Air conditioning system for rapidly maintaining constant temperature and humidity of clean room

Abstract

The invention provides an air conditioning system for rapidly maintaining the constant temperature and humidity of a clean room. The air conditioning system is composed of a fresh air pretreatment unit, a main constant-temperature and constant-humidity unit, a bypass rapid fine adjustment cycle and a multi-source sensing unit. Personnel, material vehicles and cabin door events are identified by performing target detection and instance segmentation on the visible light / thermal infrared time sequence image, the sensible heat disturbance direction and intensity are obtained based on optical flow estimation of thermal infrared convection plume, and a condensation risk index is generated in combination with an environment dew point; and the information is fused with multi-point temperature and humidity data to form a regional disturbance vector. A display / latent heat decoupling strategy of bypass priority and main loop slow following is adopted for control, N * M grid air supply distribution and belt constraint model prediction control are combined, and the adjustable flow guide shutters are driven to achieve differential air supply. The system realizes second-level to minute-level disturbance suppression, improves temperature and humidity precision, reduces energy consumption, considers pressure difference and cleanliness, and is suitable for high-level clean scenes such as semiconductors and biological medicines.

Description

Technical Field

[0001] This invention relates to the field of air conditioning technology, and more specifically to an air conditioning system for rapidly maintaining a clean, constant temperature and humidity. Background Technology

[0002] Cleanrooms are widely used in industries such as semiconductors, displays, biomedicine, and precision manufacturing. The processes require extremely high environmental stability, with typical indicators including temperature fluctuations of no more than ±0.5℃ and relative humidity of no more than ±3%RH. Some scenarios even require temperature fluctuations of ±0.2℃ and humidity fluctuations of ±1%RH. Existing cleanroom air conditioning systems mostly adopt a classic architecture of a centralized air handling unit (AHU) + mixing section + surface cooling / reheating + humidification, combined with terminal high-efficiency filters or FFUs to achieve constant temperature, humidity, and cleanliness control. To cope with high humidity in summer or low humidity in winter, some systems introduce rotary dehumidifiers, deep cooling reheat, or steam / electrode humidifiers; to accommodate differential pressure, they are often combined with constant air volume (CAV) or variable air volume (VAV) terminals and differential pressure control logic.

[0003] However, in actual operation, existing solutions still have several shortcomings: First, response lag. Centralized systems have long air ducts and large system heat capacity. Control loops are mostly based on supply air temperature or single-point indoor feedback. When faced with disturbances such as personnel entering and exiting, process equipment starting and stopping, door opening, material cart movement, and sudden changes in fresh air load, the room temperature and humidity recovery time is often measured in minutes or even tens of minutes, easily exceeding the stringent process window. Second, sensible heat / latent heat coupling. Many systems rely on "deep cooling + reheating" or "subcooling + rehumidification" for joint regulation. Temperature and humidity are difficult to optimize independently, resulting in repeated dehumidification / humidification or excessive reheating, which affects control accuracy and causes high energy consumption. Third, measurement and control granularity is too coarse. Traditional systems mainly rely on temperature, humidity, and differential pressure sensors at several points, which are difficult to reflect the intensity and location of local disturbances in key process areas; humidity sensors have drift and hysteresis, making it difficult to capture transient phenomena such as door opening and heat source plumes in a timely manner. Fourth, insufficient feedforward information. Existing controls are mostly feedback-based, rarely linked to process equipment status, fresh air fluctuations, personnel access, etc., lacking mechanisms for predicting and compensating for disturbances in advance. Fifth, there is insufficient coordination between cleanliness and differential pressure. In pursuit of stability, increasing total air volume and reheating are often used as "insurance," but fluctuations may still occur during high humidity seasons or under diurnal load changes, while also having side effects on differential pressure gradient and airflow organization. Sixth, engineering modification is difficult. In existing plants, increasing terminal local compensation capacity is often limited by space, ductwork, and power conditions, and requires compatibility with existing BMS / PLC systems, resulting in high implementation costs and long cycles.

[0004] Furthermore, existing technologies have limited means to acquire key physical phenomena such as localized convective plumes and the risk of condensation on cold surfaces, relying mostly on experience or post-event alarms, lacking high-frequency, regionalized disturbance identification and risk assessment. When localized condensation occurs, it may lead to quality risks such as moisture absorption of wafers, optical components, or chemicals, and particle adhesion. Traditional systems often only compensate for this after the event by adjusting overall setpoints. In summary, how to achieve rapid, precise, and loosely coupled temperature and humidity control in key areas of cleanrooms without significantly increasing system energy consumption and complexity, and how to coordinate this with pressure differential and cleanliness targets, remains a pressing technical problem to be solved in this field. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes an air conditioning system for rapidly maintaining constant temperature and humidity in cleanrooms. It comprises a fresh air pretreatment unit, a main constant temperature and humidity unit, a bypass rapid fine-tuning circulation system, and multi-source sensing. By performing target detection and instance segmentation on visible light / thermal infrared time-series images to identify personnel, material carts, and door events, the system obtains the direction and intensity of sensible heat disturbances based on optical flow estimation of thermal infrared convection plumes, and generates a condensation risk index by combining this with environmental dew point data. This information is then fused with multi-point temperature and humidity data to form a regionalized disturbance vector. The control employs a "bypass priority, main loop follow-up" sensible / latent heat decoupling strategy, combined with N×M grid air supply distribution and constrained model predictive control, driving adjustable guide louvers to achieve differentiated air delivery. This system achieves disturbance suppression at the second to minute level, improves temperature and humidity accuracy, reduces energy consumption, and balances pressure differential and cleanliness, making it suitable for high-level cleanroom environments such as semiconductor and biopharmaceutical industries.

[0006] This application provides an air conditioning system for rapidly maintaining a constant temperature and humidity in a clean environment, comprising:

[0007] The fresh air pretreatment unit is used to pre-cool, preheat, dehumidify, and humidify outdoor fresh air;

[0008] The main constant temperature and humidity treatment unit is connected to the fresh air pretreatment unit and is used to adjust the humidity and temperature of the fresh air to obtain the basic air supply for the clean room.

[0009] Clean room and main supply air ducts and return air ducts connected to the main constant temperature and humidity control unit;

[0010] The bypass fine-tuning circulation unit has one end connected to the clean room or return air duct to extract part of the air, and the other end supplies air to the clean room through an independent air supply branch. The bypass fine-tuning circulation unit includes a bypass fan, a rapid cooling or heating module, a rapid dehumidification module and a rapid humidification module.

[0011] The multi-point environmental sensing unit includes temperature and humidity sensors respectively installed in the cleanroom background area and at least one process area; and an image acquisition unit facing the process area, which performs target detection and instance segmentation on time-series images to identify personnel, material carts, open equipment doors and corresponding stay or passage events; performs optical flow estimation on the convection plume based on thermal infrared images to obtain the direction and intensity of local sensible thermal disturbances, and calculates and outputs a condensation risk index based on the thermal infrared temperature field and the environmental dew point, and fuses the condensation risk index with the temperature and humidity data to form a regionalized disturbance vector;

[0012] The hierarchical collaborative control unit is electrically connected to the fresh air pretreatment unit, the main constant temperature and humidity treatment unit, the bypass fine-tuning circulation unit, and the multi-point environmental sensing unit. The clean room is divided into an N×M regional grid. The hierarchical collaborative control unit maps the regional disturbance vector into an air supply distribution matrix and implements differentiated air supply to the corresponding grid through adjustable guide louvers, swirl air outlets, and variable air volume terminals.

[0013] Preferably, the system further includes a mixing section connected upstream of the main constant temperature and humidity treatment unit, used to mix fresh air and return air at an adjustable mixing ratio to form mixed air, wherein the mixing ratio is determined in real time by the hierarchical collaborative control unit based on load and energy consumption targets.

[0014] Preferably, the rapid cooling or heating module and the rapid dehumidification module in the bypass fine-tuning circulation unit are arranged independently and controlled by independent actuators to achieve functional decoupling of temperature and humidity regulation.

[0015] Preferably, the outlet of the independent air supply branch is located above the process equipment and is equipped with an adjustable guide component to change the air supply direction, so that the local compensation matches the dominant airflow direction of the disturbance source in the same or opposite direction.

[0016] Preferably, a visual processing module is provided between the image acquisition unit and the hierarchical collaborative control unit. The module uses a multi-scale convolutional neural network combined with temporal attention to perform target detection and instance segmentation, and performs optical flow estimation based on adjacent frames to reconstruct the thermal plume velocity field, and further calculates the condensation risk index.

[0017] Preferably, the outputs of the multi-point environmental perception unit and the visual processing module are subjected to Kalman filtering to obtain a regionalized perturbation vector, which includes the discrimination of outliers.

[0018] Preferably, the hierarchical collaborative control unit sets a first threshold and a second threshold, as well as a corresponding hysteresis band: when the deviation or its rate of change in the regionalized disturbance vector exceeds the first threshold, bypass priority compensation is triggered; when the deviation continues to exceed the second threshold, the main constant temperature and humidity processing unit is linked; the hierarchical collaborative control unit uses constrained model predictive control (MPC) to generate the air supply distribution matrix, with a prediction time domain of 30 to 180 seconds, and simultaneously applies dew point constraints, pressure difference constraints, and energy consumption penalties; when optimization is not feasible or the solution fails, it degenerates into hierarchical PID control.

[0019] Preferably, the image acquisition unit performs the following target detection and instance segmentation process on the time-series image:

[0020] S1) Acquire visible light or thermal infrared images at a rate of no less than 15 frames per second, and perform distortion correction, white balance and brightness normalization preprocessing.

[0021] S2) Input the preprocessed image into the backbone of a convolutional neural network containing a feature pyramid to extract multi-scale features;

[0022] S3) Generates center heatmaps, width, height and offset of each target category through anchor point free detection head on multi-scale features, and outputs candidate box confidence scores for detecting personnel, material carts and equipment doors;

[0023] S4) For each candidate box, a binary mask for the candidate instance is generated using a mask branch based on ROI alignment to obtain the instance segmentation result;

[0024] S5) Perform non-maximum suppression (NMS) on the candidate boxes, with an IoU threshold in the range of 0.3 to 0.6;

[0025] S6) Multi-target tracking using Kalman filter prediction and Hungarian matching is used to associate cross-frame instances, and the dwell time Δt of each instance in the preset area and the virtual gate crossing event are calculated. When Δt≥T1, it is determined as a dwell event, and when the gate crossing occurs, it is determined as a pass event. T1 is the preset dwell threshold.

[0026] S7) Perform segmentation on the ROI of the door and calculate the proportion of the opening mask θ. When θ≥θ_open, output the open state. θ_open is the preset opening threshold.

[0027] Preferably, the optical flow estimation process for convective plumes is as follows:

[0028] S1) Acquire thermal infrared images at a rate of no less than 15 frames per second and perform radiometric calibration and non-uniformity correction, converting pixel radiance into a temperature field T(x,y,t);

[0029] S2) Establish a background temperature model T_bg(x,y,t) based on exponential weighted moving average, calculate the temperature anomaly ΔT=T−T_bg, generate a plume candidate mask M based on the condition ΔT≥γ2, where γ2 is a configurable threshold, and perform morphological opening and closing operations to obtain the plume region ROI.

[0030] S3) At the pyramid multi-scale, the velocity field v(x,y)=(v_x,v_y) is solved by TV-L1 dense optical flow for ROI. The data terms are constructed with the assumption of temperature invariance, the regularization terms are smoothed by total variation TV, and the data terms are weighted by temperature gradient to enhance the stability of low texture region.

[0031] S4) Perform forward and backward consistency checks and median filtering on the optical flow results to remove outlier vectors, and apply physical constraints, including prioritizing the rising component coaxial with the direction of gravity, to obtain the plume velocity field.

[0032] S5) Based on the camera's geometric calibration, map v(x,y) to the cleanroom coordinate system and calculate the principal direction angle θ = atan2(v_y,v_x); and characterize the local sensible thermal disturbance intensity with the disturbance intensity index H = mean_ROI(∥v∥·ΔT); when H ≥ H_min, it is judged as a significant sensible thermal disturbance; H_min is the intensity threshold; atan2() is the two-parameter arctangent function; mean_ROI() is the spatial average of the field within the ROI region; v_x and v_y are the horizontal and vertical velocities of the velocity v, respectively;

[0033] S6) Outputs {θ, H} and together with the temperature and humidity sensing data, forms a regionalized disturbance vector, which is used by the hierarchical collaborative control unit for bypass priority compensation and air supply distribution matrix update.

[0034] Preferably, the process for calculating and outputting the condensation risk index based on the thermal infrared temperature field and the ambient dew point includes:

[0035] S1) Acquire thermal infrared time-series images and complete radiometric calibration, non-uniformity correction and bad pixel repair to obtain a time series of pixel-level surface temperature distribution;

[0036] S2) Obtain the ambient dry-bulb temperature and relative humidity from the cleanroom background temperature and humidity sensor, and determine the ambient dew point temperature according to the standard humid air properties algorithm;

[0037] S3) Compare the surface temperature and dew point temperature by pixel, define the dew point margin as the temperature difference between the two, and mark candidate dew condensation areas according to the set threshold. Perform connected component filtering and morphological opening and closing operations on the candidate areas to remove noise and isolated small blocks.

[0038] S4) Evaluate the three risk components separately within the candidate region:

[0039] Temperature difference risk component: The greater the degree to which the surface temperature is below or close to the dew point, the higher the risk level;

[0040] Cooling rate risk component: The faster the cooling occurs and the longer the duration of cooling within the preset time window, the higher the risk level.

[0041] Convection enhancement component: Combining the airflow velocity and direction obtained from optical flow estimation, the more favorable the airflow is to remove sensible heat or promote local cooling, the higher the risk level.

[0042] S5) Normalize and fuse the above three types of components according to preset weights to obtain a pixel-level condensation risk index; and perform spatial statistics on the pixel-level index in the candidate region to obtain a regional-level condensation risk index.

[0043] S6) Set two thresholds, high and low, with hysteresis: when the regional condensation risk index reaches the high threshold, trigger the dehumidification priority or reheat priority strategy; when it is between the low and high thresholds, perform bypass micro-compensation; when it is below the low threshold, only maintain monitoring;

[0044] S7) Outputs a condensation risk heat map and a regional condensation risk index, and writes it into a regional disturbance vector for the hierarchical collaborative control unit to update the air supply allocation and execution strategy.

[0045] This invention provides an air conditioning system for rapidly maintaining a constant temperature and humidity in a clean environment, and the beneficial technical effects it achieves are as follows:

[0046] 1. This application introduces a bypass fine-tuning loop on the basis of the traditional main loop, and uses the generated regionalized disturbance vector as the feedforward signal. By identifying personnel / materials / door events, estimating thermal infrared plume optical flow, and assessing condensation risk index, it achieves sensible / latent heat decoupling control with "bypass priority and main loop follow-up". Combined with N×M grid differentiated air supply, it can suppress local fluctuations on a second to minute basis. Under typical operating conditions, temperature fluctuations converge to ±0.2℃ and humidity to around ±1%RH, significantly reducing overshoot and recovery time.

[0047] 2. This application implements sensible / latent heat independently to avoid repeated offsetting of "deep cooling + reheating, excessive humidification + re-dehumidification"; model predictive control incorporates dew point, pressure difference, and energy consumption into the constraints, dynamically allocating airflow and dew point of the main / bypass and each grid according to the load, prioritizing small-scale compensation near the disturbance source, reducing large-scale system-wide adjustments and long-term maintenance of high airflow. This reduces cooling, reheating, and humidification water consumption, decreases equipment start-up and shutdown frequency and valve operation, and improves long-term operational stability and economy.

[0048] 3. This application's system adopts a modular architecture. The bypass unit, image acquisition, and vision processing module can be deployed outside existing cleanrooms, compatible with existing BMS / PLCs. Camera geometric self-calibration and robust occlusion / drift fusion ensure data reliability. Optimization failures automatically degenerate into hierarchical PID while maintaining differential pressure / cleanliness constraints. The algorithm threshold and prediction time domain can be parameterized, requiring minimal maintenance and facilitating phased implementation and rapid replication and promotion. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the steps of an air conditioning method for rapidly maintaining a clean and constant room temperature and humidity, according to the present invention.

[0051] Figure 2 This is a schematic diagram of an air conditioning system for quickly maintaining a clean, constant temperature and humidity, according to the present invention. Detailed Implementation

[0052] 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.

[0053] Example 1:

[0054] In view of the aforementioned problems mentioned in the prior art, and in order to solve the above technical problems, as shown in the appendix. Figure 1 As shown: This application provides an air conditioning system for rapidly maintaining a constant temperature and humidity in a clean environment, comprising:

[0055] The fresh air pretreatment unit is used to pre-cool, preheat, dehumidify, and humidify outdoor fresh air using one or more of the following methods: In one embodiment, the designed fresh air volume is 20,000 m³ / h, with outdoor temperatures of 35℃ dB / 28℃ WB in summer and –5℃ dB / 70%RH in winter. The pre-treatment targets (before entering the mixing section / main unit) are: summer outlet air 22–24℃, dew point 12–14℃; winter outlet air ≥8–10℃ to avoid frost formation; during transitional seasons, dew point is prioritized. Unit Composition (in order of airflow direction): Air inlet louvers + rain cap + insect screen; electric fresh air regulating valve (with 0–10V or Modbus); pre-filter section (G4 / ePM1050%, optional sand and dust pre-filter); surface cooling coil (chilled water 7 / 12℃, or straight expansion copper tube aluminum fin); demisting baffle + stainless steel water collection tray (slope ≥1%, double drain outlet); preheating coil (hot water 60 / 50℃ or electric heating / steam) with freeze protection and temperature recovery; humidifier (electrode type steam or micro-mist ultrasonic, using RO / DI water, with sterilization and sewage discharge); medium-efficiency filter section (F7 / ePM155%) as a pre-purification barrier.

[0056] The sensing and control are as follows, with sensing points including: outdoor T / RH, pre-treated outlet air T / RH, air valve opening, coil inlet / outlet water temperature / pressure difference, condensate pan level, and frosting temperature control. Summer: The main loop is set at dew point (12–14℃), with PID regulation of the chilled water valve; if the outlet dry bulb temperature is below 22℃, a small amount of preheating is used to restore temperature, reducing the main unit's reheat function. Winter: Freezing prevention is prioritized at the lowest outlet air temperature (≥8–10℃), with the preheating valve taking precedence; when RH is too low, humidification (staged / pulse flow) is activated. The system includes coil freeze protection interlock, air valve and fan interlock, level alarm, and filter differential pressure alarm. This "smooths out" the highly fluctuating fresh air condition, absorbing some latent heat load at the front end, and reducing the main unit's reliance on deep cooling and reheat.

[0057] The main constant temperature and humidity processing unit, connected to the fresh air pretreatment unit, is used to regulate the humidity and temperature of the fresh air to obtain the basic supply air for the cleanroom. The operating conditions and objectives are as follows: the designed air supply volume is 40,000 m³ / h. The cleanroom target is an indoor temperature of 20.0±0.2℃ and a RH of 45±1%; differential pressure and cleanliness are controlled collaboratively by the terminal units and the system. Unit outlet air (basic supply air condition): dry bulb temperature 19–20℃, dew point 10–11℃; ensuring that the set values ​​are maintained under the average process load without significant reheating / rehumidification after entering the room.

[0058] The structure and flow (in order of airflow direction) are as follows: The medium-high efficiency filtration section includes F7 (ePM 155%) or F8, preventing residual dust after pretreatment from entering the heat exchanger. First-stage surface cooling coil (primarily sensible heat): chilled water at 7 / 12℃ or direct expansion, aiming to reduce the mixed air temperature from ~26–28℃ to ~18–19℃, not for strong dehumidification, reducing the load on subsequent deep cooling stages. Deep dehumidification coil (primarily latent heat): independent loop, chilled water at 5 / 10℃ (or secondary refrigeration / low-temperature brine), raising the dew point to the 10–11℃ range; this coil is equipped with a high-efficiency baffle plate and a SUS304 water collection tray. Reheat coil: Prioritizes waste heat reheat (e.g., condensation heat recovery with low-temperature hot water at 35–45℃ or heat pipe regeneration); when waste heat is insufficient, it is compensated by low-power electricity / steam / hot water to bring the outlet dry bulb temperature back to 19–20℃. Basic humidification section (optional): Electrode-type steam or micro-mist humidification, serving as "chassis humidification" in winter or extremely dry conditions. It should be kept inactive during normal operation to avoid competing with dehumidification. Terminal filtration section: F9 (ePM180%) acts as the unit's terminal barrier; HEPA filters are generally installed in the terminal box / ceiling, configured according to cleanliness level. Noise-absorbing / flexible joint and maintenance section: Includes a maintenance door, differential pressure tap, and observation window.

[0059] Sensors for measuring points and actuators: Mixing section T / RH, T / RH before and after the primary coil, T / RH before and after the depth coil, T / RH after reheat (for calculating dew point), coil inlet and outlet water temperature / pressure difference, filter pressure difference, and water level in the sump tray. Actuators: Two-stage chilled water control valve (independent PID), reheat valve (prioritizing waste heat return to water, electric / steam as secondary), basic humidification valve / variable frequency pump, and fan inverter (coordinated with the system's total airflow / pressure difference strategy).

[0060] Control strategy (dew point priority + dry bulb follow, sensible / latent heat decoupling): Dew point main loop: closed loop to the target dew point (10–11℃) with the "deep dehumidification coil valve position" as the main actuator; Dry bulb slave loop: closed loop to the outlet dry bulb (19–20℃) with the "reheat coil valve position" as the actuator; First-stage coil feedforward coordination: calculated based on the enthalpy and flow rate of the mixed air, prioritizing sensible heat reduction and reducing the burden on the deep coil; Basic humidification: activated only in winter or under extremely dry fresh air conditions at the lower limit of absolute moisture content, providing small-scale humidification to avoid conflict with deep cooling dehumidification; Linkage with bypass fine-tuning: when the hierarchical control detects a critical area deviation / rate of change exceeding the threshold, the bypass loop takes priority; if the deviation continues to exceed the second threshold, the main unit dew point / dry bulb setting is fine-tuned (e.g., dew point –0.5℃, dry bulb +0.3℃), forming a hierarchical control of "bypass first, main loop follow". Energy efficiency: Waste heat reheat is prioritized; when the outdoor absolute humidity is low and the pre-treatment is close to the target dew point, the deep coil valve position is automatically lowered and the sensible heat sharing of the first-stage coil is increased to reduce deep cooling and reheating. Anti-condensation / water carryover: Baffle plate efficiency ≥98%, internal air velocity ≤2.2m / s; water collection pan ≥1% slope + deep water seal, with level and overflow alarms. Anti-freeze: Low temperature / winter minimum outlet air temperature protection (≥8–10℃), coil return water temperature linked to outdoor temperature; over-limit switching prioritizes reheat. Maintenance: Filter differential pressure alarm threshold tiers; all valves have position feedback for easy BMS self-diagnosis. Under high humidity and disturbance conditions in summer, the main unit stabilizes the "base load" to a dew point of 10–11℃ and a dry bulb temperature of 19–20℃, with the bypass only intervening briefly during events. Compared to the traditional "single coil deep cooling + large reheat" solution, it can reduce the reheat by 10–25% (depending on the operating conditions), while improving temperature and humidity stability and reducing overshoot.

[0061] Clean room and main supply air ducts and return air ducts connected to the main constant temperature and humidity control unit;

[0062] The bypass fine-tuning circulation unit has one end connected to the clean room or return air duct to extract part of the air, and the other end supplies air to the clean room through an independent air supply branch. The bypass fine-tuning circulation unit includes a bypass fan, a rapid cooling or heating module, a rapid dehumidification module and a rapid humidification module.

[0063] In one embodiment, the application scenario and cleanliness level are: ISO 6 for the main area and ISO 5 for key workstations (above equipment). Area and floor height: approximately 250 m², floor height 3.4 m, ceiling static pressure layer 0.8 m. Design air volume: main unit supply air 40,000 m³ / h (including 2,500 m³ / h fresh air), return air 37,500 m³ / h. Pressure difference: cleanroom to corridor maintained at +15 to +25 Pa. Main supply and return air duct layout: main supply air duct: unit outlet through silencer → galvanized sheet supply air duct (with sound-absorbing lining), arranged along the corridor, branching into four branches within the static pressure layer to the ceiling HEPA terminal (H14), supply air velocity 0.35 to 0.45 m / s. Return air system: Room return air louvers → ceiling return air interlayer → return air shaft → return air main duct back to the main unit mixing section; a DN250 bypass air intake tee (with electric shut-off valve and check valve) is reserved on the return air main duct. Inspection and measurement points: The supply and return air main ducts are equipped with T / RH, static pressure taps, and inspection doors; differential pressure ports are provided before and after the HEPA filter for easy maintenance and verification.

[0064] The bypass rapid-adjustment circulation unit (independent "small circulation") is located in the ceiling or equipment layer, serving key disturbance sources (such as lithography / baking / opening positions) nearby. 1) Air path connection, intake end: two options can be selected or used in combination: a. Air intake from the return air main (recommended default), tee + electric air valve + check valve; b. Air intake from the local return air vent in the cleanroom ceiling (for very close-end circulation, to speed up response). Supply end: Independent DN200~DN250 branch passes through the static pressure layer, and the end is connected to adjustable guide louvers / swirl air outlets + end HEPA (H14), directly targeting the area 0.8~1.2m above the key work positions for directional air supply; each bypass end is equipped with an electric regulating air valve (VAV) and a silencer section. Air volume configuration: 800~1,200m³ / h per bypass, this example configures 4 sets, the total bypass accounts for about 10~12% of the main supply air volume.

[0065] Functional Sections (by airflow direction): B1 Bypass Fan Section: EC or backward-curved centrifugal fan + frequency converter, static pressure margin ≥600Pa, with flexible connection and vibration damping. B2 Rapid Cooling / Heating Module (Sensible Heat): Small coil (shared or two-pipe system for chilled water 7 / 12℃ and hot water 45 / 35℃), electric two-way valve, response ≤2s; used for dry-bulb fine-tuning. B3 Rapid Dehumidification Module (Latent Heat): Surface cooling / deep cooling small coil (5 / 10℃ loop or direct expansion core), equipped with baffle plate and SUS304 water collection tray; primarily used for dew point correction. B4 Rapid Humidification Module: Electrode steam or micro-mist integrated sprayer (RO / DI), minimum pulse ≤3s, anti-drip / backflow. B5 High-Efficiency Terminal Filter: H14 grade, frame with condensation break bridge and drip edge, pre-filter F9 protection segment optional. B6 Sound Insulation / Thermal Insulation: 50mm fiberglass / rock wool enclosure, full-length piping insulation, cold bridge disconnection. 3) Electric air valves (with opening feedback) and check valves are installed at valves, air intakes, branches, and terminals; all coil valves are 0-10V adjustable with position feedback; the fan / humidifier has remote start / stop and fault feedback.

[0066] Measuring points and "fast loop" control logic: Measuring point configuration includes T / RH + wind speed above the critical workstation, T / RH of the background area; T / RH of the bypass outlet air, coil inlet and outlet water temperature, valve position, and air volume (calibrated with a radiator / airflow hood); image / thermal image acquisition (for critical workstations), providing the vision module with output events / plume direction / intensity / condensation risk. Action logic (hierarchical coordination with the main system): Triggering conditions: When any of the following occurs at the critical workstation: temperature deviation ≥ 0.3℃ or rate of change exceeds the threshold; humidity deviation ≥ 1.5%RH or rate of change exceeds the threshold; vision detects personnel / material carts passing by, doors opening, or thermal plume intensity reaching "medium level" or above; then the hierarchical control issues a "bypass priority" command.

[0067] The priority order is as follows: When humidity is too high, first use rapid dehumidification (B3) to correct the dew point; if the dry bulb is too low, use B2 for reheating to slightly restore the temperature. When humidity is too low, first use rapid humidification (B4), move the dry bulb outward, and then use B2. For dry bulb deviation only: only adjust B2. Airflow and direction: Increase the VAV opening of the corresponding grid / station to 60-100% of the rated value; the angle of the guide louvers should point upstream of the disturbance source. Exit strategy: After the disturbance is recovered and maintained for 60-120 seconds, the valve position and airflow should be ramped back to the baseline to prevent oscillation. Linkage with the main circuit: If the deviation in the same area lasts for >3-5 minutes, the main unit dew point setting should be lowered by 0.3-0.5℃ or the dry bulb setting should be raised by 0.2-0.3℃ to follow up gradually, and it will automatically return to normal after the disturbance is resolved.

[0068] The differential pressure and cleanliness synergistic graded control applies a differential pressure constraint (+15~+25Pa) to the total air volume and bypass air volume at all times; when the bypass air volume increases and causes static pressure fluctuations, the main supply air frequency converter makes a slight compensation; the bypass terminal is always equipped with HEPA filters to ensure that local air increase does not damage the gradation and flow pattern.

[0069] Performance and Commissioning Highlights: Response Time: From the detection of a trigger event by the bypass to the outlet air T / RH reaching the 90% target, typically 30–90 seconds; Bypass Ratio: Baseline 6–8%, temporarily rising to 10–12% during the event period, and then dropping back down in 60–120 seconds after the event ends; Anti-condensation: B3 / B4 actions are constrained by the "condensation risk index," prioritizing the dehumidification + reheat path; Noise Control: Silencing plates are installed in the branch circuits, with NC ≤ 55 at the terminal; Maintenance: The bypass water collection pan drainage is equipped with a deep water seal and is flushed regularly; HEPA pre-installed F9 differential pressure alarm.

[0070] Traditionally relying solely on the main unit: humidity fluctuations at critical points are within ±3%RH, and recovery takes 10–15 minutes; after introducing a bypass: humidity fluctuations are reduced to within ±1%RH, and recovery takes 2–4 minutes; dry bulb fluctuations are ≤±0.3℃; the number of times the main unit reheat valve position and large-scale adjustments are reduced, resulting in a decrease in annual energy consumption (generally up to 8–15%).

[0071] The multi-point environmental sensing unit includes temperature and humidity sensors respectively installed in the cleanroom background area and at least one process area; and an image acquisition unit facing the process area, which performs target detection and instance segmentation on time-series images to identify personnel, material carts, open equipment doors and corresponding stay or passage events; performs optical flow estimation on the convection plume based on thermal infrared images to obtain the direction and intensity of local sensible thermal disturbances, and calculates and outputs a condensation risk index based on the thermal infrared temperature field and the environmental dew point, and fuses the condensation risk index with the temperature and humidity data to form a regionalized disturbance vector;

[0072] In one embodiment, the application scenario and objectives are as follows: Scenario: ISO 6 cleanroom (critical workstations ISO 5), area approximately 250m², 6 critical workstations (above equipment). Objective: To form regionalized disturbance vectors in an N×M (10×10) grid with an update cycle of 1–2 seconds, providing "bypass priority" feedforward for hierarchical collaborative control. Hardware composition and installation: 1) Temperature and humidity sensors (background + process points), accuracy: temperature ±0.1~0.2℃, relative humidity ±1~2%RH; long-term drift ≤±0.1℃ / year, ±0.5%RH / year. Sampling cycle: 1Hz; communication: RS485 / ModbusRTU or BACnetMS / TP (local aggregation followed by uploading to BACnet / IP). Layout: 6 points in the background area (uniform grid), ≥0.5m from the wall, ≥1m from the supply and return air vents, height 1.2–1.5m; 6 points above the critical workstations, 0.6–0.8m from the top of the equipment, avoiding direct airflow.

[0073] 2) Image Acquisition Unit (Visible Light + Thermal Infrared): Visible Light Camera: Resolution ≥ 1080p, Frame Rate 15–30fps, PoE Power Supply, Low-Distortion Lens; Mounted on ceiling beams / equipment racks, overhead view of key workstations, Field of View ≥ 2.5 × 2.5m; Dust Cover Included. Thermal Infrared Camera: Resolution ≥ 320 × 240 (preferably 640 × 480), Thermal Sensitivity ≤ 50mK, Frame Rate ≥ 15fps; Installed in the same or close proximity position, FOV overlap with the visible light camera by more than 90%. Edge Computing Unit: Industrial Embedded Host (GPU / AI Acceleration), Dual Gigabit Ethernet Ports + RS485, Running Vision and Fusion Algorithms; UPS Power Supply (≥ 15 minutes).

[0074] 3) Network, clock, and power supply: Gigabit wired network from the core switch to the edge hosts; PoE for cameras; NTP time synchronization for all devices (drift ≤50ms). Flame-retardant network cables and metal cable trays are used for cabling, with stainless steel sealed boxes and cable seals installed where the cables pass through clean areas.

[0075] Calibration and self-calibration of camera intrinsic / extrinsic parameters: After installation, a checkerboard pattern is used to complete the visible light intrinsic parameter calibration; thermal infrared calibration is performed using an isothermal blackbody plate; the two cameras are spatially registered using a common calibration plate (establishing a mapping from pixels to room coordinates). Coordinate mapping: Three-point attitude markers are affixed to the ceiling or equipment housing, and the system automatically solves for extrinsic parameters; recalibration is automatically triggered after relocation / maintenance. Temperature and humidity sensor calibration: Compare with a standard source 30 minutes after power-on; drift self-check every 6 months, triggering an alert if limits are exceeded.

[0076] Data processing and algorithm flow (updated every 1-2 seconds)

[0077] 1) Preprocessing: Visible light: distortion correction, brightness normalization, motion blur suppression; Thermal infrared: radiometric calibration, non-uniformity correction, bad pixel filling, mild temporal filtering; Dew point calculation: calculate the environmental dew point from the background temperature and humidity, which is used as a parameter for subsequent risk assessment.

[0078] 2) Target detection and instance segmentation (event recognition): Lightweight multi-scale detection + mask branch (such as anchor free center detection + ROI mask) is run on the edge host to identify personnel, material vehicles, and equipment doors; multi-target tracking (Kalman prediction + Hungarian matching) is used to count dwell time and virtual door line crossing; the proportion of the opening mask in the door area is used to calculate the door opening event (opening / closing).

[0079] 3) Optical flow estimation of convective plumes (sensible thermal disturbance): Candidate plume regions (ROIs) are generated on the thermal infrared image according to temperature anomalies. The velocity field is solved by multi-scale dense optical flow, and forward and backward consistency checks and median filtering are performed. Combined with camera registration, the optical flow vector is mapped to the room coordinates to obtain the main direction of the plume and the disturbance intensity level (divided into low / medium / high, based on a combination of velocity and temperature anomalies).

[0080] 4) Condensation risk index: compare surface temperature with ambient dew point to obtain dew point margin; comprehensively evaluate the three components of temperature difference approach, short-term cooling rate and convection enhancement in candidate cold surface areas; output pixel-level risk heat map and regional risk level (low / medium / high, with hysteresis).

[0081] 5) Multi-source data fusion → regionalized perturbation vector: aggregate temperature and humidity location data, event statistics, plume direction / intensity, and condensation risk level into an N×M grid; use Kalman / Bayesian fusion to suppress noise and outliers, and generate perturbation vector fields for each grid, including: temperature and humidity deviation and short-term change rate; event count and gate status (past 10–30s window); plume main direction and intensity level; condensation risk level and confidence level; data freshness and quality markers (camera occlusion, sensor anomaly flags).

[0082] For the hierarchical collaborative control unit: a full-grid perturbation vector matrix is ​​pushed every 1–2 seconds; when any grid meets the trigger logic (deviation / rate of change / event / plume intensity / condensation risk reaches the threshold), a bypass priority compensation instruction is generated (including target grid, desired supply air dew point / dry bulb correction, VAV airflow target, and guide angle suggestion). Threshold temperature deviation ≥ 0.3℃ or change ≥ 0.2℃ within 10 seconds; humidity deviation ≥ 1.5%RH or change ≥ 1.0%RH within 10 seconds; plume intensity: medium or above; condensation risk: medium or above; event: door opening duration ≥ 3 seconds, material cart passage frequency ≥ 2 times / 30 seconds. If there is no trigger for 60–120 seconds after exiting, the airflow is attenuated back to the baseline airflow / set to prevent oscillation. In the visible light flow, face and recognizable details are anonymized / masked at the edge; only event metadata and heatmaps are uploaded, and the original human image is not saved. When the camera is obstructed / the lights are off, it automatically degenerates into a sensor-only feedback mechanism; temperature and humidity sensor drift is corrected by background self-checking and comparison; critical data is uploaded in full-duplex mode and buffered locally for ≥24 hours. Network security: cameras and edge hosts communicate in independent VLANs with whitelists; device firmware and script signature verification.

[0083] The debugging process achieves the following results: ≥95% accuracy in identifying personnel / material carts / door opening events; ≤15° error in determining the direction of thermal plumes; ≥90% consistency rate in determining intensity levels; ≥30s advance warning for high-risk condensation scenarios; and ≤90s for a significant T / RH response (90% target) to the bypass outlet airflow after a disturbance. Long-term stability: 30 days of continuous operation without frame drops or system crashes; no abnormal sensor alarms; and data clock drift ≤100ms.

[0084] The hierarchical collaborative control unit is electrically connected to the fresh air pretreatment unit, the main constant temperature and humidity treatment unit, the bypass fine-tuning circulation unit, and the multi-point environmental sensing unit. The clean room is divided into an N×M regional grid. The hierarchical collaborative control unit maps the regional disturbance vector into an air supply distribution matrix and implements differentiated air supply to the corresponding grid through adjustable guide louvers, swirl air outlets, and variable air volume terminals.

[0085] In one embodiment, the system architecture and communication control hardware are as follows: Upper-level industrial PC (IPC): Performs vision-data fusion, grid disturbance assessment, and air supply distribution optimization (including MPC). PLC / DDC: High-speed execution layer, handling closed-loop VAV airflow, guide louver angle, bypass coil valve position; acquiring T / RH / differential pressure, etc. Network and protocol: Camera / edge computing power → IPC (Gigabit Ethernet); IPC ↔ PLC / DDC (BACnet / IP or Modbus / TCP); End effectors (air valves, VAVs, guide motors, valves) use 0–10V / 4–20mA or bus modules. Cycle time: Fast loop 1s; Medium-speed coordination 10s; Host follow-up 60–120s; Energy optimization / reporting 5–10min.

[0086] Area Grid and Influence Coefficients, Spatial Division: Project the cleanroom into an N×M grid (e.g., 10×10), with each grid approximately 0.6–1.0m in side length; number each grid and establish a coverage relationship table with the terminal units (VAV, guide louvers / swirl vents, bypass terminals). Influence Coefficient Library: During the commissioning phase, perform a small-amplitude "pulse airflow increase / guide angle change" on each terminal unit, and measure the influence coefficient matrix (response weights and time constants from terminal unit to grid) through the grid T / RH / wind speed response. The coefficient library is used to map "grid requirements" to "terminal unit actions".

[0087] Input data and "regionalized perturbation vectors" are sourced from: multi-point temperature and humidity, pressure difference; visible light / thermal infrared events (personnel / materials / door opening); plume main direction and intensity; condensation risk level; and data quality markers (occlusion / drift). The perturbation vector content for each grid includes: temperature / humidity deviation and short-term rate of change; event intensity (cumulative passage / door opening over the past 10–30 seconds); plume direction (relative to terminal orientation) and intensity classification; and condensation risk level (low / medium / high) and confidence level. Update cycle: refreshed every 1–2 seconds; outlier data is denoised using Kalman / Bayesian fusion.

[0088] Air supply allocation matrix generation (core logic), urgency scoring: Calculate the "urgency" for each grid, taking into account temperature and humidity deviation, rate of change, events, plume intensity, and condensation risk, to form U(i,j) (0–1). Target action set (for each grid): Airflow target: VAV target airflow offset of adjacent terminals (relative to baseline ±0–40%); Guiding target: Angle / rotation target of guiding louvers / swirl vents to make the supply air flow in reverse / tangential direction to counteract the plume; Bypass setting: "Target dew point fine-tuning, dry bulb fine-tuning, humidification / dehumidification priority" of the bypass terminals in the corresponding area. Terminal layer synthesis: Based on the influence coefficient library, synthesize the targets of each grid into the terminal layer to obtain the air supply allocation matrix A (terminal × target airflow / angle); if multiple grids make different requirements for the same terminal, calculate the weighted average according to urgency and distance weights, and apply rate of change and step limits to prevent oscillation.

[0089] Constraints and solution pressure differential constraints: The total supply air-return air difference is maintained at +15 to +25 Pa, and total and zone limits are applied to A; Dew point / condensation constraints: When the risk of condensation in a certain grid is medium / high, the "dehumidification + reheat" path is prioritized, and direct humidification and subcooled supply air are restricted; Noise / comfort constraints: Maximum airflow step under terminal wind speed and NC limit; Minimum ventilation: The baseline airflow in each zone is not lower than the process and personnel density requirements. Optimizer: MPC with constraints (predicting 60–120s), with the objective of "urgency-weighted error + energy consumption penalty + action smoothing", to calculate the final A*.

[0090] Layered timing control (three-level linkage): L1 fast loop (1s, local priority) executes small steps of A*: VAV opening fine-tuning, guide angle fine-tuning, bypass B2 / B3 / B4 fine-tuning (sensible / latent heat decoupling: dry bulb → B2, dew point → B3, humidification → B4). The bypass airflow limit is usually set to 5-30% of the total supply airflow with local flexibility within the frame; a single step does not exceed 10%. L2 coordination loop (10-30s, regional stability): If the same region continues to exceed the threshold, the task is appropriately transferred: part of the compensation is changed from "airflow + guide" to "dew point / dry bulb setting fine-tuning", and a small offset is made in the adjacent grid to maintain the pressure difference and flow pattern. If the load change persists for 3–5 minutes without subsiding (60–120s), the L3 main unit will fine-tune the main constant temperature and humidity processing unit settings: dew point -0.3 to -0.5℃, dry bulb +0.2 to +0.3℃, etc.; after the disturbance is resolved, it will return to its ramp state.

[0091] Actuator VAV: with airflow sensing and closed-loop control; louvers / vortex vents: stepper motor or servo motor with angle feedback; bypass valve: electric two-way valve, 0–10V; bypass fan: frequency converter. Anti-jitter and hysteresis: set upper and lower thresholds and minimum hold time (60–120s) for trigger conditions to avoid frequent switching. Fault and degradation: In case of visual / network anomalies, the system degrades to "sensor-only feedback + hierarchical PID"; when exceeding the feasible range, the optimizer stops, maintaining the previous stable ratio. Baseline calibration: establish baseline airflow / guide angle for each zone to meet differential pressure and cleanliness requirements; Pulse test: measure the end-to-grid influence coefficient and fill the coefficient library; Thresholds and weights: adjust urgency weights and action limits based on process windows (e.g., ±0.2℃, ±1%RH) and event characteristics; Differential pressure linkage: set upper / lower limits for total airflow and return air compensation logic.

[0092] The critical workstation door opens for 6 seconds, allowing the material cart to pass through; thermal imaging identifies the rising plume as having a "medium-high" intensity. In the rapid loop (1 second): the target grid urgency increases, A* outputs VAV +25% for this area, and the guide angle shifts upstream of the plume; bypass B3 (rapid dehumidification) slightly cools and lowers the dew point, while B2 slightly warms up. From 30 to 90 seconds: the temperature and humidity deviation decreases, and the plume weakens; A* gradually reduces VAV and guide angle offset, maintaining stability for 60 seconds before returning to the baseline. If events occur frequently within 3–5 minutes: L3 lowers the main unit dew point setting by 0.3℃ and raises the dry bulb setting by 0.2℃, reducing bypass operation frequency and energy consumption; after load recovery, it returns to a ramp-up state in 5–10 minutes.

[0093] Key performance indicators: grid temperature fluctuation ≤ ±0.2℃, humidity ≤ ±1%RH (minute-level recovery during event period); pressure difference stable at +15~+25Pa; bypass intervention frequency and time ratio reduced compared to traditional solutions; main unit reheat / humidification energy consumption 8–15% (depending on operating conditions); event identification trigger → terminal response ≤1s; 90% convergence ≤2–4min.

[0094] Preferably, the system further includes a mixing section connected upstream of the main constant temperature and humidity treatment unit, used to mix fresh air and return air at an adjustable mixing ratio to form mixed air, wherein the mixing ratio is determined in real time by the hierarchical collaborative control unit based on load and energy consumption targets.

[0095] Preferably, the rapid cooling or heating module and the rapid dehumidification module in the bypass fine-tuning circulation unit are arranged independently and controlled by independent actuators to achieve functional decoupling of temperature and humidity regulation.

[0096] Preferably, the outlet of the independent air supply branch is located above the process equipment and is equipped with an adjustable guide component to change the air supply direction, so that the local compensation matches the dominant airflow direction of the disturbance source in the same or opposite direction.

[0097] Preferably, a visual processing module is provided between the image acquisition unit and the hierarchical collaborative control unit. The module uses a multi-scale convolutional neural network combined with temporal attention to perform target detection and instance segmentation, and performs optical flow estimation based on adjacent frames to reconstruct the thermal plume velocity field, and further calculates the condensation risk index.

[0098] In some embodiments, the complete multi-scale convolutional neural network (MS-CNN) structure is given below, consisting of "input layer—hidden layer—activation layer—output layer," with structural improvements made only at two key locations. The complete model structure (for "detection + instance segmentation + plume / dew sensing") is as follows:

[0099] The input layer receives two synchronized time-series frames: visible light and thermal infrared. It also receives auxiliary scalars from environmental sensors (indoor dry-bulb temperature, relative humidity, dew point, optical flow stability, event confidence, etc.) as conditional signals. Preprocessing includes geometric and radiometric calibration, distortion correction, brightness / temperature normalization, and mild temporal denoising; dual-modal spatial alignment.

[0100] Hidden layers (backbone and pyramid), dual-stream multi-scale backbone: constructs shallow / medium / deep three-level convolutional features for visible light and thermal infrared respectively, with residual connections between layers; each layer is followed by a standard normalization and activation layer. Cross-modal alignment and fusion: channel attention fusion is performed at the same scale to obtain shared multi-scale feature groups. Feature pyramid (FPN): aggregates semantics from bottom to top and supplements details from top to bottom, forming multi-resolution feature maps of P3 to P5. Lightweight temporal aggregation (basic version): performs simple weighted stacking in the feature space with a fixed window to enhance robustness in steady-state scenes (not an improvement in this case).

[0101] The activation layer uses SiLU / ReLU as the basic nonlinearity (see "Improvement ②" below for adjustable replacement). The output layer (multi-head) is as follows: Detection head: Detects anchor point free center points, outputting category and bounding box. Instance mask head: Outputs a binary mask of instances after ROI alignment. Plume / risk auxiliary head: Outputs the main direction and intensity level of the plume, the proportion of gate openings, and the confidence level of the condensation risk heatmap. Inference interface: Integrates detection / mask / plume / risk with multi-point temperature and humidity into a "regionalized perturbation vector," which is then used by hierarchical collaborative control.

[0102] Two structural improvements to the network and their corresponding technical effects are as follows: Improvement ①: A plume prior-guided dynamic scale routing layer (PSR-Route) (embedded between the pyramid and the output head) is added after the feature pyramid. This layer generates a plume prior map using thermal infrared temperature gradients and coarse optical flow, marking possible ascending channels, hot and cold spots, and weak plume boundaries. The routing layer assigns weights and access switches for "fine-scale / medium-scale / large-scale" channels to each spatial location: regions with significant priors and weak textures are directed to the fine-scale branch (preserving more spatial details and highlighting boundaries); large-scale, strongly perturbed regions preferentially enter the medium / large-scale branches (improving semantic convergence and robustness); non-critical regions proceed along the default channel to avoid wasting computation. The routing is lightweight gating, without introducing a heavy convolution stack; during training, consistency constraints are applied to "prior—routing—detection / segmentation results" to make it learnable and stable. The improved network identifies weak disturbances earlier: small plumes and partially obscured door gaps are directed into fine-scale branches, significantly reducing the false negative rate; bypass compensation can be triggered one control cycle earlier for early disturbances. It reduces invalid computation and false triggers: non-critical areas do not follow high-resolution paths, saving overall computing power; features in critical areas are more focused, reducing false alarms. Mask boundaries are cleaner: fine-scale features after routing make door edges, material cart outlines, and personnel limb boundaries clearer, resulting in more stable subsequent mask head output.

[0103] Improvement ②: Physically Aware Adjustable Activation Layer (PPA-Act) (replacing the standard activation layer within the backbone and pyramid). This replaces the standard activation layer in the network with a "Physically Aware Adjustable Activation Layer," whose response intensity is gated by environmental and visual priors: dew point margin, optical flow stability, event confidence, and sensor quality markers serve as external "modulation factors," amplifying or suppressing the activation output in channels or space. When the system identifies "high condensation risk," "strong and stable plume," or "high confidence in door opening / passage," it increases the sensitivity of characteristic channels related to these phenomena; when it detects "vapor obstruction, specular reflection, or low sensor confidence," it suppresses characteristic responses that may lead to false alarms. PPA-Act is a universal replacement that can be selectively enabled at various stages of the backbone and at the pyramid fusion points, with almost no change to the original network topology. Improvements result in stronger noise resistance and fewer false triggers: under highly reflective stainless steel surfaces, vapor atomization, or localized obstruction, the activation layer automatically reduces the response in visible light branches or unstable regions, reducing oversensitivity to "false anomalies." Smoother timing output: Improved response in stable plume regions, reduced transient flicker, and smoother plume direction and intensity curves, facilitating small-step control of guide angle and VAV volume, and reducing oscillations. Direct connection to control targets: When condensation risk increases, relevant characteristics are amplified, resulting in a more consistent risk heatmap; when risk uncertainty increases, the overall response is downweighted, and the control layer automatically adopts a conservative strategy, reducing energy consumption and large-scale adjustments.

[0104] The complete "input-hidden-activation-output" chain is as follows: Input layer: dual-modal temporal frames + conditional signals (dew margin, optical flow stability, event confidence, sensor quality). Hidden layer: dual-stream multi-scale convolutional backbone → cross-modal fusion → feature pyramid → improved ① PSR-Route dynamically selects scale channels; and retains basic temporal aggregation to enhance steady-state robustness. Activation layer: improved ② PPA-Act is used at key positions in the backbone and pyramid, adaptively adjusting the nonlinear response according to physical and event priors. Output layer: detection head (personnel / material cart / door), instance mask head (opening mask and dwelling area), plume / risk auxiliary head (main direction, intensity level, condensation risk heatmap and confidence) → fusion to generate "regionalized perturbation vector".

[0105] The improved convolutional neural network exhibits the following effects: With the help of PSR-Route, the system is more sensitive to weak disturbances and small openings, allowing bypass compensation to intervene earlier and reducing recovery time. With the help of PPA-Act, robustness is significantly improved in non-ideal scenarios such as steam, reflection, obstruction, and sensor drift, reducing false triggering and ineffective compensation, resulting in smoother control and lower energy consumption. Both modifications are structurally searchable features, forming a synergistic innovation loop with "bypass priority, main loop follow-up, and gridded differentiated air supply."

[0106] like Figure 2 The system structure diagram shown is as follows, including: Main airflow path (outdoor fresh air → fresh air pretreatment → mixing section → main constant temperature and humidity unit → main supply air → clean room → return air); Bypass rapid fine-tuning loop (air intake tee → bypass fan → rapid dehumidification / heating / humidification → HEPA terminal → critical workstation); Sensing-Computation-Control-Action chain (multi-point temperature and humidity / pressure difference / particle + visible light / thermal infrared cameras → edge computing and fusion → hierarchical collaborative control + PLC / DDC → actuator layer / VAV / flow guide / valve / frequency converter); Network and power supply (industrial switch, PoE, UPS, NTP); and upper-layer BMS / SCADA and historical database.

[0107] Preferably, the outputs of the multi-point environmental perception unit and the visual processing module are subjected to Kalman filtering to obtain a regionalized perturbation vector, which includes the discrimination of outliers.

[0108] Preferably, the hierarchical collaborative control unit sets a first threshold and a second threshold, as well as a corresponding hysteresis band: when the deviation or its rate of change in the regionalized disturbance vector exceeds the first threshold, bypass priority compensation is triggered; when the deviation continues to exceed the second threshold, the main constant temperature and humidity processing unit is linked; the hierarchical collaborative control unit uses constrained model predictive control (MPC) to generate the air supply distribution matrix, with a prediction time domain of 30 to 180 seconds, and simultaneously applies dew point constraints, pressure difference constraints, and energy consumption penalties; when optimization is not feasible or the solution fails, it degenerates into hierarchical PID control.

[0109] Preferably, the image acquisition unit performs the following target detection and instance segmentation process on the time-series image:

[0110] S1) Acquire visible light or thermal infrared images at a rate of no less than 15 frames per second, and perform distortion correction, white balance and brightness normalization preprocessing.

[0111] S2) Input the preprocessed image into the backbone of a convolutional neural network containing a feature pyramid to extract multi-scale features;

[0112] S3) Generates center heatmaps, width, height and offset of each target category through anchor point free detection head on multi-scale features, and outputs candidate box confidence scores for detecting personnel, material carts and equipment doors;

[0113] S4) For each candidate box, a binary mask for the candidate instance is generated using a mask branch based on ROI alignment to obtain the instance segmentation result;

[0114] S5) Perform non-maximum suppression (NMS) on the candidate boxes, with an IoU threshold in the range of 0.3 to 0.6;

[0115] S6) Multi-target tracking using Kalman filter prediction and Hungarian matching is used to associate cross-frame instances, and the dwell time Δt of each instance in the preset area and the virtual gate crossing event are calculated. When Δt≥T1, it is determined as a dwell event, and when the gate crossing occurs, it is determined as a pass event. T1 is the preset dwell threshold.

[0116] S7) Perform segmentation on the ROI of the door and calculate the proportion of the opening mask θ. When θ≥θ_open, output the open state. θ_open is the preset opening threshold.

[0117] Preferably, the optical flow estimation process for convective plumes is as follows:

[0118] S1) Acquire thermal infrared images at a rate of no less than 15 frames per second and perform radiometric calibration and non-uniformity correction, converting pixel radiance into a temperature field T(x,y,t);

[0119] S2) Establish a background temperature model T_bg(x,y,t) based on exponential weighted moving average, calculate the temperature anomaly ΔT=T−T_bg, generate a plume candidate mask M based on the condition ΔT≥γ2, where γ2 is a configurable threshold, and perform morphological opening and closing operations to obtain the plume region ROI.

[0120] S3) At the pyramid multi-scale, the velocity field v(x,y)=(v_x,v_y) is solved by TV-L1 dense optical flow for ROI. The data terms are constructed with the assumption of temperature invariance, the regularization terms are smoothed by total variation TV, and the data terms are weighted by temperature gradient to enhance the stability of low texture region.

[0121] S4) Perform forward and backward consistency checks and median filtering on the optical flow results to remove outlier vectors, and apply physical constraints, including prioritizing the rising component coaxial with the direction of gravity, to obtain the plume velocity field.

[0122] S5) Based on the camera's geometric calibration, map v(x,y) to the cleanroom coordinate system and calculate the principal direction angle θ = atan2(v_y,v_x); and characterize the local sensible thermal disturbance intensity with the disturbance intensity index H = mean_ROI(∥v∥·ΔT); when H ≥ H_min, it is judged as a significant sensible thermal disturbance; H_min is the intensity threshold; atan2() is the two-parameter arctangent function; mean_ROI() is the spatial average of the field within the ROI region; v_x and v_y are the horizontal and vertical velocities of the velocity v, respectively;

[0123] S6) Outputs {θ, H} and together with the temperature and humidity sensing data, forms a regionalized disturbance vector, which is used by the hierarchical collaborative control unit for bypass priority compensation and air supply distribution matrix update.

[0124] Preferably, the process for calculating and outputting the condensation risk index based on the thermal infrared temperature field and the ambient dew point includes:

[0125] S1) Acquire thermal infrared time-series images and complete radiometric calibration, non-uniformity correction and bad pixel repair to obtain a time series of pixel-level surface temperature distribution;

[0126] S2) Obtain the ambient dry-bulb temperature and relative humidity from the cleanroom background temperature and humidity sensor, and determine the ambient dew point temperature according to the standard humid air properties algorithm;

[0127] S3) Compare the surface temperature and dew point temperature by pixel, define the dew point margin as the temperature difference between the two, and mark candidate dew condensation areas according to the set threshold. Perform connected component filtering and morphological opening and closing operations on the candidate areas to remove noise and isolated small blocks.

[0128] S4) Evaluate the three risk components separately within the candidate region:

[0129] Temperature difference risk component: The greater the degree to which the surface temperature is below or close to the dew point, the higher the risk level;

[0130] Cooling rate risk component: The faster the cooling occurs and the longer the duration of cooling within the preset time window, the higher the risk level.

[0131] Convection enhancement component: Combining the airflow velocity and direction obtained from optical flow estimation, the more favorable the airflow is to remove sensible heat or promote local cooling, the higher the risk level.

[0132] S5) Normalize and fuse the above three types of components according to preset weights to obtain a pixel-level condensation risk index; and perform spatial statistics on the pixel-level index in the candidate region to obtain a regional-level condensation risk index.

[0133] S6) Set two thresholds, high and low, with hysteresis: when the regional condensation risk index reaches the high threshold, trigger the dehumidification priority or reheat priority strategy; when it is between the low and high thresholds, perform bypass micro-compensation; when it is below the low threshold, only maintain monitoring;

[0134] S7) Outputs a condensation risk heat map and a regional condensation risk index, and writes it into a regional disturbance vector for the hierarchical collaborative control unit to update the air supply allocation and execution strategy.

[0135] This invention provides an air conditioning system for rapidly maintaining a constant temperature and humidity in a clean environment, and the beneficial technical effects it achieves are as follows:

[0136] 1. This application introduces a bypass fine-tuning loop on the basis of the traditional main loop, and uses the generated regionalized disturbance vector as the feedforward signal. By identifying personnel / materials / door events, estimating thermal infrared plume optical flow, and assessing condensation risk index, it achieves sensible / latent heat decoupling control with "bypass priority and main loop follow-up". Combined with N×M grid differentiated air supply, it can suppress local fluctuations on a second to minute basis. Under typical operating conditions, temperature fluctuations converge to ±0.2℃ and humidity to around ±1%RH, significantly reducing overshoot and recovery time.

[0137] 2. This application implements sensible / latent heat independently to avoid repeated offsetting of "deep cooling + reheating, excessive humidification + re-dehumidification"; model predictive control incorporates dew point, pressure difference, and energy consumption into the constraints, dynamically allocating airflow and dew point of the main / bypass and each grid according to the load, prioritizing small-scale compensation near the disturbance source, reducing large-scale system-wide adjustments and long-term maintenance of high airflow. This reduces cooling, reheating, and humidification water consumption, decreases equipment start-up and shutdown frequency and valve operation, and improves long-term operational stability and economy.

[0138] 3. This application's system adopts a modular architecture. The bypass unit, image acquisition, and vision processing module can be deployed outside existing cleanrooms, compatible with existing BMS / PLCs. Camera geometric self-calibration and robust occlusion / drift fusion ensure data reliability. Optimization failures automatically degenerate into hierarchical PID while maintaining differential pressure / cleanliness constraints. The algorithm threshold and prediction time domain can be parameterized, requiring minimal maintenance and facilitating phased implementation and rapid replication and promotion.

[0139] The above provides a detailed description of an air conditioning system for rapidly maintaining a clean, constant temperature and humidity. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, those skilled in the art will recognize that, based on the ideas and methods of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An air conditioning system for rapid maintenance of clean room temperature and humidity constancy, characterized by, The system comprises: a fresh air pre-processing unit for one or more of pre-cooling, pre-heating, dehumidification, and humidification of outdoor fresh air; a main constant temperature and humidity processing unit in communication with the fresh air pre-processing unit for adjusting the humidity and temperature of the fresh air to obtain basic supply air for a clean room; a clean room and main supply air duct and return air duct in communication with the main constant temperature and humidity processing unit; a bypass fine adjustment circulation unit in communication with the clean room or return air duct at one end for extracting part of the air and in communication with the clean room through an independent supply air branch at the other end, the bypass fine adjustment circulation unit comprising a bypass fan, a rapid cooling or heating module, a rapid dehumidification module, and a rapid humidification module; a multi-point environment sensing unit comprising temperature and humidity sensors arranged in the background area and at least one process area of the clean room, and an image acquisition unit facing the process area, which performs target detection and instance segmentation on time-series images to identify personnel, material carts, open equipment hatch doors, and corresponding stay or pass events, performs optical flow estimation on thermal infrared images to obtain the direction and intensity of local sensible heat disturbance, and calculates a dew risk index based on the thermal infrared temperature field and the ambient dew point, and fuses the dew risk index with the temperature and humidity data to form a regionalized disturbance vector; a hierarchical collaborative control unit electrically connected with the fresh air pre-processing unit, the main constant temperature and humidity processing unit, the bypass fine adjustment circulation unit, and the multi-point environment sensing unit; the clean room is divided into an N×M area grid, and the hierarchical collaborative control unit maps the regionalized disturbance vector to a supply air distribution matrix and implements differentiated air supply to the corresponding grid through adjustable guide vanes, cyclone air outlets, and variable air volume terminals.

2. The air conditioning system for maintaining constant temperature and humidity in a clean room quickly according to claim 1, wherein The system further comprises a mixing section in communication with the upstream of the main constant temperature and humidity processing unit for mixing fresh air and return air at an adjustable mixing ratio to form mixed air, and the mixing ratio is determined in real time by the hierarchical collaborative control unit according to the load and energy consumption target.

3. The air conditioning system for maintaining constant temperature and humidity in a clean room quickly according to claim 1, wherein The rapid cooling or heating module and the rapid dehumidification module in the bypass fine adjustment circulation unit are independently arranged and controlled by independent actuators to realize functional decoupling of temperature and humidity adjustment.

4. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 3, wherein The outlet of the independent supply air branch is located above the process equipment, and adjustable guide members are arranged to change the direction of the supply air, so that the local compensation and the dominant airflow direction of the disturbance source are matched in the same direction or opposite direction.

5. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 4, wherein A visual processing module is arranged between the image acquisition unit and the hierarchical collaborative control unit, which uses a multi-scale convolutional neural network combined with time-series attention for target detection and instance segmentation, and performs optical flow estimation based on adjacent frames to reconstruct the thermal plume velocity field and further calculate the dew risk index.

6. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 4, wherein The output of the multi-point environment sensing unit and the output of the visual processing module are subjected to Kalman filtering to obtain the regionalized disturbance vector, and the discrimination of abnormal values is included.

7. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 1, wherein The hierarchical cooperative control unit sets a first threshold value and a second threshold value and corresponding hysteresis bands: when the deviation in the regional disturbance vector or its rate of change exceeds the first threshold value, the bypass priority compensation is triggered; when the deviation continues to exceed the second threshold value, the linkage main constant temperature and humidity processing unit is linked; the hierarchical cooperative control unit generates a supply air distribution matrix using model predictive control (MPC) with constraints, with a prediction time domain of 30-180 seconds, and simultaneously applies dew point constraints, pressure difference constraints and energy consumption penalties; when optimization is not feasible or fails, it degenerates into layered PID control.

8. The air conditioning system for maintaining constant temperature and humidity in a clean room quickly according to claim 1, wherein The image acquisition unit performs the following target detection and instance segmentation process on the time sequence images: S1) collect visible light images or thermal infrared images at no less than 15 frames per second, and perform distortion correction, white balance and brightness normalization preprocessing; S2) input the preprocessed images into a convolutional neural network backbone containing a feature pyramid to extract multi-scale features; S3) generate center heat maps, width and height, and offset for each target category on the multi-scale features through an anchor-free detection head to output candidate box confidence, which is used to detect personnel, material vehicles, and equipment hatches; S4) for each candidate box, a mask branch based on ROI alignment is used to generate a binary mask of the candidate instance to obtain the instance segmentation result; S5) non-maximum suppression (NMS) is performed on the candidate boxes, with an IoU threshold in the interval of 0.3-0.6; S6) cross-frame instances are associated using Kalman filter prediction and multi-target tracking based on Hungarian matching, and the residence time Δt of each instance in the preset area and the crossing event of the virtual door line are calculated, with Δt≥T1 being determined as a residence event and the door line crossing being determined as a passing event, T1 being a preset residence threshold; S7) the door ROI is segmented and the opening mask ratio θ is calculated, and when θ≥θ_open, the open state is output, θ_open being a preset opening threshold.

9. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 1, wherein The optical flow estimation process for the convection plume is as follows: S1) collect thermal infrared images at no less than 15 frames per second and perform radiation scaling and non-uniformity correction to convert pixel radiation to temperature field T(x, y, t); S2) establish a background temperature model T_bg(x, y, t) based on exponential weighted moving average, calculate the temperature anomaly ΔT=T-T_bg, generate plume candidate mask M from the condition ΔT≥γ2, γ2 being a configurable threshold, and perform morphological opening and closing operations to obtain the plume region ROI; S3) on the pyramid multi-scale, for the ROI, TV-L1 dense optical flow is used to solve the velocity field v(x, y)=(v_x, v_y), where the data term is constructed based on the temperature invariance assumption, the regularization term uses total variation (TV) smoothing, and the data term is weighted by the temperature gradient to enhance the stability of low texture areas; S4) perform forward and backward consistency check and median filter on the optical flow result to remove outlier vectors, and apply physical constraints, including the upward component coaxial with the gravity direction, to obtain the plume velocity field; S5) Map v(x, y) to the cleanroom coordinate system according to camera geometry calibration, calculate the principal direction angle θ = atan2(v_y, v_x); and represent the local sensible heat disturbance intensity with the disturbance intensity index H = mean_ROI(||v||·ΔT); determine a significant sensible heat disturbance when H ≥ H_min; H_min is the intensity threshold; atan2() is the two-parameter arctangent function; mean_ROI() is the spatial average of the field in the ROI region; v_x and v_y are the horizontal and vertical velocities of the velocity v, respectively; S6) Output {θ, H} and jointly form the regionalized disturbance vector with the temperature and humidity sensor data for the hierarchical collaborative control unit to update the air supply distribution matrix and execute the strategy.

10. The air conditioning system for rapid maintenance of constant temperature and humidity in a clean room according to claim 1, wherein The process for outputting a dewing risk index based on a thermal infrared temperature field and an ambient dew point includes: S1) Collect thermal infrared time-series images and complete radiation calibration, non-uniformity correction, and bad pixel repair to obtain a time series of pixel-level surface temperature distribution; S2) Obtain the ambient dry-bulb temperature and relative humidity from the cleanroom background temperature and humidity sensor, and determine the ambient dew point temperature according to the standard moist air property algorithm; S3) Compare the surface temperature and the dew point temperature according to the pixel, define the dew point margin as the temperature difference between the two, and mark the candidate dewing area according to the set threshold, and perform connected domain screening and morphological opening and closing operations on the candidate area to remove noise and isolated small blocks; S4) Evaluate three types of risk components in the candidate area: Temperature difference risk component: the greater the degree to which the surface temperature is below or approaches the dew point, the higher the risk level; Cooling rate risk component: the faster the cooling is within a preset time window, the longer the duration, the higher the risk level; Convection enhancement component: combine the air flow velocity and direction estimated by optical flow to determine the risk level, the greater the air flow is conducive to carrying away sensible heat or promoting local cooling, the higher the risk level; S5) Normalize and fuse the above three types of components according to the preset weight to obtain the pixel-level dewing risk index; and perform spatial statistics on the pixel-level index in the candidate area to obtain the regional-level dewing risk index; S6) Set high and low thresholds with hysteresis: trigger the dehumidification priority or reheating priority strategy when the regional-level dewing risk index reaches the high threshold; execute bypass micro-compensation when it is between the low and high thresholds; and only monitor when it is below the low threshold; S7) Output the dewing risk thermograph and the regional-level dewing risk index, and write them into the regionalized disturbance vector for the hierarchical collaborative control unit to update the air supply distribution and execute the strategy.

Images