Intelligent air cooling temperature control method and system

Abstract

The invention discloses an intelligent air cooling temperature control system and method, and belongs to the technical field of industrial process control. In order to solve the problems of slow response, poor precision and weak adaptability of existing PID or fixed rule control in a nonlinear and large-lag heat exchange process, a deep learning model is adopted as an intelligent control core. The system collects temperature, pressure and environmental parameters in real time through a sensor, fuses historical time sequence data, constructs feature vectors, and inputs the feature vectors to a pre-trained long-short-term memory network model. According to the model, the continuous adjusting quantity of the rotating speed of the fan and the spraying starting probability value are directly output by analyzing the time sequence dynamic state, and therefore advanced and cooperative control over the fan and the spraying device is achieved. The method overcomes the hysteresis quality and rigidity of a traditional method, has prediction adjustment and self-adaptive optimization capabilities, effectively reduces energy consumption while guaranteeing the precision and stability of the condensation temperature, and is particularly suitable for condensation link temperature control in technologies such as wine brewing and distillation.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, specifically to an intelligent air-cooled temperature control method and system. Background Technology

[0002] In distillation and condensation processes in industries such as brewing, chemical processing, and food processing, precise control of the outlet temperature after steam condensation is crucial for ensuring the quality, purity, and yield of the final product. For example, in brewing, high-temperature steam (T1>100℃) needs to be condensed into a liquid state, and the condenser outlet temperature (T3) needs to be precisely controlled and stabilized within a narrow range of 35℃ to 42℃. Traditional temperature control mainly relies on classical control theory and strategies based on fixed rules.

[0003] Currently, common industrial air-cooled temperature control solutions mainly include the following: First, using a proportional-integral-derivative (PID) controller, which adjusts the cooling fan speed in real time based on the deviation between the measured value of the temperature sensor (such as T2 or T3) and the set value. This method has a simple structure, but it often performs poorly when facing the large inertia and nonlinear characteristics of the condensation heat exchange process. Due to the lag between temperature detection and actuator response, PID control is prone to overshoot or continuous oscillation, making it difficult to achieve fast and stable adjustment.

[0004] Secondly, some systems employ on / off control rules based on fixed thresholds. For example, Chinese patent application CN118224817A discloses a control scheme that specifies that when the temperature at a certain monitoring point (e.g., T2) exceeds 50°C, the sprinkler system is triggered to execute a fixed cycle of "on for 10 seconds, off for 5 seconds." While this method is logically intuitive, it lacks adaptability. It cannot self-adjust based on complex disturbances such as ambient temperature fluctuations, transient changes in steam load (pressure P), and equipment aging, easily leading to insufficient control or energy waste.

[0005] Third, in systems with multiple actuators (such as variable frequency fans and spray valves), existing technologies typically set independent and fixed control logic for each actuator. In this mode, there is a lack of intelligent coordination between fan speed regulation and spray switch, which may even lead to control conflicts. For example, unnecessary spraying may occur even when the fan has provided sufficient cooling capacity, which not only does not help stabilize the temperature but also increases water and electricity consumption.

[0006] In summary, existing technologies (including but not limited to PID control and fixed-rule control, such as CN118224817A) generally suffer from three major drawbacks when dealing with industrial heat exchange processes characterized by nonlinearity and large time lag: slow response, insufficient adaptability, and weak coordination among actuators. This results in temperature control accuracy failing to meet the requirements of high-end processes and low energy efficiency. Therefore, there is an urgent need in this field for a novel temperature control method that can sense dynamic changes in the environment and process, predict temperature trends, and intelligently coordinate multiple actuators to achieve fast, accurate, adaptive, and energy-efficient temperature control. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent air-cooled temperature control method and system to solve the three major defects of existing industrial control technologies mentioned in the background art: slow response, insufficient adaptability, and weak coordination of actuators.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A smart air-cooled temperature control method includes the following steps: Step S1: Collect system status data in real time, including temperature values ​​T0, T1, T2, T3 and steam pressure P at multiple temperature monitoring points; Step S2: Construct a feature vector containing real-time status data, historical time-series data, control target and actuator status; Step S3: Input the feature vector into the pre-trained deep learning model for inference; Step S4: The deep learning model outputs the fan speed setpoint. and the probability of the sprinkler system being turned on ; Step S5: Based on the fan speed set value Control the variable frequency fan and adjust the spray activation probability accordingly. The comparison result with the dynamic threshold controls the opening and closing of the spray solenoid valve.

[0009] Furthermore, the deep learning model is a long short-term memory network, and its input feature vector includes at least: the current time T0, T1, T2, T3, P, the time sequence of T0, T1, T2, T3, P, fan speed S, and spray switch state K in the past 60 seconds, and the target temperature range of T3.

[0010] Furthermore, the deep learning model performs inference through the following forward propagation process: The input feature sequence is first standardized: ; The standardized sequence is input into the LSTM layer to extract temporal features; The features output from the LSTM layer are processed by the Dropout layer. ; Features after Dropout Multi-task output is achieved through a fully connected layer, where: The fan speed output is:

[0011] The spray probability output is: ,in This is the Sigmoid activation function.

[0012] Furthermore, the LSTM layer consists of at least two layers, with at least one layer containing 128-256 hidden units.

[0013] Furthermore, deep learning models are obtained through offline training, and the loss function used in training... The weighted mean square error function is expressed as follows:

[0014] in, For the sample size, The T3 value predicted by the model. For the T3 target value, The wind turbine speed predicted by the model. The optimal fan speed is indicated. This represents the probability of the sprinkler system being activated. These are the weighting coefficients.

[0015] Furthermore, dynamic threshold An adaptive mechanism is used for adjustment, and the adjustment formula is as follows:

[0016] in, The deviation between the actual value and the target value of T3. This is for adjusting the coefficient.

[0017] Furthermore, based on the sprinkler activation probability The steps for controlling the spray solenoid valve include: when When the value exceeds the dynamic threshold, the spray system is activated, and based on... Adjust the value to control the duration of the sprinkler operation. ,in This is the scaling factor; when If the temperature remains below the dynamic threshold for a preset duration, the sprinkler system will be shut off.

[0018] An intelligent air-cooled temperature control system is used to execute the aforementioned intelligent air-cooled temperature control method, the system comprising: The sensor unit is used to collect the status data in real time; The execution unit includes the variable frequency fan and the spray solenoid valve; The intelligent control unit has a built-in pre-trained deep learning model for processing the feature vectors and generating control commands.

[0019] Furthermore, the intelligent control unit includes an industrial computer deployed with a deep learning model, which is deployed in ONNX format and performs online inference once per second.

[0020] Furthermore, the system is applied to the condensation stage of brewing or distillation processes to control the temperature monitoring point T3 between 35°C and 42°C.

[0021] Compared with the prior art, the beneficial effects of the present invention are: (1) Achieve predictive feedforward control and significantly improve response speed and temperature control accuracy: By analyzing historical and real-time time series data through deep learning models, accurately predict the temperature change trend, adjust the fan speed and spray action in advance, overcome the lag problem of large inertia heat exchange process, effectively avoid overshoot and oscillation in traditional feedback control, and make the control of key temperature points faster and more stable. (2) It has strong adaptive capability and can adapt to complex working conditions and environmental disturbances: The model can automatically learn and adapt to nonlinear and time-varying disturbances such as steam pressure fluctuations and ambient temperature changes, and dynamically optimize control parameters. This solves the defect that traditional fixed rules or PID parameters cannot be adaptively adjusted according to working conditions, and enhances the robustness of the system. (3) Achieving intelligent collaboration and global energy efficiency optimization among multiple actuators: As a unified intelligent decision-making center, the model can comprehensively coordinate the working status of the fan and the sprinkler system, automatically find the best matching strategy, and reduce unnecessary energy consumption (such as ineffective sprinkler) while ensuring temperature control accuracy, thus achieving a balance between control quality and operational economy. (4) Promote the evolution of control strategies towards flexibility and intelligence: upgrade the spray control from a rigid fixed time cycle mode to a flexible decision-making mechanism based on real-time probability assessment, so that the control actions are more in line with the actual heat exchange load requirements, and improve the intelligence level and control flexibility of the system. Attached Figure Description

[0022] Figure 1 This is a flowchart of an intelligent air-cooled temperature control method according to the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent air-cooled temperature control system according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 2 This invention provides a technical solution: a deep learning-based air-cooled temperature control system, comprising: Sensor unit: used to acquire system data in real time, including temperature sensors (monitoring T0, T1, T2, T3) and pressure sensors (monitoring steam pressure P).

[0025] The execution unit includes a variable frequency fan (for adjusting speed S) and a spray solenoid valve (for controlling opening and closing K).

[0026] Intelligent control unit: Its core is a pre-trained deep learning model used to process sensor data and generate control commands.

[0027] The input feature vectors of deep learning models include: Real-time status data: T0, T1, T2, T3, P at the current moment.

[0028] Historical time series data: T0, T1, T2, T3, P, S, K sequences over a past period of time (e.g., the past 60 seconds).

[0029] Control target: The T3 target range set for the current process stage (e.g., 35-40℃).

[0030] Actuator status: current fan speed, spray switch status.

[0031] The output of a deep learning model is a control policy, including: Fan speed setting: A continuous value, ranging from 0 to 1500 r / min ¹.

[0032] Sprayer on / off probability or timing: Outputs a probability value between 0 and 1, representing the expected time before the spray system will turn on (5-10 seconds is recommended, as the heat exchange process has a large lag, allowing the actuator to act in advance; the time window size can be adjusted according to the actual system response time). The control unit dynamically decides the spray on / off timing based on this probability value, replacing the fixed "T2>50 on" rule.

[0033] After comprehensive consideration, this solution preferably uses a Long Short-Term Memory (LSTM) network and / or a temporal convolutional network as the deep learning model to effectively capture the temporal dynamic characteristics of the control process. Specifically: Input layer: Receives feature vectors with dimensions of (sequence length × number of features). For example, inputting time-series data from the past 60 seconds (including temperatures T0, T1, T2, T3, steam pressure P, fan speed S, and spray status K) and the current target temperature results in approximately 10 features and a sequence length of 60.

[0034] LSTM layer: Consists of 2-3 stacked LSTM units, with 128-256 hidden units per layer. Purpose: To capture temporal dependencies, such as the inertia of temperature changes. Input is sequential data, output is the hidden state.

[0035] Dropout layer: placed after the LSTM layer, with a dropout rate of 0.2~0.5, to prevent overfitting.

[0036] Fully connected layer: Maps the LSTM output to control inputs. The output layer has two headers: Fan speed output: continuous value, range 0-1500 rpm. ¹, Using a linear activation function. Spray probability output: values ​​between 0 and 1, using the sigmoid activation function. This architecture can effectively learn nonlinear temporal dynamics and achieve multi-step prediction.

[0037] The workflow of the intelligent control unit consists of two stages: offline training and online inference. (1) Offline training: Use historical normal operation data (including state features and corresponding optimal control results) as the training set to train the model by minimizing the loss function of control effect (such as temperature stability and energy consumption).

[0038] Data preprocessing includes the following steps: Data cleaning: Remove sensor outliers (such as invalid data with T3>100℃) and use sliding window filtering to smooth noise.

[0039] Normalization: Scaling each feature to the range of 0 to 1, the formula is as follows. ,in and This represents the minimum and maximum values ​​of historical data.

[0040] Feature engineering: Construct a time series window: concatenate the data from the current moment and the past 60 seconds into a feature vector (length 60 × number of features).

[0041] Add derived features such as the rate of temperature change (ΔT / Δt) and cumulative energy.

[0042] Vectorization: The final feature vector is a two-dimensional array (number of samples × number of time steps × number of features), which is directly input into the LSTM model.

[0043] The loss function uses weighted mean square error (MSE), combined with optimization of temperature control accuracy and energy consumption:

[0044] Symbol interpretation: Number of training samples.

[0045] : The T3 temperature value predicted by the model.

[0046] : Target temperature value for T3.

[0047] : The wind turbine speed predicted by the model.

[0048] Optimal fan speed (from labeled data).

[0049] : Sprinkler activation probability. This item penalizes excessive spraying to save energy.

[0050] Weighting coefficients (e.g.) This balances temperature error, speed deviation, and energy consumption.

[0051] (2) Online inference: The feature vectors collected in real time are input into the trained model, and the model directly outputs the fan speed and spray control suggestions, which are then converted into execution instructions by the control unit.

[0052] See Figure 1 This solution also provides a temperature control method based on the above system, including the following steps: Step 1: Data Acquisition: Real-time acquisition of T1, T2, T3, P, ambient temperature, and actuator status.

[0053] Step 2: Feature construction: Combine the current data with historical data to form a feature vector that meets the model input requirements.

[0054] Step 3: Model Inference: Input the feature vectors into the pre-trained deep learning model.

[0055] Step 4: Strategy Output: The model outputs the fan speed setpoint. and the probability of the sprinkler system being turned on .

[0056] The specific process is as follows: 1. Input feature vector construction The multi-source feature vectors received by the model contain information in four dimensions: (1) Real-time status data: five parameters at the current time: T0, T1, T2, T3, and P. Historical time-series data: The time-series sequence of system parameters over the past 60 seconds, forming a 60×7 matrix. (2) Control target: The target range of T3 for the current process stage (e.g., 35-40℃) (3) Actuator status: current fan speed S, spray status K (4) The total feature dimensions are: 60×7 (historical sequence) + 5 (current state) + 2 (control target) + 2 (actuator state) ≈ 430 feature points.

[0057] 2. Forward propagation reasoning process

[0058] Step 1: Feature Preprocessing and Embedding: The input features are first standardized to unify parameters of different dimensions to the same scale.

[0059] Step 2: Temporal Feature Extraction (LSTM Layer): The preprocessed feature sequence is input into a three-layer LSTM network. The first layer of LSTM has 128 hidden units and extracts short-term temporal patterns (such as the instantaneous trend of temperature changes). The second layer of LSTM has 256 hidden units, capturing medium- to long-term dependencies (such as the hysteretic effect of steam pressure changes on temperature). The third LSTM layer has 128 hidden units, further abstracting high-level temporal features. Step 3: Feature Dimensionality Reduction and Regularization (Dropout Layer): The high-dimensional features output by the LSTM are passed through a Dropout layer (dropout rate 0.3), randomly masking some neurons to prevent overfitting.

[0060] in, This is the hidden state vector output by the LSTM layer. Features after Dropout.

[0061] Step 4: Multi-task output (fully connected layer)

[0062] The features after dropout are input into a two-branch fully connected layer: Fan speed branch:

[0063] Using a linear activation function, the output range is mapped to [0, 1500]r·mi via a scaling factor. ¹. Among them This is the weight matrix for predicting wind turbine speed. This is the bias term for the fan speed.

[0064] Spray probability branch:

[0065] Using the Sigmoid activation function The output value range is [0,1], representing the expected probability of the sprinklers turning on within the next 5-10 seconds. This is the weight matrix for predicting the spray probability. This is the bias term for the spray probability.

[0066] 3. Explanation of the physical meaning of the reasoning process

[0067] Fan speed inference logic: The model analyzes the historical trend of T3 (such as dT3 / dt) and the deviation between the current temperature and the target temperature (T3- This includes the influence of ambient temperature on heat exchange efficiency, predicting the fan power required to maintain an ideal temperature. For example: When the model detects an upward trend in T3 and a high ambient temperature, it will increase the fan speed in advance. When the steam pressure P suddenly increases, the model predicts that the temperature will rise, and the cooling intensity will be increased in advance. Spray probabilistic inference logic: The model comprehensively evaluates the temperature level, rate of change, and current cooling capacity of the fan at T2. A high T2 value (>45℃) combined with a rapid upward trend will result in a high probability of spraying. Even if the absolute value of T2 is not high, if a step change in pressure is detected, the probability of spraying will be increased to prevent potential problems. (5) Execution control: Will The order was directly given to the variable frequency fan. Compare with a dynamic threshold, if If the threshold is exceeded, the sprinkler system will be turned on immediately, and the duration of the "on for 10 seconds" will be adjusted according to the probability value. The shutdown logic is also determined by the decreasing probability trend output by the model, achieving more flexible control than a fixed cycle.

[0068] The dynamic threshold is used for sprinkler control, initially set to 0.7 (an empirical value), and subsequently adjusted based on real-time control performance. Setting principle: Based on the temperature change pattern when the sprinkler is turned on in historical data, calculate the probability threshold to minimize the control error.

[0069] Adjustment method: A PID adaptive mechanism is adopted, and the formula is as follows:

[0070] in, The deviation between the actual value and the target value of T3. , This is to adjust the coefficient. For example, if T3 remains high, the threshold is lowered to activate the sprinklers earlier.

[0071] The strategy for adjusting the sprinkler activation duration is as follows: Base duration: fixed at 10 seconds (refer to the original rules).

[0072] Probability-weighted: On-time (seconds), where k is a scaling factor (e.g., k=5), meaning that for every 0.1 increase in probability, the duration is extended by 0.5 seconds.

[0073] Close logic: when The sprinkler system shuts off when the value is less than 0.3 and remains so for 3 seconds. The duration is determined by monitoring probability trends through a sliding window to avoid frequent on / off switching.

[0074] The following description uses two specific embodiments as examples: Example 1.

[0075] refer to Figure 2 The system includes a condenser, sensors (T0-T3, P), a variable frequency fan, a spray system, and an intelligent control cabinet. The control cabinet houses an industrial computer equipped with a deep learning model.

[0076] The training process of the model is as follows: 1. Data collection: Collect data every second for several months during the brewing condensation process, including T0-T3, P, S, K, and ambient temperature.

[0077] 2. Data cleaning and labeling: Cleaning: Use the 3σ rule to remove outliers and impute missing values.

[0078] Note: The period when T3 is stable at 35-40℃ and energy consumption is low is marked as the "optimal control" sample (accounting for about 30%).

[0079] 3. Model Training: Using the PyTorch (an open-source deep learning framework), the LSTM has 256 hidden layers and 100 training cycles.

[0080] Optimizer: Adam, learning rate 0.001, batch size 64.

[0081] Validation set: Divided by time, the last 20% of the data is used for validation to prevent overfitting.

[0082] 4. Deployment: The model is converted to ONNX (Open Neural Network Exchange Format) format and deployed on an industrial computer.

[0083] When used online, the model performs inference once per second and outputs control commands to achieve real-time dynamic adjustment.

[0084] Example 2.

[0085] Compared to Example 1, the main difference in this example lies in the specific implementation of the spray control: The model outputs the spray activation probability Pspray. A threshold is set, for example, 0.7. When... When the temperature reaches >0.7, the sprinkler system is activated. After activation, the system continuously monitors the sprinkler system. Value. When If the value remains below 0.3 for 3 seconds, the spray system will shut off. This value is determined based on system thermal inertia experiments: if the duration is too short (e.g., 1 second), it may cause frequent spray on / off cycles; if it is too long (e.g., 5 seconds), the response will be lag-induced. In practical applications, this can be optimized through adjustments, for example, shortening it to 2 seconds when steam pressure fluctuates significantly. This mechanism ensures that the spray on / off cycle is no longer a fixed 10s / 5s cycle, but rather dynamically adjusted according to actual cooling needs, making it more intelligent and efficient.

[0086] As can be seen from the above embodiments, this solution combines deep learning technology with traditional air-cooled temperature control technology, which effectively improves the control quality and intelligence level of the system and has high industrial application value.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent air-cooled temperature control method, characterized in that, Includes the following steps: Step S1: Collect system status data in real time, including temperature values ​​T0, T1, T2, T3 and steam pressure P at multiple temperature monitoring points; Step S2: Construct a feature vector containing real-time status data, historical time-series data, control target and actuator status; Step S3: Input the feature vector into the pre-trained deep learning model for inference; Step S4: The deep learning model outputs the fan speed setpoint. and the probability of the sprinkler system being turned on ; Step S5: Based on the fan speed set value Control the variable frequency fan and adjust the spray activation probability accordingly. The comparison result with the dynamic threshold controls the opening and closing of the spray solenoid valve.

2. The method of claim 1, wherein, The deep learning model is a long short-term memory network, and its input feature vector includes at least: the current time T0, T1, T2, T3, P, the time sequence of T0, T1, T2, T3, P, fan speed S, and spray switch state K in the past 60 seconds, and the target temperature range of T3.

3. The method of claim 2, wherein, The deep learning model performs inference through the following forward propagation process: The input feature sequence is first normalized: ; The standardized sequence is input into the LSTM layer to extract temporal features; The features output by the LSTM layer are processed by a Dropout layer ; Features after dropout Multi-task output through fully connected layers, where: The fan speed output is: The spray probability output is: where is a Sigmoid activation function.

4. The method of claim 3, wherein, The LSTM layer consists of at least two layers, with at least one layer containing 128-256 hidden units.

5. The method according to any one of claims 1-4, characterized in that, The deep learning model is obtained through offline training, and the loss function used in training is... The weighted mean square error function is expressed as follows: in, For the sample size, The T3 value predicted by the model. For the T3 target value, The wind turbine speed predicted by the model. The optimal fan speed is indicated. This represents the probability of the sprinkler system being activated. These are the weighting coefficients.

6. The method of claim 1, wherein, The dynamic threshold The adjustment is made by using an adaptive mechanism, and the adjustment formula is: wherein, is the deviation of the actual value of T3 from the target value, is the adjustment factor.

7. The method according to claim 1 or 6, characterized in that, According to the spray opening probability The step of controlling the spray solenoid valve includes: When greater than the dynamic threshold, the sprinkler is turned on, and the duration of the sprinkler being turned on is adjusted based on the value of the dynamic threshold wherein is a scaling factor; When When the current is continuously below the dynamic threshold for a preset duration, the spraying is turned off.

8. An intelligent air-cooled temperature control system, characterized in that, The system for performing the intelligent air-cooled temperature control method according to any one of claims 1-7, the system comprising: The sensor unit is used to collect the status data in real time; The execution unit includes the variable frequency fan and the spray solenoid valve; The intelligent control unit has a built-in pre-trained deep learning model for processing the feature vectors and generating control commands.

9. The system of claim 8, wherein, The intelligent control unit includes an industrial computer deployed with a deep learning model, which is deployed in ONNX format and performs online inference once per second.

10. The system according to claim 8 or 9, characterized in that, The system is applied to the condensation stage of brewing or distillation processes to control the temperature monitoring point T3 between 35°C and 42°C.

Images