithm adjusts actuator power according to initial control strategy, real-time monitoring sensor data by isolated forest model, when detecting abnormality, adjust initial control strategy, obtain target control strategy.Predict in advance environmental demand change, solved the hysteresis problem of traditional control.

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 depends 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 as follows: 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:

[0005] 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 the equipment state vector;

[0006] The improved Transformer model is used to analyze the equipment state vector, predict the environmental parameter demand trend in the culture process, and obtain the prediction result;

[0007] According to the prediction result, the control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework, the power distribution of the heater, the cooler, the humidifier and the fan is dynamically optimized to minimize the energy consumption and the temperature fluctuation as the objective function, and the initial control strategy is obtained;

[0008] The fuzzy PID algorithm is used to adjust the actuator power according to the initial control strategy, and the isolated forest model is used to monitor the sensor data in real time. When an abnormality is detected, the initial control strategy is adjusted, and the target control strategy is obtained.

[0009] 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 state data in the incubator, the preprocessing of the collected data, and the obtaining of the equipment state vector comprise:

[0010] Real-time acquisition of temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator;

[0011] The sliding window mean filtering is used to eliminate sensor fault values and short-time noise, and the multi-source heterogeneous data is mapped to a unified range to obtain processing data.

[0012] The time sequence features of the processing 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 a device state vector.

[0013] 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 device state vector, predict the demand trend of environmental parameters in the cultivation process, and obtain a prediction result, including:

[0014] The space-time attention mechanism is added to the Transformer model, wherein the space attention analyzes the spatial dependence of the sensor network, and the time attention captures the long-period law, and the 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.

[0015] The improved Transformer model is used to analyze the device state vector, predict the demand trend of environmental parameters in the cultivation process, and obtain a prediction result.

[0016] Further, in the above-mentioned intelligent control method for constant temperature incubator based on deep learning, the DRL deep reinforcement learning framework is used to model the control strategy as an MDP Markov decision process according to the prediction result, and the power distribution of the heater, cooler, humidifier and fan is dynamically optimized to minimize the energy consumption and temperature fluctuation as the objective function, thereby obtaining an initial control strategy, including:

[0017] The state space of the MDP Markov decision process is defined, including the device state vector and the prediction result, and the action space is defined as the heater power, cooler power, humidifier power and fan speed, and each action is subjected to physical constraints.

[0018] The reward function is established by taking the minimization of energy consumption and temperature fluctuation as the core and comprehensively considering the control effect of CO2 and humidity.

[0019] Further, in the above-mentioned intelligent control method for constant temperature incubator based on deep learning, the DRL deep reinforcement learning framework is used to model the control strategy as an MDP Markov decision process according to the prediction result, and the power distribution of the heater, cooler, humidifier and fan is dynamically optimized to minimize the energy consumption and temperature fluctuation as the objective function, thereby obtaining an initial control strategy, including:

[0020] The DDPG deep deterministic policy gradient algorithm is used to learn the optimal control strategy through the Actor-Critic architecture.

[0021] The Actor network is used for continuous actions that take the state as input and output the actuator power, while the Critic network is used to evaluate the value of the current action and guide the optimization of the Actor.

[0022] When selecting actions, Ornstein-Uhlenbeck noise balance exploration and utilization are incorporated. 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.

[0023] Furthermore, in the aforementioned intelligent control method for a constant temperature incubator based on deep learning, the step of adjusting the actuator power according to the initial control strategy using a fuzzy PID algorithm, monitoring sensor data in real time through an isolated forest model, and adjusting the initial control strategy when an anomaly is detected to obtain the target control strategy includes:

[0024] The 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 as output.

[0025] Furthermore, in the aforementioned intelligent control method for a constant temperature incubator based on deep learning, the step of adjusting the actuator power according to the initial control strategy using a fuzzy PID algorithm, monitoring sensor data in real time through an isolated forest model, and adjusting the initial control strategy when an anomaly is detected to obtain the target control strategy includes:

[0026] The isolated forest model was used to monitor whether sensor data deviated from the normal distribution and whether actuator feedback was abnormal. The model was trained with historical normal operation data and the abnormality score threshold was set to 0.8.

[0027] If the anomaly is minor, the strong correction mode of fuzzy PID is activated, and the deviation is recorded for subsequent model calibration; if the anomaly is severe, the safety policy is immediately triggered and an audible and visual alarm is issued, and the alarm information is pushed to the management personnel through the IoT module.

[0028] Furthermore, in a deep learning-based intelligent control system for a constant temperature incubator, the deep learning-based intelligent control system for a constant temperature incubator includes the following modules:

[0029] 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. The collected data is preprocessed to obtain the device state vector.

[0030] 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.

[0031] The control strategy generation module is configured to model a control strategy as a Markov decision process (MDP) through a deep reinforcement learning (DRL) framework based on the prediction result, dynamically optimize power distribution of the heater, the cooler, the humidifier, and the fan with a target function of minimizing energy consumption and temperature fluctuation, and obtain an initial control strategy.

[0032] The control strategy optimization module is configured to adjust the power of the actuator 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.

[0033] Further, in the intelligent control system for the constant-temperature incubator based on deep learning, the control strategy optimization module comprises the following sub-modules:

[0034] The correction sub-module is configured to use a fuzzy PID algorithm to take the deviation of the initial control strategy from a current measured value and a rate of change of the deviation as input, and output a correction amount of the power of the actuator.

[0035] Further, in the intelligent control system for the constant-temperature incubator based on deep learning, the control strategy optimization module comprises the following sub-modules:

[0036] The judgment sub-module is configured to monitor whether the sensor data deviates from a normal distribution and whether the actuator feedback is abnormal through an isolation forest model, train the model with historical normal operation data, and set an anomaly score threshold to 0.8.

[0037] The control sub-module is configured to enable a strong correction mode of the fuzzy PID if the anomaly is mild, and record the deviation for subsequent model calibration; and immediately trigger a safety strategy and issue an audible and visual alarm if the anomaly is severe, and push alarm information to a manager through an Internet of Things module.

[0038] The beneficial effects are as follows: 1. The improved Transformer model is used to predict the trend of multi-parameter time series data, to predict the change in environmental demand 60 minutes in advance, and to solve the hysteresis problem of traditional control. The fuzzy PID algorithm combines the initial strategy of DRL to achieve double regulation, so that the temperature fluctuation is controlled within ±0.1℃, the control accuracy of humidity and CO2 concentration is improved, and the high-precision cultivation demand is met. 2. The DRL framework takes minimizing energy consumption as a target function, optimizes the power distribution of the heater, cooler, and other actuators, avoids invalid energy consumption, and realizes advance adjustment in combination with parameter prediction, reduces the frequent start-stop of the actuator, and reduces energy consumption and long-term operation cost through experimental verification. 3. The isolation forest model monitors sensor data and actuator state in real time, can identify an anomaly within 1 second, and triggers an adaptive adjustment strategy. When the anomaly is mild, the fuzzy PID parameter correction is used to maintain stability, and when the anomaly is severe, a safety mechanism is automatically started and an alarm is issued, thereby reducing the risk of cultivation failure. BRIEF DESCRIPTION OF DRAWINGS

[0039] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the application.

[0040] Figure 1 A first embodiment schematic diagram of a deep learning-based intelligent control method for a constant-temperature incubator according to an embodiment of the present application;

[0041] Figure 2 A second embodiment schematic diagram of a deep learning-based intelligent control method for a constant-temperature incubator according to an embodiment of the present application;

[0042] Figure 3 A first embodiment schematic diagram of a deep learning-based intelligent control system for a constant-temperature incubator according to an embodiment of the present application. EMBODIMENTS

[0043] In order to make the objects, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and should not be used to limit the present application.

[0044] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "one", "a", and "the" as used herein can include plural forms. It should be further understood that the use of the terms "comprise", "comprises", "comprising", "include", "includes", "including", "contain", "contains", "containing", "have", "has", "having", or "comprised of" in the specification of the present application means that there is existence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0045] The present application will be described below with reference to the accompanying drawings, Figure 1 As shown in the drawings, a deep learning-based intelligent control method for a constant-temperature incubator includes the following steps:

[0046] In step 101, temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator are collected in real time, the collected data is preprocessed, and a device state vector is obtained. Specifically, in this embodiment, temperature, humidity, CO2 concentration, air flow speed and door state data in the incubator are collected in real time. A sliding window mean filter is used to eliminate 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, the preprocessed temperature, humidity, CO2 concentration, air flow speed and door state are integrated into a multi-dimensional vector, and a device state vector is obtained.

[0047] To achieve comprehensive perception of the incubator environment, multiple types of high-precision sensors need to be deployed and a synchronous acquisition mechanism needs to be built: temperature acquisition uses 3-5 distributed platinum resistance sensors PT1000, covering key areas such as the top, middle, and sample area of the incubator, to ensure capturing the spatial temperature gradient; humidity monitoring uses a capacitive sensor installed in a stable airflow area to avoid direct interference from the humidifier, with an accuracy of ±2% RH; CO2 concentration is detected in real time by a non-dispersive infrared (NDIR) sensor with a sampling frequency of 1 Hz to track changes in concentration caused by culture metabolism; air flow speed is collected by a thermal anemometer near the fan outlet to reflect the uniformity of the circulating airflow; and the door status is sampled at a high frequency of 10 Hz by a magnetic proximity switch to accurately capture transient changes in the switch. All sensor data is received by an embedded control system such as the STM32H7 series, which relies on a hardware clock to synchronize timestamps and ensure that the time deviation of multi-source data is ≤10 ms, providing a spatiotemporal consistent raw data basis for subsequent analysis.

[0048] Sliding window mean filtering is used to remove sensor fault values and short-term noise and map multi-source heterogeneous data to a unified range to obtain processed data. For random noise such as instantaneous jumps caused by electronic interference and fault values such as abnormal readings generated by sensor offline in the sensor raw data, sliding window mean filtering is used for preprocessing: setting the window size to 5-10 sampling points corresponding to 5-10 seconds, adapting to the 1 Hz conventional sampling frequency, replacing the current point value with the mean value within the window to smooth short-term fluctuations; for obvious outliers such as a temperature rise of more than 10°C, first mark as fault points, then use linear interpolation of valid values at both ends of the window to fill in, avoiding contamination by fault data. At the same time, due to the significant differences in the physical dimensions of each parameter such as temperature in °C and CO2 in %, standardization processing is needed to map to a unified range: based on historical normal operation data, calculate the mean μ and standard deviation σ of each parameter, convert the data to standardized values with a mean of 0 and a variance of 1, eliminate the influence of magnitude differences on subsequent model analysis, and ensure that temperature, humidity, and other parameters are balanced in the feature space.

[0049] The time sequence features of the processed data are extracted, and the preprocessed temperature, humidity, CO2 concentration, air flow speed and door body state are integrated into a multi-dimensional vector to obtain the equipment state vector. On the basis of the preprocessed data, key time sequence features need to be extracted to enrich the state description. For temperature, the mean value of the distributed sensor reflects the overall temperature level, the maximum deviation reflects the spatial uniformity, and the 5-minute sliding window change rate reflects the temperature rising / trending. The humidity, CO2 concentration and air flow speed are extracted respectively, and the current value, 10-minute cumulative change amount reflects the long-term trend and the 30-second fluctuation amplitude reflects the stability. In addition to retaining the real-time on / off state 0 / 1, the door body state increases the number of switchings within the last 5 minutes and the cumulative opening time to reflect the degree of environmental disturbance. Then, these features are integrated with the time labels of the culture process, such as the culture time, the stage: preheating / stable / ending, to form a multi-dimensional vector containing temperature features, humidity features, CO2 features, air flow features, door body state features and time labels. Finally, a 17-20-dimensional equipment state vector is constructed to comprehensively describe the real-time running state of the incubator, providing structured input for subsequent model prediction and strategy optimization.

[0050] Step 102, using the improved Transformer model to analyze the equipment state vector, predicting the trend of environmental parameter demand in the culture process, obtaining the prediction result;

[0051] Specifically, in this embodiment, a space-time attention mechanism is added to the Transformer model, where the spatial attention analyzes the spatial dependence of the sensor network, and the time attention captures the long-period law. At the same time, a sine-cosine position coding is used to enhance the model's time sequence perception of time series data, obtaining an improved Transformer model. The improved Transformer model is used to analyze the equipment state vector, predict the trend of environmental parameter demand in the culture process, and obtain the prediction result. A space-time attention mechanism is added to the Transformer model, where the spatial attention analyzes the spatial dependence of the sensor network, and the time attention captures the long-period law. At the same time, a sine-cosine position coding is used to enhance the model's time sequence perception of time series data, obtaining an improved Transformer model.

[0052] To improve the ability of the Transformer model to analyze the multi-source heterogeneous time series data of the incubator, the model is improved by fusing the spatio-temporal attention mechanism and position encoding: the spatial attention module is aimed at the layout characteristics of the distributed sensor network, constructs a sensor position correlation matrix, and takes the physical coordinates of temperature probes and humidity sensors in different areas as the basis to capture the spatial dependence relationship by calculating the attention weight, for example, the numerical correlation degree between the top temperature sensor and the sample area temperature sensor, and the spatial coupling between the fan outlet air speed and the humidity distribution in the incubator, so that the model can identify the spatial conduction law of environmental parameters; the time attention module adopts a multi-layer multi-head structure, and by setting different attention windows such as a short-term window of 30 minutes and a long-term window of 24 hours, it can capture long-period rules, such as the influence of diurnal environmental temperature fluctuations on the incubator interior and the periodic accumulation trend of CO2 concentration in the culture metabolic process; at the same time, the sine-cosine position encoding is introduced to add a time sequence label to each time step, so that the model can distinguish the sequence of different time points and enhance the dynamic change perception of time series data. Through the above improvements, the model can analyze the spatial distribution characteristics and time evolution law of environmental parameters at the same time, and solve the problem of insufficient capture of spatial correlation of multi-source sensor data by the standard Transformer.

[0053] Using the improved Transformer model to analyze the equipment state vector and predict the trend of environmental parameter demand in the cultivation process to obtain a prediction result

[0054] Based on the deep analysis of the device state vector by the improved Transformer model, the precise prediction of the environmental parameter demand trend in the cultivation process is realized: the device state vector with dimensions 17-20 containing temperature, humidity, CO2 concentration, air flow speed, door status and other multi-dimensional characteristics is input into the model. First, the embedding layer is used to convert discrete features such as door status labels into continuous vectors, which are concatenated with continuous features such as temperature mean to form the model input sequence. When processed by the spatio-temporal attention mechanism, the spatial attention layer focuses on the spatial correlation of sensor data, such as the spatial response relationship between sample area temperature and heater power, and the temporal attention layer excavates long-period patterns in historical data, such as the temperature demand surge feature in a specific cultivation stage. Combined with the time sequence information provided by the sine-cosine position encoding, the model can analyze the dynamic coupling rules between parameters. The decoder part uses the autoregressive generation mode to predict the environmental parameter demand trend in the next 60 minutes step by step, with the current and historical state vectors as conditions, and outputs the time sequence sequence including temperature set value and spatial uniformity requirement, humidity target range, and CO2 concentration metabolic demand threshold, while also providing confidence scores for each predicted value to reflect the reliability evaluation of the model's prediction results. This prediction result not only provides quantitative targets for future environmental parameters, but also provides a basis for early adjustment of subsequent control strategies through trend change rates such as a temperature increase of 0.5°C in the next 10 minutes, avoiding the lag of traditional passive control.

[0055] Step 103, according to the prediction result, the control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework, with the objective function of minimizing energy consumption and temperature fluctuation, dynamically optimizing the power distribution of the heater, cooler, humidifier and fan, and obtaining an initial control strategy.

[0056] Specifically, in this embodiment, the state space of the MDP Markov decision process is defined, including the device state vector and the prediction result, and the action space is defined as the heater power, cooler power, humidifier power and fan speed, each action is subject to physical constraints; taking energy consumption minimization and temperature fluctuation minimization as the core, a reward function is established by combining CO2 and humidity control effect. The DDPG deep deterministic policy gradient algorithm is used to learn the optimal control strategy through the Actor-Critic architecture; the Actor network is used to input the state and output the continuous action 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; Ornstein-Uhlenbeck noise is added during action selection to balance exploration and utilization, and the noise intensity decays with the training process, outputting the actuator power sequence in the next 5 minutes to obtain the initial control strategy.

[0057] The state space of the MDP Markov Decision Process is defined, including the device state vector and the prediction result, and the action space is defined as the heater power, the cooler power, the humidifier power and the fan speed, each action is physically constrained, in order to build the MDP model that fits the incubator control scene, the state space needs to fully integrate the current device state and the future demand trend: specifically, it includes the pre-processed device state vector 17~20 dimensions, covering temperature spatial distribution, humidity fluctuation, CO2 concentration, air flow speed and door body state characteristics, and the future 60-minute environmental parameter prediction result output by the improved Transformer model 180 dimensions, including temperature, humidity, CO2 time series and confidence, both of which are spliced into a state space with a dimension of about 200, ensuring that the model can make decisions based on the current state + future trend. The action space focuses on the controllable variables of the actuator, setting the heater power 0~100%, the cooler power 0~100%, the humidifier power 0~100%, and the fan speed 0~100% as continuous action dimensions, and imposing strict physical constraints: the heater and the cooler cannot work at the same time to avoid energy offset, the power switching rate ≤20% / s to prevent the actuator from being overloaded; the fan speed needs to be coordinated with temperature and humidity regulation, such as increasing the speed to evenly distribute water vapor when humidifying; the upper limit of the humidifier power is dynamically adjusted according to the current humidity, and the humidity ≥90%RH, the power is forced to drop below 30% to avoid dew. These constraints are embedded in the model through the action mask mechanism to ensure that the output action meets the physical characteristics of the device.

[0058] With the minimization of energy consumption and temperature fluctuation as the core, the reward function is established by comprehensively considering the control effect of CO2 and humidity. The design of the reward function needs to balance the control accuracy and energy efficiency, forming a multi-objective optimization guide: the core items include energy cost and temperature fluctuation penalty, where the energy item is quantified by converting the power of each actuator, such as heater power × rated power coefficient + cooler power × energy efficiency correction; the temperature fluctuation item uses the absolute deviation of the current measured temperature and the set value, and reflects the uniformity by combining the maximum temperature difference in the space and weighted calculation. At the same time, the CO2 concentration deviation ΔCO_2_t and the humidity deviation are included as auxiliary items, and the priority is distinguished by the weight coefficient: temperature weight β=0.4, energy α=0.3, CO2 γ=0.2, humidity δ=0.1, to ensure that the core parameters are met first. The final reward function is defined as R_t = βΔT_t + αE_t + γΔCO_2_t + δΔH_t, where the minimization goal is converted into a maximization reward, guiding the agent to learn a control strategy with low energy consumption and high accuracy. In addition, for interference scenes such as door opening, a reward compensation mechanism is set, such as temporarily reducing the energy consumption weight to 0.1 when the door is opened, to prioritize temperature recovery and enhance the adaptability of the strategy.

[0059] DDPG deep deterministic policy gradient algorithm is adopted to learn the optimal control policy through the Actor-Critic architecture. Since the power and speed of the incubator actuators are continuous values from 0 to 100%, the DDPG algorithm is selected as the reinforcement learning framework, which has the core advantage of directly outputting deterministic continuous actions without the need for discretization. The Actor-Critic architecture learns through the cooperation of double networks: the Actor network is responsible for policy generation, taking the MDP state space as input, passing through a 3-layer fully connected network with hidden layer dimensions of 256 and 128, and using ReLU as the activation function to map to the action space, outputting the specific power / speed of each actuator; the Critic network is responsible for value evaluation, inputting the state+action pair, calculating the Q value of the action, i.e., the expected long-term cumulative reward, as the feedback signal for Actor optimization. To avoid overfitting and unstable training, an experience replay pool is introduced to store historical state, action, reward, and next state samples, and random sampling is used to break the sample correlation; at the same time, a target Actor and Critic network are set up, with a soft update rate τ = 0.001 to slowly track the main network parameters, ensuring the stability of the training process. This architecture enables the model to efficiently explore in continuous action space and gradually approach the optimal control policy.

[0060] The Actor network is used to input the state and output the continuous action 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; the Actor network serves as the carrier of the policy function, with its input being the complete state of the MDP containing the device state vector and the prediction result, which is compressed to 128 dimensions after embedding, then the key features are extracted through the fully connected layer with BatchNorm to prevent gradient disappearance, and finally the action value is constrained to the range of 0-100% through the output layer activation function with Sigmoid, directly outputting the continuous values of the heater, cooler, humidifier power and fan speed. The Critic network takes the current state and the action output by the Actor as the joint input, captures the relevance between the two through the cross-attention mechanism of state features and action features, such as the synergistic effect of high humidity and low fan speed, calculates the Q value through the multilayer perceptron, and quantifies the long-term value of the action in the current state. During training, the Critic updates the parameters by minimizing the Q value prediction error target Q value = immediate reward + γ × target Critic evaluation of next state Q value; the Actor adjusts the network weights by using the Q value gradient of the Critic through the policy gradient theorem to maximize the Q value of the output action, forming a closed loop of generation-evaluation-optimization. This cooperative mechanism ensures that the action generated by the Actor not only meets the current state requirements, but also optimizes the cumulative reward in the long term.

[0061] The Ornstein-Uhlenbeck noise is added in the action selection to balance the exploration and utilization, the noise intensity is attenuated with the training process, the actuator power sequence of the next 5 minutes is output, and the initial control strategy is obtained. The exploration tries new actions to find better strategies in the balance strategy learning, and the known effective actions are used. The Ornstein-Uhlenbeck noise is superimposed on the action output by the actor: the noise has time correlation, the current noise is related to the noise of the last moment, which is more in line with the continuity requirement of the incubator environment control to avoid action mutation. Its mathematical model is θ is the regression coefficient, σ is the noise intensity, and N is the normal distribution. A higher noise intensity σ=0.2 is set in the early training to encourage the exploration of the effects of different power combinations; every 1000 steps of training iteration, σ is exponentially attenuated with an attenuation rate of 0.95, and finally reduced to 0.01, so that the strategy gradually converges to a stable solution. The model receives the updated state and prediction result every 10 minutes, and outputs the actuator power sequence of the next 5 minutes. The time step is 1 minute, there are 5 groups of action values, and the initial control strategy is formed. The sequence contains both the optimization direction learned based on historical data and a certain exploration to adapt to new scenarios, providing a benchmark for the fine adjustment of the fuzzy PID.

[0062] Step 104, adjusting the actuator power according to the initial control strategy using the fuzzy PID algorithm, and monitoring the sensor data in real time through the isolation forest model, adjusting the initial control strategy when an anomaly is detected, and obtaining the target control strategy.

[0063] Specifically, in the present embodiment, the fuzzy PID algorithm is used to take the deviation and deviation change rate of the initial control strategy and the current measured value as input, and output the actuator power correction amount. 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, the 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.

[0064] The fuzzy PID algorithm takes the deviation and the rate of change of the deviation between the initial control strategy and the current measured value as input, and outputs the actuator power correction amount. The fuzzy PID algorithm combines the nonlinear processing capability of fuzzy logic and the precise adjustment characteristics of PID to dynamically modify 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 value, as well as the rate of change of the deviation, which reflects the trend of the deviation, are taken as inputs to the fuzzy controller. These inputs are divided into seven fuzzy subsets (negative large, negative medium, negative small, zero, positive small, positive medium, and positive large). Based on expert experience, 56 fuzzy rules are designed. For example, if the deviation is positive large and the rate of change is positive small, the heater power needs to be increased significantly. Through fuzzy reasoning and defuzzification, the gravity method is used to output the PID parameter correction amount. These correction amounts adjust the proportional, integral, and derivative coefficients of the PID controller in real time. For example, when the deviation is large and changes slowly, the proportional coefficient is increased to speed up the response. When the deviation is small and fluctuates frequently, the integral coefficient is increased to suppress oscillation. Finally, the actuator power correction amount output by the fuzzy PID, which is in the range of ±10%, is added to the power value of the initial control strategy to form an execution instruction that considers both global optimization and local fine-tuning. This approach retains the energy optimization advantages of the DRL strategy while eliminating local deviations through PID feedback adjustment to improve control accuracy.

[0065] The isolated 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 abnormality, the strong correction mode of the fuzzy PID is enabled, and the deviation is recorded for subsequent model calibration. If it is a severe abnormality, the safety strategy is triggered immediately and an audible and visual alarm is issued. The Internet of Things module pushes the alarm information to the management personnel. The isolated forest model achieves rapid detection of abnormal data by constructing multiple random binary trees. The training set is obtained by excluding special periods such as frequent opening and closing of the door body and actuator maintenance from the past three months of normal operation data. It covers the normal distribution characteristics of parameters such as temperature ±0.3℃ fluctuation, humidity ±3%RH fluctuation, and CO2 concentration ±0.2% fluctuation. The model constructs a tree structure by randomly selecting features and splitting points, and calculates the abnormal score of each sample. Normal data is isolated in the deeper layer of the tree with a low score due to dense distribution, while abnormal data is isolated in the shallower layer with a high score due to deviation from the group. The abnormal score threshold is set to 0.8. When the sensor data, such as a sudden temperature rise of 0.5℃ / s, or the actuator feedback, such as the deviation between the instruction power and the actual output, exceeds 20%, triggers an abnormal response when the score exceeds the threshold. Mild abnormality score 0.8~0.9, such as slow deviation caused by sensor drift, the fuzzy PID automatically switches to the strong correction mode, increases the weight to 1.5 times the original, and shortens the integral time to quickly offset the deviation. At the same time, the sensor data, control instructions, and deviation values during the abnormal period are stored in the log for the next day's model calibration to update the training data of Transformer and DRL. Severe abnormality score >0.9.

[0066] Its beneficial effects are that 1, the improved Transformer model is used for trend prediction of multi-parameter time series data, and the environmental demand change is predicted 60 minutes in advance, solving the hysteresis problem of traditional control; the fuzzy PID algorithm combines the initial strategy of DRL to realize double regulation, so that the temperature fluctuation is controlled within ±0.1℃, the humidity and CO2 concentration control precision is improved, and the high-precision culture demand is met. 2, the DRL framework takes minimizing energy consumption as the objective function, optimizes the power distribution of the heater, cooler and other actuators to avoid invalid energy consumption; combined with parameter prediction to realize advance regulation, reduce the frequent start-stop of the actuator, and experimentally verified to reduce energy consumption and significantly reduce long-term operation cost. 3, the isolation forest model monitors the sensor data and actuator state in real time, can identify abnormalities within 1 second, and trigger adaptive adjustment strategy; when the abnormality is slight, the fuzzy PID parameter correction is used to maintain stability, and when the abnormality is severe, the safety mechanism is automatically started and an alarm is given, reducing the risk of culture failure.

[0067] Please refer to Figure 2 In a deep learning-based intelligent control method of constant-temperature incubator, according to the prediction result, the control strategy is modeled as an MDP Markov decision process through a DRL deep reinforcement learning framework, the power distribution of the heater, cooler, humidifier and fan is dynamically optimized to minimize energy consumption and temperature fluctuation, and the initial control strategy includes the following steps:

[0068] Step 201, using DDPG deep deterministic policy gradient algorithm, learning optimal control strategy through Actor-Critic architecture;

[0069] Step 202, the Actor network is used for input state and output continuous action of actuator power, and the Critic network is used for evaluating the value of current action to guide Actor optimization;

[0070] Step 203, when the action is selected, Ornstein-Uhlenbeck noise is added to balance exploration and utilization, the noise intensity decays with the training process, the actuator power sequence of the next 5 minutes is output, and the initial control strategy is obtained.

[0071] The above describes an embodiment of the deep learning-based intelligent control method of constant-temperature incubator, please refer to Figure 3 In a deep learning-based intelligent control system of constant-temperature incubator, the deep learning-based intelligent control system of constant-temperature incubator includes the following modules:

[0072] The state vector acquisition module is used for real-time acquisition of temperature, humidity, CO2 concentration, air flow velocity and door body state data in the incubator, and pre-processing of the collected data to obtain the device state vector;

[0073] a device state prediction module configured to analyze the device state vector using an improved Transformer model to predict a trend of the environmental parameter requirement in the cultivation process and obtain a prediction result;

[0074] a control strategy generation module configured to model a control strategy as a Markov decision process (MDP) based on the prediction result through a deep reinforcement learning (DRL) framework, and dynamically optimize power distribution of a heater, a cooler, a humidifier and a fan to obtain an initial control strategy, with the objective function being to minimize energy consumption and temperature fluctuation;

[0075] a control strategy optimization module configured to adjust power of an actuator according to the initial control strategy using a fuzzy PID algorithm, and to monitor sensor data in real time through an isolation forest model and adjust the initial control strategy when an anomaly is detected to obtain a target control strategy.

[0076] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent control of a constant temperature incubator based on deep learning, characterized in that, The intelligent control method for a constant temperature incubator based on deep learning includes the following steps: Real-time data collection of temperature, humidity, CO2 concentration, airflow speed, and door status within the incubator; preprocessing of the collected data to obtain the device status vector. The improved Transformer model is used to analyze the state vector of the device, predict the trend of environmental parameter demand during the cultivation process, and obtain the prediction results. 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. The fuzzy PID algorithm is used to adjust the actuator power according to the initial control strategy. The isolated forest model is used to monitor sensor data in real time. When an anomaly is detected, the initial control strategy is adjusted to obtain the target control strategy.

2. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, The system collects real-time data on temperature, humidity, CO2 concentration, airflow velocity, and door status within the incubator. The collected data is preprocessed to obtain a device status vector, including: Real-time data collection of temperature, humidity, CO2 concentration, airflow speed, and door status within the incubator; 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. Extract the time series features of the processed data, and integrate the preprocessed temperature, humidity, CO2 concentration, airflow speed and door status into a multi-dimensional vector to obtain the equipment status vector.

3. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, The improved Transformer model is used to analyze the device state vector and predict the environmental parameter demand trend during the cultivation process, yielding prediction results including: A spatiotemporal attention mechanism is added 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 is used to enhance the model's temporal perception of time series data, resulting in an improved Transformer model. The improved Transformer model is used to analyze the state vector of the device and predict the trend of environmental parameter requirements during the cultivation process, thus obtaining the prediction results.

4. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, 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, including: Define the state space of the MDP Markov decision process, including the device state vector and the prediction result. Define the action space as heater power, cooler power, humidifier power and fan speed. Each action is physically constrained. With energy consumption minimization and temperature fluctuation minimization as the core objectives, a reward function is established by integrating the effects of CO2 and humidity control.

5. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, 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, including: The DDPG deep deterministic policy gradient algorithm is used to learn the optimal control policy through the Actor-Critic architecture; The Actor network is used for continuous actions that take the state as input and output the actuator power, while the Critic network is used to evaluate the value of the current action and guide the optimization of the Actor. When selecting actions, Ornstein-Uhlenbeck noise balance exploration and utilization are incorporated. 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.

6. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, The process involves using a fuzzy PID algorithm to adjust the actuator power according to the initial control strategy, monitoring sensor data in real time using an isolated forest model, and adjusting the initial control strategy when an anomaly is detected to obtain the target control strategy, including: The 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 as output.

7. The intelligent control method for a constant temperature incubator based on deep learning as described in claim 1, characterized in that, The process involves using a fuzzy PID algorithm to adjust the actuator power according to the initial control strategy, monitoring sensor data in real time using an isolated forest model, and adjusting the initial control strategy when an anomaly is detected to obtain the target control strategy, including: The isolated forest model was used to monitor whether sensor data deviated from the normal distribution and whether actuator feedback was abnormal. The model was trained with historical normal operation data and the abnormality score threshold was set to 0.

8. If the anomaly is minor, the strong correction mode of fuzzy PID is activated, and the deviation is recorded for subsequent model calibration. If the anomaly is severe, the safety policy is triggered immediately and an audible and visual alarm is issued. The alarm information is pushed to the management personnel through the IoT module.

8. A deep learning-based intelligent control system for a constant temperature incubator, characterized in that, The deep learning-based intelligent control system for the constant temperature incubator 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. The collected data is preprocessed 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 using the DRL deep reinforcement learning framework. With minimizing energy consumption and temperature fluctuations as the objective function, it dynamically optimizes the power allocation of the 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.

9. The intelligent control system for a constant temperature incubator based on deep learning as described in claim 8, characterized in that, 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.

10. The intelligent control system for a constant temperature incubator based on deep learning as described in claim 8, characterized in that, 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.

Images